Repositorio Institucional de Documentos

2014 96 Jaime Ibáñez Pereda Development of EEG-based technologies for the characterization and treatment of neurological diseases affecting the motor function Departamento Director/es Informática e Ingeniería de Sistemas Mínguez Zafra, Javier Castillo Sobrino, María Dolores del Serrano Moreno, José Ignacio Director/es Tesis Doctoral Autor Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA Departamento Director/es Jaime Ibáñez Pereda DEVELOPMENT OF EEG-BASED TECHNOLOGIES FOR THE CHARACTERIZATION AND TREATMENT OF NEUROLOGICAL DISEASES AFFECTING THE MOTOR FUNCTION Director/es Informática e Ingeniería de Sistemas Mínguez Zafra, Javier Castillo Sobrino, María Dolores del Serrano Moreno, José Ignacio Tesis Doctoral Autor Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA Departamento Director/es Director/es Tesis Doctoral Autor Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA Development of EEG-based technologies for the characterization and treatment of neurological diseases affecting the motor function TESIS DOCTORAL Autor Jaime Ib´a˜nez Pereda Directores Mar´ıa Dolores del Castillo Sobrino Jos´e Ignacio Serrano Moreno Javier M´ınguez Zafra UNIVERSIDAD DE ZARAGOZA Escuela de Ingenier´ıa y Arquitectura Departamento de Inform´atica e Ingenier´ıa de Sistemas Zaragoza - 2014 Escuela de Ingenier´ıa y Arquitectura Departamento de Inform´atica e Ingenier´ıa de Sistemas Development of EEG-based technologies for the characterization and treatment of neurological diseases affecting the motor function Jaime Ib´a˜nez Pereda Thesis submitted in fulfillment of the requirements for the degree of Doctor of Philosophy in Computer Science and Systems Engineering Supervised by: Mar´ıa Dolores del Castillo Sobrino Jos´e Ignacio Serrano Moreno Javier M´ınguez Zafra Universidad de Zaragoza September, 2014 Contents Resumen 1 Abstract 3 1 Introduction 5 1.1 Purpose ...................................... 7 1.2 Research lines and projects giving rise to this thesis . . . . . . . . . . . . . 8 1.3 Motivation .................................... 9 1.4 Objectives..................................... 11 1.5 Methodology ................................... 13 1.6 Chaptersdescription............................... 15 2 EEG-based systems to study the motor function and BCI technologies in the neurorehabilitation field 17 2.1 Structures of the central nervous system involved in the generation of volitionalmovements................................. 18 2.2 The use of the EEG signal to extract cortical activity related to the movement 20 2.2.1 Cortical patterns related to the motor function and visible in the EEGsignal ................................ 22 2.3 BCI systems for motor rehabilitation . . . . . . . . . . . . . . . . . . . . . . 27 3 EEG-based predictive classification of analytical upper-limb movements 31 3.1 Abstract...................................... 31 3.2 Introduction.................................... 33 3.3 Methods...................................... 34 3.3.1 Participants and experimental procedure . . . . . . . . . . . . . . . . 34 3.3.2 DataAcquisition............................. 35 3.3.3 Data processing and classifier design . . . . . . . . . . . . . . . . . . 36 3.3.4 Additional experiments to prove the validity of the classification results 38 3.3.5 Statistical analysis of the features selected . . . . . . . . . . . . . . . 39 CONTENTS 3.4 Results....................................... 40 3.4.1 Classification results of 7 analytical movements . . . . . . . . . . . . 40 3.4.2 Results of validation experiment 1: Estimation of the significance of the obtained class description . . . . . . . . . . . . . . . . . . . . . . 41 3.4.3 Results of validation experiment 2: Analysis of the influence of the time segments location . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.4.4 Results of validation experiment 3: Extraction of other tentative sources of information . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.5 Discussion..................................... 43 3.6 Chapterconclusions ............................... 46 4 Study of alprazolam-induced changes in cortical oscillations and tremors of patients with ET 49 4.1 Abstract...................................... 49 4.2 Introduction.................................... 50 4.3 Methods...................................... 51 4.3.1 Patients, data acquisition and experimental procedure . . . . . . . . 51 4.3.2 Data processing and analysis . . . . . . . . . . . . . . . . . . . . . . 52 4.3.3 Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.4 Results....................................... 54 4.5 Discussion..................................... 59 4.6 Chapterconclusions ............................... 61 5 Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor 63 5.1 Abstract...................................... 63 5.2 Introduction.................................... 65 5.3 Methods...................................... 67 5.3.1 Experimental protocols . . . . . . . . . . . . . . . . . . . . . . . . . 67 5.3.2 Dataacquisition ............................. 68 5.3.3 Detection of the movement onset with the gyroscopes . . . . . . . . 69 5.3.4 Description of the ODIM architecture . . . . . . . . . . . . . . . . . 69 5.3.5 Design of the mHRI with the ODIM system . . . . . . . . . . . . . . 72 5.3.6 Results in Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . 73 5.3.7 Results in Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . . 76 5.4 Discussion..................................... 77 5.5 Chapterconclusions ............................... 81 ii List of Tables 3.1 Tasks classification results. The Recall (R) and Precision (P) results are presented for each subject and task. The last row shows the average results across subjects. The last column shows the average results across joints moved........................................ 41 4.1 Main baseline demographic and clinical variables. Fahn, Tolosa, Marin Essential Tremor Rating Scale (ETRS). . . . . . . . . . . . . . . . . . . . . 51 5.1 Advantages and disadvantages of EEG, EMG and gyroscopes to detect and track tremors and voluntary movements. . . . . . . . . . . . . . . . . . . . . 73 5.2 Selected features (Channel-Frequency pairs) by the ODIM. . . . . . . . . . 74 5.3 Classification results of the ODIM. . . . . . . . . . . . . . . . . . . . . . . . 75 5.4 Classification results of the ODIM in Experiment 2. . . . . . . . . . . . . . 78 6.1 Demographic table of the patients participating in the present study. . . . . 85 6.2 Detection results obtained with control subjects and patients. . . . . . . . . 96 6.3 Features selected by the ERD-based detector for the control group. . . . . . 96 6.4 Features selected by the ERD-based detector for the patients. . . . . . . . . 97 6.5 Gain in the performance of the detector (GT in %) when using the combined information of the ERD and BP compared to the use of either of these patternsalone. .................................. 98 6.6 ISI and FMI scales of the patients before and after the BCI intervention. . . 100 Resumen La electroencefalograf´ıa, o registro de potenciales el´ectricos generados por el cerebro mediante electrodos superficiales colocados sobre el cuero cabelludo, es una t´ecnica ampliamente utilizada a nivel cl´ınico, siendo tradicionalmente ´util para un primer diagn´ostico en alteraciones en la corteza cerebral, trastornos del sue˜no o para la b´usqueda de focos epil´epticos en pacientes con este tipo de crisis. En las ´ultimas d´ecadas, la incursi´on de las tecnolog´ıas de la informaci´on y comunicaci´on en los campos cl´ınicos y, en este caso, de estudios electrofisiol´ogicos, han contribuido a que este tipo de herramientas hayan sido propuestas para un gran n´umero de aplicaciones, entre las que destacan los estudios de b´usqueda de correlatos neuronales de la actividad motora en los seres humanos y las interfaces cerebro-computador que proporcionan una l´ınea directa de interacci´on entre la actividad cerebral y sistemas automatizados. Las propiedades ´optimas en cuanto a resoluci´on temporal de la se˜nal de electroencefalograf´ıa hacen que esta t´ecnica permita conocer, sin pr´acticamente retraso en el tiempo, las caracter´ısticas de los procesos el´ectricos de poblaciones de neuronas en la corteza motora. Estos procesos pueden desencadenarse como consecuencia de la realizaci´on de una acci´on voluntaria o ser provocados por el comportamiento patol´ogico del cerebro, causando dificultades en el control motor. De este modo, se abre la puerta al desarrollo de nuevas t´ecnicas que caracterizan los procesos corticales patol´ogicos y sanos asociados con el procesamiento motor, como puede ser la ejecuci´on, visualizaci´on o imaginaci´on de un movimiento voluntario, la realizaci´on de tareas funcionales por parte de pacientes con alteraciones cerebrales causadas por una lesi´on, o la manifestaci´on de movimientos involuntarios que alteran la capacidad funcional del paciente. En todos estos casos, la se˜nal de electroencefalograf´ıa puede presentar patrones, observables en las variaciones de las actividades oscilatorias o de baja frecuencia de ciertas componentes de la se˜nal, que permiten conocer informaci´on relevante acerca de los procesos corticales que desencadenan el movimiento. Este trabajo de tesis presenta un conjunto de estudios en los que se aplican t´ecnicas de procesamiento de la se˜nal y de miner´ıa de datos en sistemas en tiempo real para el registro, caracterizaci´on y condicionamiento de la actividad de la corteza motora en sujetos Resumen sanos y en pacientes con des´ordenes neurol´ogicos que afectan a la capacidad motora. En concreto, la presente tesis incluye estudios con pacientes de dos de las patolog´ıas de origen neurol´ogico m´as extendidas: pacientes con temblor esencial y pacientes que han sufrido un accidente cerebrovascular. Los mecanismos neuronales de acci´on y tecnolog´ıas para el tratamiento de ambas patolog´ıas son en la actualidad ampliamente investigados por numerosos grupos en todo el mundo. A lo largo de los cap´ıtulos que conforman este trabajo de tesis se presentan resultados sobre la actividad cortical normal relacionada con la planificaci´on y ejecuci´on de acciones motoras con el miembro superior, y ´esta se contrapone a la actividad patol´ogica que los pacientes presentan y que est´a directamente relacionada con la incapacidad motora que se puede observar y cuantificar por medio de t´ecnicas de estimaci´on de la actividad muscular y/o de movimiento. En los cap´ıtulos iniciales se presenta una revisi´on de los conceptos b´asicos del papel de la corteza cerebral en el control motor y de c´omo la actividad electroencefalogr´afica permite su an´alisis y su condicionamiento, se propone un estudio de interacci´on cortico-muscular a la frecuencia del temblor en pacientes con temblor esencial con el objetivo de conocer los efectos de un f´armaco en estos pacientes y, por ´ultimo, se presenta un estudio basado en algoritmos evolutivos para la identificaci´on de patrones corticales asociados con la planificaci´on de tareas motoras realizadas con un mismo brazo. En la segunda parte del trabajo se presentan dos propuestas de interfaces cerebro-computador para ser utilizadas en entornos de rehabilitaci´on en pacientes con temblor esencial o con un ictus. En la primera propuesta se plantea el uso de un sistema de electroencefalograf´ıa para anticipaci´on de movimientos voluntarios como parte integrada de una plataforma multimodal de estimaci´on y supresi´on del temblor. En la segunda propuesta se plantea un paradigma de condicionamiento basado en la identificaci´on de la intenci´on motora con precisi´on temporal para pacientes con ictus, y ´este es evaluado en un grupo de pacientes durante un intervalo de un mes a lo largo del cual se realizan hasta ocho intervenciones. De este modo, el objetivo general de esta tesis es proponer soluciones tecnol´ogicas que permitan profundizar en el conocimiento de los mecanismos neuronales que permiten la generaci´on de la acci´on motora voluntaria, la caracterizaci´on de la actividad cortical patol´ogica en pacientes con des´ordenes motores causados por afecciones nerviosas, la b´usqueda de t´ecnicas ´optimas para la estimaci´on de la planificaci´on e intencionalidad motora y la propuesta de nuevas formas de rehabilitaci´on de pacientes con las patolog´ıas previamente indicadas. Se espera que los resultados aqu´ı presentados sirvan de base para el posterior desarrollo de plataformas cercanas al ´ambito cl´ınico y que supongan un avance en el diagn´ostico, pron´ostico y tratamiento de patolog´ıas del sistema nervioso central que conllevan alteraciones en el control motor. 2 Abstract The electroencephalography consists in the recording of electric potentials generated in the brain acquired by means of surface electrodes distributed on the scalp. This technique is widely used in the clinical field, and it is of relevance for the first diagnosis of damages in the brain cortex, the study of sleep disorders or the localization of seizure foci in patients with epilepsy. During the last decades, the use of the information and communication technologies in the clinical field, and in this case in electrophysiological studies, has contributed to broaden the field of applications of the electroencephalographic systems. Among these, studies on the neural correlates of the motor activity in human beings and the development of brain-computer interfaces (providing an interaction line between the brain activity and automatic systems) stand out. Due to the optimal properties in terms of temporal resolution of the electroencephalographic signal, it is now possible to study, with almost no temporal delay, the characteristics of the electrical processes produced by neuronal populations in the motor cortex. These processes may appear as a consequence of the execution of a certain voluntary action, or may be caused by the pathological function of the brain, leading to an affected motor function. As a result of these advances in electroencephalographic systems new techniques are developed, studying the movement-related healthy and pathological cortical processes during the execution, visualization or imagination of a voluntary movement, the performance of functional tasks by patients suffering brain damages due to a certain lesion, or the presence of involuntary movements, such as tremors. In all these cases the electroencephalographic signal presents certain patterns, observed in the variations of the oscillatory or low-frequency components of the signal, that may allow the extraction of relevant information regarding the cortical processes giving rise to the studied movement. This thesis presents a set of studies applying signal processing and data mining techniques in real-time working systems to register, characterize and condition the movementrelated cortical activity of healthy subjects and of patients with neurological disorders affecting the motor function. Patients with two of the most widespread neurological affections impairing the motor function are considered here: patients with essential tremor Abstract and patients who have suffered a cerebro-vascular accident. The neurophysiological action mechanisms and treatment technologies for both pathologies are currently under extensive research by a number of groups around the world. The different chapters in this thesis present results regarding the normal cortical activity associated with the planning and execution of motor actions with the upper-limb, and the pathological activity related to the patients’ motor dysfunction (measurable with muscle electrodes or movement sensors). The initial chapters of the book present i) a revision of the basic concepts regarding the role of the cerebral cortex in the motor control and the way in which the electroencephalographic activity allows its analysis and conditioning, ii) a study on the cortico-muscular interaction at the tremor frequency in patients with essential tremor under the effects of a drug reducing their tremor, and finally iii) a study based on evolutionary algorithms that aims to identify cortical patterns related to the planning of a number of motor tasks performed with a single arm. In the second half of the thesis book, two brain-computer interface systems to be used in rehabilitation scenarios with essential tremor patients and with patients with a stroke are proposed. In the first system, the electroencephalographic activity is used to anticipate voluntary movement actions, and this information is integrated in a multimodal platform estimating and suppressing the pathological tremors. In the second case, a conditioning paradigm for stroke patients based on the identification of the motor intention with temporal precision is presented and tested with a cohort of four patients along a month during which the patients undergo eight intervention sessions. To summarize, the general objective of this thesis is to propose technological solutions that lead to i) a better understanding about the neuronal mechanisms that mediate voluntary motor actions, ii) the characterization of the cerebral cortical activity in patients with neurological affections, iii) the search of optimal techniques to estimate the motor planning and intention, and iv) the proposal of new rehabilitation strategies in patients with the aforementioned pathologies. It is expected that the results presented become the basis for future developments of technological platforms that can be integrated in the clinical practice and allow an improved diagnosis, prognosis and treatment of pathologies of the central nervous system. 4 Chapter 1 Introduction Exploring the nervous system implies studying the essence of human beings. It entails understanding the mechanisms through which two people perceive in a different way the same song [Ramachandran et al., 2001], how a child’s personality is conditioned by the mechanisms of language acquisition [Pinker and Jackendoff, 2005], or how is it possible to achieve, by means of intensive practise, that a tennis forehand results in a winner point by slightly touching the side line of the opponent’s field [Kandel et al., 2000]. Clocks tick, skyscrapers and bridges vibrate... and neurons oscillate [Buzs´aki, 2006], and with their oscillation they communicate with other neurons, giving rise to highly complex associations that result into all kinds of mental processes, from which we consciously recognize a small fraction [Dijksterhuis and Nordgren, 2006]. Discovering the correlates between electrical processes observed in the brain and observable or measurable human acts leads to understanding the neuronal principles that rule human behaviour and to further describing the neurophysiological mechanisms of neurological disorders. Nowadays there exist different windows on the brain, covering many and complementary spatial regions and time intervals. The electroencephalogram provides a window on the mind, albeit one that is often clouded by technical and other limitations [Nunez and Srinivasan, 2006]. Since Hans Berger placed in 1924 the first scalp electrodes to observe alpha rhythms [Berger, 1929] up to now, the number of possibilities that electroencephalographic systems provide has grown exponentially. Indeed, the electroencephalography is now an essential tool in any neurology department and new potential applications are expected to be a reality in the near future, allowing an improved analysis and treatment of certain neurological pathologies. The analysis of the nervous system and the description of how it works may be carried out from, a priori, independent research fields such as medicine, informatics, robotics, physiotherapy, chemistry, etcetera. This thesis aims at being a small contribution in the neuroscience field from the biomedical engineering perspective. To that end, a group of four independent studies using electroencephalographic systems are proposed in the framework Chapter 1. Introduction of the analysis and treatment of neurological diseases causing motor disabilities. The global objectives of the entire work are to further understand the cortical mechanisms of voluntary movement planning and execution, how they may be affected by the pathological brain structures of patients with neurological diseases (specifically patients with essential tremor and patients who have suffered a stroke) and, eventually, how functional recovery may be achieved either with assistive technologies taking advantage of the online characterization of the motor cortex, or using conditioning paradigms of the cortical activity that elicit plastic changes resulting in an improvement of the motor function. In order to meet these goals, advanced signal processing techniques and data mining algorithms are used to characterize the cortical changes of subjects performing movement actions, and the observed results are used to characterize the action mechanisms and evolution of the two studied pathologies. Additionally, and of special relevance in this thesis, real-time systems characterizing the cortical activity related to motor-planning online are programmed and used to implement brain-computer interfaces that allow the patients to use the brain activity to control external neuroprosthetic devices. 6 1.1 Purpose 1.1 Purpose The past decades have witnessed the achievement of important advances in the electrophysiological study of the nervous system, both in terms of advances in the technology used to acquire neurophysiological information and in terms of new algorithms developed to process this information. These advances have allowed the acquisition, storage, characterization and even the conditioning of the nervous system activity at a local level (measuring neurons spike trains) or at a global level (considering the activity of populations of neural networks), and focusing on the central nervous system (brain and spinal cord) or the peripheral nervous system (characterising the activity of motor neurons, reflexes etc.). All these achievements provide new opportunities to analyse the function and the structure of the nervous system associated with the execution of daily-living activities by the human beings. Focusing on the applications using electroencephalographic (EEG) systems to analyse and treat neurological diseases affecting the motor function, a wide variety of studies have been published during the last two decades. These studies can be divided in two generic research lines: neurophysiological studies characterising the motor cortex processes associated with neurological conditions and treatments, and experiments aimed to validate rehabilitation technologies for the motor function. In the first case, the goal is to relate the neurological pathologies and their evolution with altered cortical activation patterns in the patients, so that it becomes possible to find precise descriptions of the neurological mechanisms that lead to affected motor control. Examples among the large amount of studies in this field are the experiments analysing the cortico-muscular interaction at the tremor frequency in patients suffering from tremor-related pathologies (see for example [Hellwig et al., 2001; Timmermann et al., 2002]), or the experiments characterizing the altered cortical activation patterns in patients with brain damages to provide ways of predicting the degree of recovery (see [Burghaus et al., 2007]). The second group of EEG-based applications for patients with motor disorders are those in which the main goal is to provide new means of achieving functional rehabilitation of the patients. In this case the main objective is to develop brain-computer interfaces (BCIs), i.e. devices using the electrical activity acquired from specific scalp regions to provide a feedback (typically visual or proprioceptive) to the patient. This group of EEGbased applications may in turn be further divided into BCIs for motor compensation and BCIs for motor recovery. In the first case, BCI systems extract information from the cortical activity and convert it into control signals used to operate external devices such Jaime Ib´a˜nez Pereda 7 Chapter 1. Introduction as robotic or prosthetic arms, wheelchairs or spellers. An illustrative example in this field may be taken from BCI technologies providing a communication channel for patients with complete locked-in syndrome, which constitutes a significant gain in the possibilities of these patients to interact with the environment [Hinterberger et al., 2003; Birbaumer, 2006]. On the other hand, BCI systems aimed to recover the lost motor capacity are mainly focused on finding ways to condition the neural activity of specific cortical regions, resulting in an improvement of the patient’s motor function. In this line, the main area of research at present is oriented towards new interventions for patients suffering from a spinal cord injury or with a stroke [Silvoni et al., 2011; Ramos-Murguialday et al., 2013]. The purpose of this thesis is to evaluate the potential uses of EEG-based systems for the analysis and treatment of neurological diseases affecting the motor capacity. To that end, four studies associated with the aforementioned research lines with EEG systems (neurophysiological studies of the normal and pathological motor function and studies of BCI technology either assisting or recovering the lost motor capacity) are presented and validated with control subjects and patients. The critical validation of the potential applications of this type of systems is built upon the obtained results and reached conclusions. 1.2 Research lines and projects giving rise to this thesis Currently, one of the most active research areas with EEG systems is the development of applications for patients with neurological disorders affecting the motor function. In this regard, EEG systems provide a window on the cortical electrical activity with high temporal precision, which is a critical factor when trying to study and model the nervous system. In addition, EEG systems are widely used in neurophysiological fields due to their advantages as compared to other alternatives: EEG systems are cost-effective practical systems for clinical environments that do not require ample rooms or restrictive conditions in terms of vulnerability against electrically noisy environments. All these factors have made these systems an attractive solution to carry out experimental procedures studying the neurophysiological characteristics of movement disorders and developing neurorehabilitation technologies, which has in turn received an important economical support from funding institutions in these research lines, both in the national and European domains. This thesis is the result of studies carried out in the framework of some of these funded projects. The first experiments performed for this thesis were carried out in the framework of the TREMOR European project (FP7-ICT-2007-224051, An ambulatory BCI-driven tremor suppression system based on functional electrical stimulation). This project proposed for 8 1.5 Methodology –Interaction with clinical environments supporting them in the design of ethical committees for the proposed studies, in the definition of patients inclusion criteria and in the recruitment of patients for the experiments. Jaime Ib´a˜nez Pereda 15 Chapter 1. Introduction 1.6 Chapters description The thesis is organized around four chapters, each of which related to the four proposed studies. Preceding these four chapters, Chapter 2 presents the neurophysiological basis and EEG-related knowledge regarding cortical control of the movement, which is used in the subsequent parts of the document. In the beginning of this chapter, a brief description of the basic nervous system structures involved in the generation of voluntary movements and how they interact with each other to achieve motor control of the body limbs is reviewed. After this description, the advantages and disadvantages of EEG technology as compared to other electrophysiological measurements of the brain are presented, thus justifying its use in the subsequent chapters. BCI systems for motor rehabilitation are briefly described in the end of the chapter. The first study is presented in Chapter 3 and it explores, using healthy subjects, the possibilities of modelling cortical activation patterns related to different analytical motor tasks with spatially close somatotopic representations. The methods section of this chapter describes in detail the data mining methodology used to this end. Results show both, the performance of a classifier of these movements and a set of tests to evaluate the validity of the results obtained. The final part of the chapter presents a critical analysis of the obtained results and describes possible scenarios in which the presented EEG-based application may be of interest. In Chapter 4 a novel application of EEG systems in ET patients is presented. In this case, a neurophysiological study of the effects of a therapeutic drug (alprazolam) in the tremor manifestation and in specific cortical oscillations in patients with ET is carried out. The first part of the chapter provides an up-to-date summary of neurophysiological studies in patients with ET, presenting the main hypotheses of the mechanisms and structures leading to tremor manifestation. Next, the experimental design used to analyse the effects of the drug along a certain time after its intake is described, and results are presented, which lead to establishing a hypothesis regarding the way in which alprazolam reduces the tremor: the increased presence of fast cortical rhythms (specially in the beta band) in patients with ET caused by a single dose intake of a benzodiazepine is tightly related with tremor reduction. The critical discussion of these results, comparing them with previous studies by other authors, as well as the presentation of the main technical limitations of the study are included in the final part of the chapter. Chapter 5 presents the development of a multimodal brain-neural-computer interface in which EEG technology is integrated with other movement-related sensors to predict vol16 1.6 Chapters description untary movements. In order to justify the proposed system concept, the potential benefits of a multimodal interface to control a tremor-compensation neuroprosthesis are exposed. After this, the global architecture of the multimodal interface is presented, giving special relevance to the EEG-based subsystem, and preliminary results with healthy subjects and ET patients are shown. In addition, results of the multimodal interface are included, in order to provide a proof of concept of the cooperative interaction between different subsystems combining information from cortical, muscular and gyroscopic sensors. A critical evaluation of the reached results and the future lines of study are included in the final part of the chapter. In Chapter 6, the last of the four proposed studies is presented. This study is aimed to develop an EEG-based BCI intervention for the motor recovery of the upper-limbs of stroke patients. To achieve this, the first part of the chapter includes a brief review of related studies in the field. The proposed intervention integrates 4 different technologies: an EEG amplifier, an EMG amplifier, inertial sensors and an electrical stimulator. The experimental set-up and the used protocol are presented in the chapter. After this, preliminary results of the system function are presented. The chapter also includes results of a small clinical validation carried out with the system with four patients during eight sessions. The critical analysis of the obtained results and the possible therapeutic effects of the BCI system are presented in the final part of the chapter. Finally, the general results obtained in the proposed studies are summarized in the last chapter (Chapter 7). This chapter also includes a detailed discussion regarding the achieved results here, the main limitations that have been found throughout the process of carrying out the thesis and the future research lines that may continue what has been here proposed and tested. Jaime Ib´a˜nez Pereda 17 Chapter 2 EEG-based systems to study the motor function and BCI technologies in the neurorehabilitation field This chapter presents the theoretical basis supporting the starting points of the subsequent chapters with the four studies included in the present thesis. The first part of the chapter describes the main regions of the central nervous system that are involved in the generation of voluntary movements. After, the use of EEG technology in neurophysiological studies of the human motor cortex is justified according to its advantages as compared to other alternatives. Finally, the chapter presents the basic concepts regarding BCI systems and refers to some of the most important published studies up-to-date using BCI technology for rehabilitation purposes. Chapter 2. EEG-based systems to study the motor function and BCI technologies in the neurorehabilitation field 2.1 Structures of the central nervous system involved in the generation of volitional movements The adult human brain weights around 1.3 kg and it is comprised of around 1011 neurons with approximately 2 ∗1014 connection points between them. The neuron is the basic cell of the nervous system, and there exist at least over one thousand different kinds of neurons, although they all share a basic architecture. The complexity of human behaviour does not rely on the neural specificity, but on the ability of these cells to wire together, building highly precise anatomical circuits. Four main aspects of the nervous system are essential to understand its function: the mechanisms by which the neurons produce signals, the connection patterns between neurons, the relationship between this connection patterns and human behaviour and the means by which neurons and connections are modified through experience [Kandel et al., 2000]. Macroscopically, the central nervous system has seven main parts: spinal cord, medulla oblongata, pons, cerebellum, midbrain, diencephalon and cerebral hemispheres. The most important parts of the central nervous system, according to size and development, are the cerebral hemispheres, which consist of the cerebral cortex (outermost part of the brain formed by neural tissue) and three deep lying structures: the basal ganglia, the hippocampus and the amygdaloid nuclei. The two hemispheres of the human brain can be further divided into four different regions separated by the so-called cerebral sulci: the frontal lobe, the parietal lobe, the occipital lobe and the temporal lobe (see Fig. 2.1). Cortical neurons are highly interconnected, giving rise to human behavior through functions such as sensory processing, movement planning, preparation and execution, language processing, memory retrieval and many other cognitive functions. Among the different regions in the cerebral cortex, the primary motor and somatosensory cortices are located anterior and posterior to the Rolandic fissure, which divides the frontal and parietal lobes. Elaborating a motor strategy to perform a voluntary action is a complex task (see Fig. 2.1). The first step in a volitional movement is intent generation and planning. The prefrontal cortex is connected to many other cortical regions. This allows accessing to all required information for decision making and associated motor action. The prefrontal cortex receives information about past experience stored in memory from the temporal cortex. This information is consciously kept available, thus allowing the subject making inferences and predictions about the outcome of the tentative action. The parietal cortex, area 7, receives in turn information from vision-related cortical areas. This allows the planning of the action in space. In addition, area 5 of the parietal cortex can access 20 2.1 Structures of the central nervous system involved in the generation of volitional movements information about body situation in space. Considering this whole network, the subject is self-aware and perceives himself as an agent who can act in the environment, predicting the outcomes of his actions based on his past experience. Performing the planned action is still an even more complex process among interacting brain areas. Once the decision of performing an action is made, the upper motor controller (prefrontal cortex) let the fine-grained control to other frontal areas, such as the premotor and supplementary motor areas. Then, the primary motor area receives commands about the sequence of movements to perform the intended action and projects relevant movement commands directly to the motorneurons through the corticospinal tract and indirectly through the extrapyramidal system (specially through the rubrospinal tract for upperlimb movements). Descending projections to the muscles are also generated from other cortical regions such as the premotor and somatosensory cortices, thus building a complex model of motor control. The motor information travelling along these complex network defines the beginning and end of motor sequences, and on-line corrections. For this on-line correction, sensory information (visual, haptic, proprioceptive) is collected and introduced into the control model. Certain subcortical regions are also involved in the beginning of a motor sequence. These subcortical regions are the basal ganglia and the cerebellum. The basal ganglia consist of the striatum (caudate nucleus plus putamen), the globus pallidus, the subthalamic nucleus and the substantia nigra. The basal ganglia, and specially the striatum, receive information from the whole sensory and motor cortex. Therefore, the striatum integrates and overlaps the sensory and motor images of the self body. Those images split in the striatum, representing body parts in a redundant way. When relating pieces of sensory information to pieces of motor actions (such as muscle area activations), the striatum builds action plans coherent with the sensory signaling pattern. The internal globus pallidus inhibits paths that link the thalamus and the frontal cortex. The external globus pallidus and the subthalamic nucleus reinforce the globus pallidus activation. Therefore, the output path from the striatum is double. There is a direct path that inhibits the internal globus pallidus, thus allowing movement, and there is also an indirect path that activates the external globus pallidus and the subthalamic nucleus, which in turn reinforces the internal globus pallidus and blocks movement. Finally, the substantia nigra provides flexibility to motor plans by the connection to the striatum. If the motor plan has been successful then the substantia nigra does not activate the path to the striatum. Otherwise, if the motor plan failed the substantia nigra activates the path to the striatum. This produces the release of dopamine in the striatum, making it more plastic and sensitive to be modified and corrected. The cerebellum modulates movement and posture indirectly, adjusting the outputs of Jaime Ib´a˜nez Pereda 21 Chapter 2. EEG-based systems to study the motor function and BCI technologies in the neurorehabilitation field Supplementary Motor Area (SMA) Primary Motor Area (M1) Primary Somatosensory Area (S1) Premotor Area Substantia nigra Subthalamic nucleus Internal and external Globus Pallidus Putamen Caudate nucleus Ventral Thalamic Nuclei Frontal lobe Parietal lobe Temporal lobe Occipital lobe Striatum External Globus Pallidus Subthalamic Nucleus Substantia Nigra Internal Globus Pallidus Brain Stem Ventral Thalamic Nuclei Figure 2.1: Depiction of cortico-cortical (dashed lines) and cortico-subcortico-cortical (solid lines) information flow between brain regions in movement planning and execution. Brain stem projects fibers to muscles. Lines ending with arrows denote excitatory connections. Lines ending with dots denote inhibitory connections. the main encephalic motor structures. It acts as an online comparator between projected and performed movements, and it is therefore mainly involved in the on-line modulation of the movement, once it has already started. 2.2 The use of the EEG signal to extract cortical activity related to the movement There are a number of techniques that allow the acquisition of the cortical and subcortical electrophysiological activity. The main difference among these alternatives is the size and location of the neural population that is being “listened to”. This way, it is possible to measure •the spiking activity of single neurons; •the local field potentials reflecting the summation of nearby synaptic and neuronal activity; 22 2.2 The use of the EEG signal to extract cortical activity related to the movement •the electrical activity directly measured from the cerebral cortex; •the EEG activity, which is the cortical signal obtained from electrodes placed on the scalp; •the magnetoencephalographic (MEG) activity, acquiring the magnetic fields elicited by cortical dipole sources tangentially oriented to the head surface; From this list, only MEG and EEG systems are non-invasive solutions characterizing cortical processes. The main advantages of using the EEG signal are the fact that it is a non-invasive technique, commercially available, easy to set-up and robust to possible external interferences, which makes it perfectly suitable in clinical environments and with wearable robotic systems. The main handicap of the EEG signal is its low spatial resolution caused by the fact that cortical signals are acquired from electrodes on the scalp, a few millimetres away from the actual cortical surface. As previously defined, the EEG consists in the acquisition of the cortical electrical activity with scalp electrodes over specific points of the brain. In its origin, the EEG was developed to analyse mental processes, but its clinical applications rapidly followed. Cortical neurons are connected to thousands of other neurons through excitatory and inhibitory synapses, spreading throughout the dendritic part of the neuron. The transmission of an action potential from one neuron to the next one produces in the latter an excitatory or inhibitory postsynaptic potential. These potentials caused by ionic imbalances are summed in the dendritic bodies of the neurons, giving rise to field potentials in the nearby region. This process is assumed to be one of the main sources of the EEG activity acquired by the scalp electrodes [Niedermeyer and da Silva, 2005]. Therefore, the EEG activity will represent mental processes with sufficiently large enough groups of neurons spiking synchronously, so that the amplified version of their activity can reach the scalp. The EEG morphology depends on multiple factors such as the age, vigilance, performance of cognitive tasks, motor tasks, etcetera. Its similarity with a chaotic process and its small amplitude (10-100 µV), have definitive influence on the way these signals are analysed. As with other biological signals, the EEG activity presents a number of characteristics that make its analysis a complex task. The EEG signal is nonlinear and it is considered a stochastic and non-stationary process. The high variability present in an EEG signal is caused mainly by the noisy environment and the acquisition techniques used, the neurophysiological phenomena that produce the signal, the biological phenomena that appear in parallel with the studied process and that contribute to the recorded signal and the response of biological mechanisms to external agents. Jaime Ib´a˜nez Pereda 23 Chapter 2. EEG-based systems to study the motor function and BCI technologies in the neurorehabilitation field The EEG signal is traditionally described by means of its power spectrum, characterized by the presence of a number of cortical oscillations (cortical rhythms) associated with different frequency bands. The main rhythms of the brain are located in different frequency bands: the theta band (4-7 Hz), the delta band (1-3 Hz), the alpha band (8-12 Hz), the beta band (13-28 Hz) and the gamma band (29-100 Hz). In addition to these bands, the cortical changes with frequencies under 1 Hz (slow and ultra-slow rhythms) may also to be considered as a relevant source of information, specially in the analysis of movement-related potentials. 2.2.1 Cortical patterns related to the motor function and visible in the EEG signal Cortical patterns measured with EEG may be classified according to the nature of the stimuli that generate them: they may appear as a response to external stimuli or they may be electrical processes endogenously generated. Cortical patterns can also be classified according to their morphology and the components that form them. In this line, two sources of information may be distinguished: slow cortical changes and information contained in cortical rhythms. The information obtained from the processing of the cortical rhythms can in turn be acquired either by analysing the power changes in a specific cortical region, or by studying how two different regions interact through these oscillations. Attending to these classification criteria, cortical patterns associated with the voluntary movements may be classified in different groups, described in the forthcoming sections. 2.2.1.1 Movement related cortical potentials (MRCPs) Before and during self-initiated movements of healthy subjects, slow changes (with frequencies under 1 Hz) appear in the EEG activity. In most subjects these changes have amplitudes of few µV and therefore they are difficult to observe in the raw EEG signal. When averaging across a number of similar movements performed by a same subject, a sequence of defined temporal patterns with a specific spatial distributions can be observed. This sequence of patterns is termed movement-related cortical potential (MRCP) [Shibasaki and Hallett, 2006] and it constitutes one of the main sources of information to evaluate certain aspects of the mental activity before and during voluntary movements. Each of these MRCP components are found either before or after the onset of voluntary movements and are characterized by positive or negative deflections of the EEG signal. The MRCPs in self-initiated actions begin with a slow negative deflection of the EEG signal amplitude starting about 1.5 s before the onset of the movement, called the Bereitschaftspotential (BP). The BP precedes voluntary movements and it has proven to 24 2.3 BCI systems for motor rehabilitation brain activity to drive a device providing proprioceptive feedback. This sensory feedback is expected to induce plasticity leading to restoration of the normal motor control. This second strategy relies on the idea that brain activity can guide activity-dependent central nervous system plasticity in the same way as the standard repetitive movement practice carried out by therapists or robots influences it [V´arkuti et al., 2013]. The potential relevance of the second BCI-based strategy for changes in motor behaviour is exemplified particularly well in the context of stroke rehabilitation: assuming that the connection between peripheral muscles and the sensorimotor cortex has been disrupted due to a cortical or sub-cortical stroke, a concurrent activation of sensory feedback loops and primary motor cortex may reinforce previously silent cortical connections by Hebbian learning (repeatedly coincident activation of pre-synaptic and post-synaptic cells reinforces synaptic strength, tending to become associated) and thus support functional recovery [Mrachacz-Kersting et al., 2012; Niazi et al., 2012]. According to this arguments, as will be shown in subsequent studies, the fact that the EEG allows a precise location of the onsets of voluntary movements becomes a relevant aspect of this technology to be applied in neural rehabilitation interventions for stroke patients. Jaime Ib´a˜nez Pereda 31 Chapter 3 EEG-based predictive classification of analytical upper-limb movements 3.1 Abstract Chapter 2 has presented the most commonly used EEG patterns associated to motor processing functions (mainly the MRCPs and ERD/ERS patterns). Yet, the analysis of the characteristics and dynamics of the cortical rhythms originated from distributed points in the sensorimotor cortex and measured with EEG may allow an advanced characterization of how different motor-related cortical regions activate or deactivate when performing different kinds of motor actions. In this regard, one of the main limitations of EEG systems to characterize task-related cortical processes is their low spatial resolution. This limitation reduces the possibilities of distinguishing among mental states that present similar somatotopic representations. On the other hand, as has been commented before, EEG systems present a great potential to characterize relatively simple mental states preceding the onset of volitional movements. So far, the majority of BCI systems that have been proposed to classify different movement-related mental states have frequently presented paradigms in which movements of distant parts of the body (and therefore, with distant somatotopic representations) were to be distinguished (examples in this line are BCI systems distinguishing between movement imagery of the right hand, the left hand and the feet proposed by several BCI groups). In this chapter it is studied the possibility of classifying a number of simple movements, all of them performed with the same limb, based on premovement EEG signal segments. To do so, advanced data mining techniques are applied on a dataset with a large number of examples to find the optimal subset of features that allow a differentiation of classes over the chance level of the study. The scientific interest of experiments like this one in neurorehabilitation environments ranges from further understanding the cortical mechanisms underlying the generation of simple movements, to achieving new EEG processing techniques that can be integrated in rehabilitation BCI systems to test the pa- Chapter 3. EEG-based predictive classification of analytical upper-limb movements tients’ involvement in the rehabilitation process and to provide an adequate proprioceptive feedback associated to the movements they intend to do. 34 3.2 Introduction 3.2 Introduction The electroencephalographic (EEG) activity allows the description of cortical processes associated to volitional motor actions [Chatrian et al., 1959; Kornhuber and Deecke, 1965; Pfurtscheller and da Silva, 1999; Libet et al., 1982]. A number of studies with EEG have demonstrated its potential use to locate intervals of motor-related cortical activation and deactivation [Neuper et al., 2006; Pfurtscheller and Solis-Escalante, 2009], to anticipate the instants at which voluntary movements begin [Bai et al., 2011; Niazi et al., 2011; Ib´a˜nez et al., 2013], to decode movement parameters such as velocity, strength, etc [Gu et al., 2009a], and to distinguish between different classes of movements [Morash et al., 2008; Pfurtscheller et al., 2006]. Yet, it remains unclear the extent to which the EEG activity allows the description of motor-related mental processes. Advances in this area will lead to further understanding the relevant parts of the brain taking part in the generation of volitional actions [Desmurget et al., 2009; Obhi et al., 2009], and to new ways of inducing neural rehabilitation by integrating EEG in novel clinical interventions [Buch et al., 2008; Daly and Wolpaw, 2008], either for passive monitoring the motor therapy, or for active mobilization with robotic devices. In this context, EEG technology is of great interest since it allows the real-time characterization of the motor-related cortical activity to obtain predictive information regarding intended actions. Such information has proven to be valuable to provide natural proprioceptive feedback inducing cortical plasticity [MrachaczKersting et al., 2012; Niazi et al., 2012]. Recent studies have proposed methodologies to decode 3D kinematics of the upper-limb based on slow potentials measured with EEG [Bradberry et al., 2010]. Nonetheless, metrics applied to validate the results in these studies are subject of discussion [Antelis et al., 2013]. Previous studies using invasive recordings have pointed out that brain-machine interfaces (BMIs) based on the dynamics like those of muscles seem to be more robust and easier to learn than BMIs commanding forces or movements in external coordinates [Oby et al., 2013]. Several works have taken advantage of the changes of the cortical rhythms measured with EEG to estimate muscle activations and joint rotations [Morash et al., 2008; Pfurtscheller et al., 2006]. In[Deng et al., 2005] it was proposed a methodology to distinguish between movements performed with the shoulder and elbow of the dominant upper-limb. These two tasks present similar cortical representations, which makes them difficult to be distinguished from each other based on non-invasive recordings as the EEG activity. No previous works have tried to identify EEG-signal patterns classifying more than two different movements performed with the same arm. Jaime Ib´a˜nez Pereda 35 Chapter 3. EEG-based predictive classification of analytical upper-limb movements In this chapter, results of a classifier of analytic movements performed with the upperlimb (7 different classes) based on pre-movement EEG activity are presented. The system is evaluated on 6 participants who performed 350 self-initiated movements during the experiments. To develop the classifier, data mining techniques extracting optimal features selected with a genetic algorithm were applied. The feature space considered were the power spectral values of the alpha and beta bands of the EEG signal (information of the activation or deactivation of cortical regions associated to movement tasks [Pfurtscheller and da Silva, 1999; Deng et al., 2005; Morash et al., 2008]). The average accuracy obtained with all subjects was above the chance accuracy level obtained by randomly labelling the acquired examples. Further analyses (discussed in the last part of the chapter) discard the hypothesis that other sources of information, different from the task-related cortical activity, were used to reach the classification results. The study supports the idea that EEG can supply with predictive information about upper-limb analytical movements. 3.3 Methods 3.3.1 Participants and experimental procedure Six healthy male subjects, right-handed and with ages between 25 and 35 years-old were recruited for the experiments carried out in this study. They were seated in a comfortable chair and, during the exercises, they were asked to remain relaxed without performing any movements other than the tasks studied in the experiments. A screen was placed in front of the participants to guide them during the experiments. Each subject performed seven analytic movement tasks with the dominant upper-limb: shoulder abduction (SA), shoulder extension (SE), shoulder rotation (SR), elbow extension (EE), forearm pronation (FP), wrist extension (WE) and wrist rotation (WR). For each one of these tasks, two runs of 25 trials each were executed, leading to 350 trials (7 tasks and 50 examples per task). Each trial was divided into two parts: during the first part of 12 s, the participants were asked to start a single movement when they wanted, trying to wait at least 2 s before performing it (during this part the word “Movement” was shown in the screen). The second part of the trial lasted 3 s and the participants relaxed and got prepared for the subsequent trial while the word “Rest” was shown in the screen. Each run lasted 6 minutes and 15 seconds in which the participants performed self-initiated movements (trials) of one of the seven tasks. The runs were interleaved as follows: the first type of analytical movement was performed in runs 1 and 8, the second type in runs 2 and 9, etc. (see Fig. 3.1). A session lasted around 2 hours. Participants adopted three different starting positions of the arm to perform the move36 3.3 Methods Rest Self-initiated movement 3s 12s Trial #1 Trial #2 Trial #N Trial #25 … … Run #1 (Mov. Type 1) Run #2 (Mov. type 2) rest … Run #7 (Mov. Type 7) Run #8 (Mov. Type 1) rest … Run #14 (Mov. Type 7) Figure 3.1: Scheme of the recording sessions. A trial (top), a run (middle) and the distribution of tasks along the session (down) are represented. ments (see Fig. 3.3): A) the arm was left hanging and relaxed for tasks SA and SE, B) the arm was resting on the arm of the chair for tasks EE, FP and WE, and C) the arm was resting on an auxiliary desk for tasks SR and WR. EMG/IMUs Sensors Stimulus presentation EEG recording B) C) A) Figure 3.2: Schematic representation of the three positions adopted by the participants to perform the analytical movements. Jaime Ib´a˜nez Pereda 37 Chapter 3. EEG-based predictive classification of analytical upper-limb movements 3.3.2 Data Acquisition Three synchronized gUSBamp amplifiers (g.Tec gmbh, Graz, Austria) were used to amplify and digitize the EEG and EMG data at a sampling frequency of 512 Hz. The EEG montage consisted of 32 electrode positions (see Fig. 3.4) and active Ag/AgCl scalp electrodes were used. The ground and reference electrodes were placed on FPz and on the left earlobe, respectively. EMG activity was recorded with bipolar derivations on 8 muscular groups: extensor digitorum, extensor carpi ulnaris, palmaris longus, biceps brachii, triceps brachii, frontal part of the deltoid, lateral part of the deltoid and back part of the deltoid. Gyroscopes were placed on the third metacarpal, the edge of the forearm (dorsal side), and above the olecranon process. The gyroscopic data were digitized at 50 Hz and synchronized with the EEG and EMG data by means of an external digital signal. 3.3.3 Data processing and classifier design This section describes the methodology for building, for each participant, a classifier of the 7 possible analytical movements performed with the upper-limb during the measurements. 3.3.3.1 Detection of the movements’ onsets The onsets of the movements were obtained from the gyroscopes data as follows: the data were low-pass filtered (Butterworth, order 2, ≤6 Hz) and the rotation angle of each joint moved was obtained as the absolute value of the difference between the gyroscope measurements of the two adjacent sensors (the hand and forearm for wrist movements, the forearm and arm for elbow movements and the arm and trunk for shoulder movements). The threshold for the detection of the onset was set at 5 % of the maximum rotation speed of all movements of each type. The gyroscopes information was used to detect the onsets instead of the EMG because it was more robust for all movements with all the three possible initial positions. Notice that the latency between the EMG-based and the gyroscopes-based onset detections is expected to be small, given that the electromechanical delay for upper-limb tasks is in the order of tens of milliseconds [Norman and Komi, 1979]. EMG data served to assert that the onsets of the movements detected with the gyroscopes were correctly located, and that there was no muscular activity in the different initial positions during the resting intervals before the movements. 3.3.3.2 EEG signal processing and feature extraction 38 3.3 Methods Small Laplacian filtering [Hjorth, 1975] was applied to the EEG channels that were surrounded by 4 neighbouring positions and a Common Average de-referentiation was applied to the boundary positions of the used electrodes set-up. The EEG data were band-pass filtered (Butterworth, order 2, passed band 5 - 45 Hz). Each trial was segmented in the following time intervals: i) a 2-seconds segment starting 2 s before the movement onset (referred to as “Whole”), ii) a 1-second window starting 2 s before the movement onset (“Early”), and iii) a 1-second window starting 1 s before the movement onset (“Late”). The features from these three windows were used in combination by the classifier. The “Whole” window was expected to supply the classifier with global and low-variance information of the cortical activity related to the voluntary movement, while the “Early” and “Late” windows were expected to provide information regarding transitory mental processes before the voluntary movement initiation. For each of these segments, the Power Spectral Density (PSD) values of the EEG signal of each channel were obtained in the frequency range from 7 - 30 Hz (alpha and beta bands), with a frequency resolution of 1 Hz (Welch’s method with Hamming windowing, 75 % overlap, no zero-padding). Therefore, 23 power values were extracted per window, channel and trial, leading to 2208 features extracted per trial. The logarithms of the PSD values extracted were computed as the features fed to the data mining process, aimed to construct the EEG-based classifier of analytic movements performed. The logarithmic power values were used to convert the extracted features into normal distributions. 3.3.3.3 Classifier implementation Firstly, the feature space was reduced eliminating features correlated ≥0.75 in the training dataset. Feature selection was performed using a genetic algorithm that maximized the accuracy of a Bayesian classifier of independent features (the scheme is presented in Fig. 3.3). The algorithm was programmed to run 1000 generations, with 500 new individuals generated in each generation. The number of features of each the individual was set to be between 50 and 100. A 4-fold cross-validation was used to evaluate the classifier’s performance for each individual in each generation, avoiding singular solutions of the classification problem. In the classification stage, a Bayesian Classifier of independent features with Gaussian modelling was also selected to classify the examples because it showed better performance than neural networks and support vector machines with the data of these experiments. Moreover, similar studies have also obtained optimal results with Bayesian classification methods [Bai et al., 2007]. A 4-fold cross-validation was used to obtain the classification Jaime Ib´a˜nez Pereda 39 Chapter 3. EEG-based predictive classification of analytical upper-limb movements results. EEG CAR/ LAP PSD Feature filtering Feature subsets Cross-validation Bayesian Classifier Training (75%) Test (25%) X-val. Avg. Results Classif. Results Best Avg. Results Genetic Algorithm Nr. of generations Population (crossover mutation) Figure 3.3: Schematic representation of the three positions adopted by the participants to perform the analytical movements. 3.3.4 Additional experiments to prove the validity of the classification results Three additional experiments aimed to further validate the classification results were performed and are described here. 3.3.4.1 Validation experiment 1: Estimation of the significance of the obtained class description The experiment was performed to obtain a referential chance level, in order to compare it with the accuracy results obtained with the classifier of analytical tasks. To get the chance level, the following process was repeated 10 times for each subject: firstly, the labels of the examples in the dataset were reorganized randomly, and secondly, the new dataset was applied the classification procedure detailed in 3.3.3.3. Mean ±SD of the accuracy results over the 10 repetitions was computed for each participant and compared 40 3.5 Discussion and initial positions was less than 20 %. This suggests that a large portion of different information is used by the classifiers. Notice that although this could also be due to the fact that similar but not equal features are being selected, the feature space reduction performed before the feature selection process discarded features with high correlations. Therefore it is hypothesized that the first alternative (the classifiers of tasks and initial positions are different) is more likely. Secondly, the EEG signal cannot be considered stationary along long measurement intervals due to the changes in the system’s set-up (as for example changes in the electrodes impedances due to deterioration of the conductive gel) [Shenoy et al., 2006] or to fluctuations in the patients’ vigilance or involvement [Blankertz, 2008]. Therefore, variations in the features extracted from the EEG signal during the measurement sessions are expected. Nevertheless, in these experiments it is unlikely that this phenomenon is introducing any bias in the classification results, since the experimental design alternated the types of movements performed in consecutive runs (the two runs of the same movement were separated by one run of each of the rest of the movements). In addition, a 4-fold cross-validation was used, which separated the data in 4 randomly generated testing groups. It is therefore expected that, in general terms, training and validation subsets were randomly collected from session intervals all along the whole experiment, and the effects of the EEG non-stationarities were marginal. Thirdly, it cannot be asserted that the developed classifier of tasks is using cortical activity directly involved in the motor actions. However, it can be analysed whether certain characteristics of the classifier’s design and of its behavior fit what may be expected from a neurophysiological point of view. On the one hand, a higher number of features were selected from EEG channels around the contralateral rolandic fissure (i.e. from the motor and somatosensory areas of the cortex) in four out of the six subjects. This was specially observed with participants 03 and 06, who also returned the best classification results, whereas with participant 04 the classification results were poorer and the selected featuers presented a different distribution over the scalp. This is in line with previous studies regarding the spatial distribution of the cortical rhythms associated with voluntary motor activities [Bai et al., 2005; Desmurget et al., 2009; Pfurtscheller and da Silva, 1999; Pfurtscheller et al., 2003; Urbano et al., 1996]. In addition to this, the number of features in the upper-beta bands was higher than in other bands, which is in line with studies showing that these rhythms are involved in the preparation of voluntary movements [Salmelin et al., 1995; Morash et al., 2008; Engel and Fries, 2010]. Besides, given that they present higher frequencies than the mu rhythm, they are associated with smaller neural associations [Buzs´aki and Draguhn, 2004], therefore allowing for a finer description of the cortical representation associated with the performed task. This is in turn desired in the Jaime Ib´a˜nez Pereda 47 Chapter 3. EEG-based predictive classification of analytical upper-limb movements present study of classifying among tasks with similar cortical representations. Moreover, a temporal dependency of the classification results on the location of the time segments for feature extraction was observed (see Section 3.4.3). The segment starting -1.5 s before the onset of the movement returned the best classification results for all subjects and the segment starting at -3 s with respect to the onset returned worse accuracies than the other two conditions. This is in line with what is documented about EEG activity associated to voluntary actions: the first changes in the signal start around 2 s before the movement becomes apparent and the significance of these changes is greatest with the beginning of the movement [Pfurtscheller and da Silva, 1999; Bai et al., 2005; Pfurtscheller et al., 2003]. Finally, it may be argued that the dataset of examples is so small (50 trials per class in the case of classifying analytic movements) that suboptimal solutions are obtained with the genetic algorithm, and that these solutions only adapt the measured examples of the dataset, but would fail generalizing to new unseen data (i.e. overfitting). Nevertheless, given that a Bayesian Classifier has been used in combination with a cross-validation methodology, suboptimal classification solutions are highly unlikely. Furthermore, the results obtained with the original dataset outperform the classification results with the randomly labelled dataset. In summary, the posterior analysis of the methods used and results obtained here demonstrates that the motor-related cortical activity associated to the execution of voluntary movements performed with a single limb can be characterized to a certain degree. Nevertheless, since these results have been obtained with an offline analysis of the data acquired in a single session with each participant, it still needs to be studied whether the performance of the developed classifiers remain stable along different sessions. Furthermore, variations of the classifier’s performance were observed when features were extracted from time segments at different locations with respect to the actual movements. Higher accuracies were obtained when the initial part of the apparent movements (the first 500 ms) were considered, probably suggesting either an influence of the somatosensory information in the classification results, or an increased activation of the primary motor cortex and other cortical areas related to movement execution, once the movement starts. Future experiments also including motor imagery tasks of the analytical movements may help to gain knowledge in this regard. Finally, a system capable of decoding the cortical activity related to the kind of upperlimb movements can be of interest for neural rehabilitation protocols [Daly and Wolpaw, 2008]. The system proposed could serve as a monitoring tool of the patient’s involvement in the rehabilitation task or it could also be included as an additional input to a controller of assistive robotic devices. Therefore, analogous studies need also to be performed with patients presenting altered cortical activity due to lesions in the nervous systems [Wiese 48 3.6 Chapter conclusions et al., 2004; Stepien et al., 2010; Serrien et al., 2004], to test the reliability of the classifier in such conditions. 3.6 Chapter conclusions It has been proposed a classifier of self-paced analytical movements performed with the upper-limb and based on premovement EEG information. An average accuracy of 62.9 ± 7.5 % has been reached in the classification of the seven analytical movements performed with the dominant arm, which was above the chance level (30.2 ±4.3 %). Several tests have been performed to discard the hypothesis that the information used by the classifiers could come from different sources than the cortical activity. This chapter has described an innovative experimental procedure regarding the decoding of mental states related to 7 different movement actions performed with a single limb. It is therefore a step forward in the use of EEG signals to model cortical patterns related to planning and execution of movements and it is expected to improve future BCI systems aimed to respond in close association with users’ intentions to move. Jaime Ib´a˜nez Pereda 49 Chapter 4 Study of alprazolam-induced changes in cortical oscillations and tremors of patients with ET 4.1 Abstract In the first chapters of this thesis it was emphasised the potential capacity of EEG systems to acquire, with high temporal resolution, cortical processes associated with sensorimotor states, and it was also indicated that current electrophysiological systems (such as EEG and EMG devices) allow the concurrent measurement of neuronal information from different body regions and with negligible synchronization errors. These advances open a door to studies of interaction between central (cortical) and peripheral (muscular) neural activity. While in Chapter 3 the distribution of cortical rhythms associated to the execution of different voluntary movements was characterized , in this chapter it is presented a study of the effects of a clinically used drug (alprazolam) on pathological (involuntary) tremors and cortical oscillations of patients with ET. The study analyses tremor changes after alprazolam intake and how they are related to other changes in the cortical activity. This chapter is therefore aimed to provide new insights about the mechanisms of tremor generation in ET and to propose a novel application of EEG systems to analyze the effects induced by a drug in patients with tremor. Chapter 4. Study of alprazolam-induced changes in cortical oscillations and tremors of patients with ET 4.2 Introduction ET is a neurological disease characterized by postural and action tremor of the arms with a frequency of 4-12 Hz [Benito-Le´on and Louis, 2006]. Although it is the most prevalent movement disorder [Louis et al., 1998; Thanvi et al., 2006; Benito-Le´on et al., 2003], and constitutes one of the most common neurological disorders among adults [Benito-Le´on et al., 2003, 2005], the exact mechanisms of tremor generation in ET are still unknown [Elble and Deuschl, 2009; Louis et al., 2013]. A number of studies using different brain imaging techniques point to a neuronal loop involving cerebello-thalamocortical pathways as the structures involved in the generation of the pathological tremor-related neural activity [Benito-Le´on et al., 2009; Hua et al., 1998; Hellwig et al., 2001; Raethjen et al., 2007; Raethjen and Deuschl, 2012; Schnitzler et al., 2009]. In particular, studies of coherence between the cortical activity, measured with electroencephalography (EEG), and muscle activation, measured with electromyography (EMG), have demonstrated the implication of cortical structures in the pathological neural network [Hellwig et al., 2001], and have even allowed to postulate how such interaction may change over time [Raethjen et al., 2007]. ET is commonly treated either with neurosurgery or with drugs. However 50 % of the ET population does not benefit from any of the available alternatives [Deuschl et al., 2011]. All the pharmacological treatments for ET were discovered by chance [Deuschl et al., 2011] and are still limited and only partly effective [Benito-Le´on and Louis, 2006, 2011]. The action mechanisms of these drugs are barely understood, although it is assumed that they attenuate tremor by interfering with the widespread pathological oscillations occurring throughout the motor system [Deuschl et al., 2011]. Among the pharmacological alternatives to treat ET, alprazolam is a short-acting benzodiazepine accepted by the Quality Standards Subcommittee of the American Academy of Neurology as a probably efficacious (level B) agent [Zesiewicz et al., 2011]. Two studies using clinical rating scales found that alprazolam reduced the limb tremor in a 2-4 week monotherapy trial [Gunal et al., 2000; Huber and Paulson, 1988]. Nevertheless, its use is recommended in patients who require only intermittent therapy, due to its abuse potential, and to the risks of developing tolerance [Huber and Paulson, 1988]. As in the case of other pharmacological treatments for ET, the way in which alprazolam alleviates tremor is unknown. Previous studies with healthy subjects have observed an increased activity in the cortical beta rhythms (around 13-30 Hz) after benzodiazepine intake [Baker and Baker, 2002; Hall et al., 2010; Jensen et al., 2005; Lindhardt et al., 2001]. It is known that benzo52 4.3 Methods diazepines increase the affinity of the λ–aminobutyric acid (GABA)-A receptor toward its neurotransmitter, increasing the size of the inhibitory postsynaptic potentials that it generates [Connors et al., 1988]. However, it is not intuitive how enhancing inhibition increases the power of the beta and gamma rhythms, and why such increase is observed in the somatosensory cortex [Minc et al., 2010; Hall et al., 2010]. In this regard, it has been proposed that mutual inhibition between interneurons, and the reciprocal loop between excitatory and inhibitory cells provide two general mechanisms for rhythmogenesis, especially for fast cortical oscillations [Wang, 2010]. Whether the expected changes in the cortical beta activity of ET patients after alprazolam intake are part of the neural process that alleviates tremor or they rather represent a side effect in the ET treatment with alprazolam is an open question. Since voluntary motor commands are projected to the targeted motor unit populations at the beta band [Conway et al., 1995; Kilner et al., 2000; Petersen et al., 2012], it is hypothesized that the increase in the cortical beta activity due to benzodiazepines alters the transmission of descending motor commands. It is further expected that such an increase of oscillatory beta activity in turn impedes the appearance of pathological tremor-related cortical activity. Therefore, this study analyses the interplay between the cortical activity in the beta band and in the tremor frequency after alprazolam intake, and how this interaction is associated with the drug effects on the tremor and the cortico-muscular coupling at the tremor frequency. 4.3 Methods 4.3.1 Patients, data acquisition and experimental procedure Eight patients (two female, age 64.1 ±13.2 years; mean ±SD) were included from a general neurology outpatient clinic (details in Table 4.1). All of them had been diagnosed as ET according to the Movement Disorders Society Diagnostic Criteria [Deuschl et al., 1998]. Patients with severe tremor at the hands or the head were excluded from the study to avoid interferences with the recordings. None of the patients had any other neurological condition apart from ET, or suffered from psychiatric disorders. None of them was taking medication to treat their tremor, or any other drugs that could alter it. Wrist tremor at the most affected side was measured with solid-state gyroscopes and surface EMG. Two gyroscopes, placed on the hand dorsum and the distal third of the forearm, measured wrist tremor by computing the difference between them [Gallego et al., 2010]. The data were sampled at 50 Hz. Surface EMG was recorded using a grid of 13 X 5 electrodes (1 missing electrode), Jaime Ib´a˜nez Pereda 53 Chapter 4. Study of alprazolam-induced changes in cortical oscillations and tremors of patients with ET Patient 01 02 03 04 05 06 07 08 09 Gender Male Female Male Female Male Male Male Male Male Age (years) 76 80 44 63 45 65 77 69 58 ET family history Y Y Y Y Y Y Y N N Disease duration (y) 5 32 15 7 4 10 2 3 4 Dominant side of tremor L R R L R L L L R EMG tremor freq. (Hz) 6.2 5.2 7.0 6.2 - 6.2 7.0 6.2 8.2 Leg tremor N Y N Y N N N N N Head tremor N N N Y Y Y N N N ETRS 45 32 17 38 18 15 14 16 22 Table 4.1: Main baseline demographic and clinical variables. Fahn, Tolosa, Marin Essential Tremor Rating Scale (ETRS). with 8 mm inter-electrode distance. The electrode grid was placed on the wrist extensors, centred on the muscle exhibiting the clearest tremorogenic activity; the common reference was set to the wrist using a humidified bracelet. The data were amplified, band-pass filtered (10-750 Hz), and sampled at 2.048 Hz. EEG signals were recorded from 16 positions (F2, F4, FCz, FC2, FC4, FC6, Cz, C2, C4, C6, T8, CP2, CP4, CP6, Pz, and P4, according to the International 10-20 system, when the left arm was recorded; the symmetric positions were employed when the right arm was recorded) using passive Au electrodes. The cortical activity at the contralateral hemisphere was recorded because it is where significant cortico-muscular coherence at the tremor frequency [Hellwig et al., 2001, 2003; Raethjen et al., 2007; Timmermann et al., 2002], and the beta band [Conway et al., 1995; Negro and Farina, 2011], is best observed. The reference was set to the common voltage of the two earlobes. AFz was used as ground. The signal was amplified, band-pass (0.5-60 Hz) and notch filtered (50 Hz), and sampled at 256 Hz. The recording systems were synchronized with a common digital signal. The study was performed in a sound and light-attenuated room. Patients sat in a comfortable chair with the arms supported. During the measurements, they were asked to remain relaxed, keeping their eyes open and fixing their gaze on a point in the wall. Patients were instructed not to eat or drink anything (water was allowed) from 2 h before the recordings. In order to evaluate the effects of alprazolam, patients were measured during four 4-min runs, as follows: before the administration of alprazolam (Run0), immediately after it (Run1), 40 min after it (Run2), and 80 min after it (Run3). Postural tremor was elicited by asking patients to hold the measured hand outstretched, with palms down, and parallel to the ground. In patients who exhibited a very mild tremor before the experiments (patients 02 and 04), weight loads of 0.5 Kg were attached to the hand to enhance it [Hellwig et al., 2001; Raethjen et al., 2007]. 54 4.3 Methods A single dose of 0.50 mg of alprazolam was administered to patients who weighed less than 75 kg; the rest (5 patients) received a single dose of 0.75 mg. No patient reported adverse effects. Two patients were discarded due to technical problems with EEG acquisition (patient 03) and to the absence of tremor during the measurement session (patient 05), respectively. 4.3.2 Data processing and analysis The EEG signals were spatially filtered using Hjorth transform [Hjorth, 1975]. The resultant channels (FC2, FC4, C2, C4, C6, CP2 and CP4) were used in the subsequent analyses. Artefacts were removed based on visual inspection. After examination of the amplitude spectra of the gyroscope and EMG data, the defined tremor frequency range for the group of patients was 4-9 Hz (see Fig. 4.1). This range was used to estimate both the tremor power (measured with gyroscopes and EMG), and the power of the tremor-related cortical activity. To select the surface EMG channel that best characterized the tremor, the criterion of maximizing the signal-to-noise ratio (SNR) of the tremor component of the EMG signal was used. To select the surface EMG channel that best characterized the tremor, the criterion of maximizing the signal-to-noise ratio (SNR) of the tremor component of the EMG signal was used. This value was defined as the ratio of the integral of the power spectral density (PSD) within the tremor frequency range, to the integral of the PSD of the rest of the signal, similarly to [Hellwig et al., 2001]. This channel was used throughout the whole analysis. The percentage of tremor reduction between the first and last runs (Run0 and Run3) was computed by analyzing the gyroscope data. Tremor severity was defined as the integral of the PSD of the signal in the tremor frequency range. Before, the data were band-pass filtered (2-15 Hz) to extract the tremor [Gallego et al., 2010]. It was also calculated how the neural drive to the muscles related to tremor was reduced after alprazolam intake by computing, with the EMG data, the percentage of tremor power decrease in Run1, Run2 and Run3 with respect to Run0. Cortico-muscular coherence was computed to assess how the tremor-related cortical drive to the muscle varied due to the effect of alprazolam. The coherence between all the processed EEG channels and the rectified EMG at the electrode previously selected was calculated [Farina et al., 2013], and the EEG channel exhibiting the largest coherence peak at the tremor frequency was chosen for subsequent calculations. It was used the method for coherence estimation proposed in [Halliday et al., 1995]: the signals were divided into epochs of 1 s, and their individual spectra and cross-spectra were computed (Hanning Jaime Ib´a˜nez Pereda 55 Chapter 4. Study of alprazolam-induced changes in cortical oscillations and tremors of patients with ET window of 1 s and 0.125 Hz resolution, achieved with zero-padding). The coherence |Rxy(λ)|2was estimated as |Rxy(λ)|2=|Cxy(λ)|2 Cxx(λ)Cyy(λ) with |Cxy(λ)|2being the magnitude squared cross-spectrum, and Cxx(λ) and Cyy(λ) the individual power spectra [Halliday et al., 1995; Hellwig et al., 2001]. The confidence limit was obtained as: 1−(1 −α 100) 1 N−1 where N is the number of epochs used to calculate the coherence and αis the significance level [Rosenberg et al., 1989]. To study how alprazolam affected the tremor-related cortical activity and the cortical activity in the beta band, the changes in the EEG spectra were assessed by calculating the integral of the PSD at the selected channel in the tremor frequency range (4-9 Hz, see above) and in the beta band (13-30 Hz). 4.3.3 Statistical analysis The Wilcoxon rank sum test was used to compare the tremor severity measured by the gyroscopes before (Run0) and 75 min after the administration of alprazolam (Run3). The Kruskal-Wallis test was used to compare the tremor-related neural drive to the muscle in Run3, Run2 and Run1 with respect to Run0. Significant differences between pairs of data were assessed with the Games-Howell test assuming non-equal variances. The same test was used to compare the changes of the power of the cortical activity in the beta band and in the tremor-frequency range, and to compare the changes of the ratio between the activity in these two bands; in all cases changes in Run3, Run2 and Run1 were obtained with respect to Run0. Finally the Spearman’s rank correlation was calculated between the decrease in tremor severity (in terms of neural drive to the muscle, i.e. EMG) and the changes of the ratio between the EEG activity in the beta band and in the tremor frequency range, using the data of Run3, Run2 and Run1 with respect to Run0, to investigate whether both phenomena were related. Results are reported as mean ±SD, and considered significant if P<0.05. 56 4.5 Discussion contractions to hold their hands extended, which explains the lack of meaningful results in this regard [Baker and Baker, 2012; Chakarov et al., 2009]. It was observed that, during a period after alprazolam intake, there was a significant increase in cortical beta activity in all ET patients, similarly to what was previously reported in studies addressing the effects of other benzodiazepines in healthy subjects [Baker and Baker, 2002; Jensen et al., 2005]. Interestingly, the power of the EEG activity at the beta band is also enhanced in alcoholics [Rangaswamy et al., 2002] or after a small single dose of alcohol [Ilan and Gevins, 2001], and in 50-90 % of ET patients alcohol acts by reducing tremor amplitude [Growdon et al., 1975; Zesiewicz et al., 2011]. While the precise mode of action of ethanol in ET has not been established [Boecker et al., 1996], its principal effect is likely produced via the potentiation of GABA-A receptors [Wallner et al., 2003]. Although the effect of these substances increasing the beta power measured with EEG could not be related to its antitremorogenic effects, a significant relationship between the increase of the ratio between the beta and tremor-related cortical activity and the reduction of the contralateral postural tremor has been observed. Indeed, this dependency was seen for each patient individually (see Fig. 4.5). It is acknowledged that this finding could be an epiphenomenon or the consequence, albeit not necessarily direct, of the biochemical effect of alprazolam upon the brain. Nevertheless, considering that during maintained motor contraction the cortical motor areas and the muscles are synchronized in beta-range [Baker, 2007; Brown, 2000; Conway et al., 1995; Halliday et al., 1998; Kilner et al., 2000], it is hypothesized that the increased physiological beta activity at the primary motor cortex may be partially interfering the coupling of pathological oscillatory networks involved in the generation of tremor in ET. The main structures in the central nervous system believed to be involved in the generation of the tremor in ET [Boecker et al., 1996; Jenkins et al., 1993; Park et al., 2010; Wilms et al., 1999] are controlled by GABAergic connections. Due to this reason, it is noted that the GABAergic effect of alprazolam would be not only limited to the sensorimotor cortex, but could be spread through other subcortical structures. Indeed, localized microinjections of the GABA-A agonist muscimol into the ventral intermediate nucleus (in areas where tremor-synchronous cells were identified electrophysiologically) of ET patients undergoing stereotaxy, were effective in reducing tremor [Pahapill et al., 1999]. This study presents some limitations, but their impact on the conclusions is expected to be minor. Firstly, a small group of patients was recruited for the experiments, and thus the obtained results might not be generalized to population dwelling ET cases. However, the homogeneity of the results obtained in all the patients (see Fig. 4.6), reinforce the hypothesis proposed in this study. Second, no placebo group was measured. However, Jaime Ib´a˜nez Pereda 63 Chapter 4. Study of alprazolam-induced changes in cortical oscillations and tremors of patients with ET the obtained data does not indicate that any of the patients experimented placebo effects, given that the observed reduction of tremor severity 4 min after alprazolam intake was negligible compared to subsequent runs (see Fig. 4.1). It is considered that the results were not influenced by expectancy bias since patients had never been previously treated with alprazolam, and did not know how and when the drug could improve the tremor. 4.6 Chapter conclusions It has been shown that alprazolam attenuates tremor in ET at the same time that it increases the ratio between the beta and the tremor-related cortical activity, and decreases the strength of cortico-muscular coupling at the tremor frequency. It is hypothesized that the increase in the cortical beta activity due to the effects of alprazolam acts as a blocking mechanism of the pathological neural networks, which in turn helps reducing the tremor in ET. This is the first study of the neurophysiological changes occurring in ET patients after the intake of a drug used to alleviate the tremor, and it is expected that further experiments with other drugs reducing the tremor in ET will help understanding the pathophysiology of this disease and its response to the different treatments. The study represents an example of possible clinical applications of EEG technology to characterize and/or monitorize the effects of certain pharmacological treatments in patients with neurological disorders affecting their motor capacity. 64 Chapter 5 Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor 5.1 Abstract It has been previously shown that cortical changes can be observed in the EEG activity when movement tasks are performed, and that these changes may appear up to 1.5-2 s before voluntary movements are initiated. Chapter 4 also showed that patients with tremor may present altered cortical activity at certain frequency bands as a result of the existing tremor. In this chapter, an EEG-based design predicting voluntary movements and integrated with other sources of information also related to the execution of motor tasks is presented. The ultimate goal is to build up a multimodal BCI system managing pathological tremors. In this multimodal interface, anticipated information regarding intended motor actions is extracted from the EEG signals and supplied to other subsystems (based on EMG and gyroscopic signals) to finely track and cancel the tremor superimposed on the voluntary movement. Results of two experiments are presented in the chapter. The first experiment is aimed to validate the EEG system anticipating voluntary movements with healthy subjects and patients with essential tremor, and to compare an adaptive and a fixed design of the system. The second experiment in the chapter is aimed to validate the idea of a multimodal interface integrating EEG data with other sources of movement information. In this second case, results are given for a group of patients and the main objective is to study the advantages of a multimodal system integrating EEG information with EMG and gyroscopic data. The proposed system represents the first approach to EEG-based systems anticipating voluntary movements under a fully asynchronous and continuously evaluated paradigm, and it also represents the first BCI application for patients with tremors and the Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor first time that EEG and EMG signals are fused to improve the performance of a system tracking motor tasks. 66 5.2 Introduction 5.2 Introduction Multimodal Human-Robot Interfaces (mHRI) for motor compensation take advantage of complementary sources of information to drive external devices. In such applications, a major goal is to provide the patient with a communication channel that behaves in a natural way. A natural human-robot interface controlling movement compensation devices aims at reducing the impact of the technology on the user, and to do so it must meet three objectives: 1) the system must reliably distinguish the user’s intentions to move from the periods of non-intended activity (when the controlled device is in an idle state), 2) it must react with minimum latency with respect to the user’s intentions to move, and 3) the assistive technology must rely on the biosignals that appear when the user performs an action in a normal way, i.e. the user does not need to learn artificial strategies to control the device. To achieve these goals, the multimodal interface needs to make use of as many movement-related sources of information as possible, with these sources reflecting complementary aspects regarding movement generation. As it was presented in previous chapters, EEG activity acquired from regions around the central sulcus reflects cortical activity related to movement intentions and motor awareness [Desmurget et al., 2009]. Therefore, its integration with other noninvasive sensor modalities that track actual human movements, like EMG (analysing musle activation) and gyroscopic information (analysing rotations of body parts), makes it possible to characterise a voluntary movement during the planning and execution stages. In this chapter it is presented an Online EEG-based Detector of the Intention to Move (ODIM) and its integration (with EMG and gyrosopic technology) in a mHRI aimed to cancel pathological tremors by means of electrical stimulation. In such integrated platform, the proposed ODIM distinguishes resting states from intervals preceding the execution of upper-limb movements, and therefore it is aimed to provide the EMG/gyroscopesbased systems with predictions of voluntary movements. Having anticipated information about intended actions, the pathological tremor can be characterized and tracked from right before intended actions begin, giving rise to a successful detection of the voluntary movement onset and to a precise tremor tracking and cancellation from the exact moment at which the movement begins. The proposed platform helps to meet the aforementioned requirements of a natural interface. First, given that EEG holds information on the patient’s intentions to move, it enables the gyroscopes/EMG-based movement tracking systems to detect voluntary actions. Besides, providing the EMG and gyroscopic systems with predictive information on voluntary actions is useful to detect movement onsets Jaime Ib´a˜nez Pereda 67 Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor with short delays (this can be complicated if tremor is present before the movements begin). Finally, the ODIM proposed here is based only on the EEG patterns present before a subject self-initiates a movement with the upper-limb. Therefore, learning artificial mental strategies to command the interface is not required. The integration of EEG technology in such a mHRI is hence justified. Nevertheless, the EEG-based system must demonstrate a proper function providing cortical information that allows anticipation of intended actions and being robust against false activations during long periods of nonactivity. Additionally, the system must demonstrate its suitability for tremor patients, given that EEG movement-related patterns in the most typical tremor-related pathologies may be somewhat different to those observed in healthy subjects [Tam´as et al., 2006; Lu et al., 2010; Magnani et al., 1998, 2002]. As it was described in Chapter 2, two EEG patterns are suitable for movement intention detection: the BP and the ERD. Although both cortical processes appear approximately 2 s before the onset of voluntary movements, to detect the intention to move using the BP presents an important drawback: the “early-BP” presents small amplitudes (2-3 µV) [Bai et al., 2011], which are barely detectable in a single-trial analysis. For that reason, the robust online single-trial detection of the BP relies on “late-BP” detection. This makes it difficult to anticipate the onset of the movements using this cortical pattern. ERD, on the other hand, overcomes this problem since the switch between the synchronised and the desynchronised states is faster and more pronounced [Bai et al., 2011; Morash et al., 2008]. Several previous works have dealt with the problem of detecting the intention to move [Niazi et al., 2011; Bai et al., 2011; Lew et al., 2012]. On the one hand, [Niazi et al., 2011] and [Lew et al., 2012] used the BP to locate the onsets of voluntary movements performed with the ankle and the arm, respectively. A high percentage of movements was detected, although no anticipation was achieved due to the aforementioned characteristics of the BP. On the other hand, Bai et al. [Bai et al., 2011] used subject-specific ERDpatterns to detect the intention to move in healthy subjects. High prediction periods (0.62±0.25 s) were obtained with an average precision of 75±10 %, but a small number of movements was detected with most subjects analyzed (less than 50 % of the movements were detected with the best subject). Importantly, most of these studies provide results of paradigms in which rest intervals preceding voluntary actions last on average ˜ 5 s, thus reducing the chances of the systems to generate false detections. ODIM uses the ERD pattern to anticipate movements and it is validated with an asynchronous paradigm (no external cues are used to indicate when to move) on 6 healthy subjects and 4 patients with ET, which consitutes the most common tremor-related neurological disease, typically implicating postural and action tremor of the arms [Louis et al., 1998; Benito-Le´on and Louis, 2006]. Besides, the results of the integrated function of a 68 5.3 Methods mHRI taking advantage of the EEG information are provided in an additional experiment with 5 patients with ET. 5.3 Methods 5.3.1 Experimental protocols 5.3.1.1 Experiment 1 Six healthy subjects (one female), all right-handed and between 27 and 36 years old, and four ET patients, males, right-handed and between 75 and 85 years old were recruited. The patients and 2 control subjects were measured in a single session, while the rest of the control subjects participated in two measurement sessions performed over different days. Patients were diagnosed as ET according to the Movement Disorders Society Diagnostic Criteria [Deuschl et al., 1998]. They presented bilateral postural and action tremor of mild and moderate severity. Patients P01 and P02 presented also mild rest tremor. None of them had other neurological symptoms. The patients were asked not to take antitremorogenic drugs within the 24 hours before the experiments. During the experiments, subjects were seated in a comfortable chair and with the arms supported. One measurement session of one subject was divided into 3-minute-long runs. In each run, the subject was asked to stay steady and to repeat a motor task consisting of focusing on the dominant hand and performing a single wrist extension followed by a return to the resting position (with the arm and hand relaxed on the armrest of the chair). The subjects were asked to stare at a fixation cross presented on a wall in front of them to avoid ocular artifacts. An acoustic signal sounded 10 s after each movement onset to indicate that a new trial was starting. The subjects were asked to wait more than 3 s between the acoustic signal and the execution of the movement. A valid trial contained an initial acoustic signal followed by a period of no motor activity (before the subjects decided to start the movement), an execution of the motor task and an additional 10 s time period without motor activity (see Fig. 5.1). All patients and two control subjects (C05 and C06) completed six to eight 3-minute runs in a single session. The rest of the measured subjects completed two sessions on two different days. For those participating in a single session, the runs performed during this session were divided into runs for training (first 2 runs) and for classification (remaining runs in the session). As for the rest of the control subjects, the first session was used as the training dataset and the whole second session was used for validation. On average, 35±19 trials were used to calibrate the ODIM and 56±11 trials were used to validate it. In each trial, 87.4±2.8 % of the time corresponded to intervals with the subjects presenting a resting state, and these resting periods of time between Jaime Ib´a˜nez Pereda 69 Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor movements lasted more than 15 s. This is a relevant information in order to objectively validate the precision of the EEG system in asynchronous paradigms; the longer the idle states, the more likely it is that the system generates false activations. Acoustic signal Voluntary movement time Look at the fixation cross Subject-dependent period of time before the movement 10 s Acoustic signal > 3 s Figure 5.1: Graphical representation of one trial. 5.3.1.2 Experiment 2 This experiment was aimed to validate the combined function of the EEG and EMG and gyrscopic systems in the mHRI. Five essential tremor patients (2 female and 3 male) between 47 to 79 years old were recruited. All patients presented postural and kinetic tremor of mild or moderate severity. Medications were continued at the time of the recordings. Similar conditions to the first experiment were given. Subjects were asked to perform a series of exercises that are commonly employed in the clinic to assess tremor: finger to finger and finger to nose tests, and elevating both arms and keeping them outstretched against gravity. Each patient performed 6 repetitions of each exercise. In this case, the trials were separately recorded. The execution of all the trials followed the same scheme: patients were asked to stay relaxed avoiding eye movements, and self-initiate the exercise after allowing for a sufficient repose time after the trial started. In this case, the resting intervals preceding the self-initiated movements were significantly shorter than in Experiment 1. The system validation was performed offline using a leave-one-out procedure (to test the system on each trial, the rest of the trials were used to calibrate the ODIM). To evaluate the multimodal platform, only those classified trials with visible tremor were used to present the results of this experiment. On average, results of 10.0 ±5.6 trials per patient are presented. 5.3.2 Data acquisition EEG signals were recorded with passive Au electrodes from positions FC3, FCz, FC4, C5, C3, C1, Cz, C2, C4, C6, CP3, CPz and CP4 according to the extended international 10/20 system. Impedances were kept below 7 KOhm. The reference was set to the common 70 5.3 Methods potential of the two earlobes and Fz was used as ground. The amplifier filtered the signal between 0.1 and 60 Hz, and an additional 50 Hz notch filter was used. The sampling frequency was 256 Hz. Reference-free estimations of the EEG signals were obtained by spatially filtering the 13 channels acquired. A Laplacian filter was applied to the C3, C1, Cz, C2, and C4 positions [Hjorth, 1975], i.e. for each electrode position the average voltage of the four equally close neighbours was subtracted. For boundary channels, a common average reference was used (the average voltage of all channels was subtracted). Wrist extension/flexion was monitored by means of two gyroscopic sensors placed on the hand and forearm. Wrist rotation was obtained by computing the difference between both gyroscopes [Gallego et al., 2010]. Both measuring systems were acquired in two different computers and they were synchronised by means of a pulse signal that was generated by the computer storing the gyroscopes data and sent through a DAQ to the EEG (two pulses at the start and the end of the recordings and one pulse each time the IMUs detected a wrist extension). In addition to EEG and gyroscopic data, in Experiment 2 tremor was recorded from the most affected side (with which tasks were performed) using surface EMG. EMG signals were recorded over the wrist extensors and flexors with a 128-channel amplifier in differential configuration. A 64-channel array electrode was placed on the muscle belly, and a humidified wrist bracelet served as common reference. The signal was amplified, band-pass filtered (10-500 Hz), and sampled at 2048 Hz by a 12 bit A/D converter. Synchronization of the different systems was controlled by a digital clock signal. Only results from those trials with visible tremor are presented here. 5.3.3 Detection of the movement onset with the gyroscopes In order to detect the time at which each movement started in the training data (the recorded data used to calibrate the EEG system), wrist movement in the resting condition was characterised at the beginning of each session, and the threshold amplitude was set as two times the maximum amplitude value in this interval. The data from the gyroscopes were low-pass filtered (Butterworth, order 2, ≤6 Hz). Movements incorrectly detected by the online gyroscopes-based algorithm were either corrected or discarded manually after the sessions, to ensure a rigorous evaluation of the ODIM. 5.3.4 Description of the ODIM architecture The core of the ODIM consisted of a Bayesian Classifier (BC) fed by the logarithmic Power Spectral Density (PSD) values. Previous results presented in [Bai et al., 2007] showed that these techniques (the BC and PSD estimations) provide the best classification Jaime Ib´a˜nez Pereda 71 Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor performances in similar experiments. The BC also presents the advantage of requiring low computational load during its online function and also during its training process. During the function of the ODIM (see Fig. 5.2), the logarithmic power values of three selected channel/frequency pairs (see 5.3.4.1) were extracted from the EEG signal every 125 ms using 2-s windows.The power estimations were performed using Welch’s method (Hamming windows, 128 samples, 75 % overlap). A single class Bayesian Classifier (BC) was fed with these values and the three output probabilities were combined to generate the final output probability. An optimized threshold (see 5.3.4.3) was then used to convert this probability into a binary signal, and a Refractory Period (RP) was applied in order to maintain each positive output interval of the ODIM active for at least 2.5 s, thus generating a stable output of movement predictions [Townsend et al., 2004]. EEG Channels selection PSD Frequency selection Bayesian classifier Threshold Refractory period Movement prediction Gyroscopes Movement locations Training dataset with the previous movements Figure 5.2: Flowchart of the ODIM. The arrows crossing the blocks represent the adaptive design of the parameters in these blocks. 5.3.4.1 Selection of subject-specific optimal channels and frequencies for the ERD characterization based on the training data The process was aimed to search for the channel/frequency pairs with largest and most anticipative ERD. The process was divided in two steps. First, the system looked for the frequency at which the largest ERD was observed in each channel. This frequency was the one that maximized the ratio between the average frequency spectra of the basal and movement states. In previous tests with the training data of the control subjects, using this criterion provided better results in the selection of optimal frequency components than the Bhattacharyya index, the two-sample t-test and the Kullback-Leibler distance. The frequency spectrum of the movement state was characterised by averaging the PSDs of all the movement intervals included in the training dataset. The movement intervals were taken from 2 s before the onsets of the movements (when the average ERD is expected to begin in most subjects [Pfurtscheller and da Silva, 1999]) until they ended. Similarly, the frequency spectrum of the basal state was characterised by averaging the PSDs of all 72 5.3 Methods C01 C02 C03 C04 C05 C06 P01 P02 P03 P04 0 50 100 Recall (%) C01 C02 C03 C04 C05 C06 P01 P02 P03 P04 0 1 2 3 FPMR Figure 5.6: Comparison of the Recall and FPMR results for three conditions: 1) Both the model of the BC and the threshold are adapted (black), 2) only the model is adapted (grey), 3) only the threshold is adapted (white). Mean and standard deviations across runs are presented. 5.3.7 Results in Experiment 2 Table 5.4 shows the detection results of the ODIM with the valid trials in Experiment 2 (those in which tremor was visible). Results for patient 02 are not supplied since he did not exhibit a visible ERD. In general terms, although in this case a significantly lower number of examples was used to validate the system than in Experiment 1, higher recalls are obtained with equivalent number of false activations and similar amounts of anticipation. This was in part due to the fact that here, the resting periods of time preceding the movements were shorter. In summary, the results indicate that the mHRI was capable of consistently anticipating the intention to move (in those patients that exhibited ERD). Moreover, the delay in the detection of both voluntary movement and tremor was considerably increased in the patient 02, who did not present a detectable EEG-based movement anticipation (average delay 1.83±1.77 s and 1.79±0.91 s for the voluntary activity and the tremor, respectively) when compared to the other patients (average delay in all trials 0.88±0.45 s and 0.77±0.45 s for the voluntary activity and the tremor respectively). The outcome of increasing the overlapping of the windows that the EMG algorithm used from the 50 % to the 75 % was also evaluated. When the ODIM was used, a statistically lower (P<0.05) delay in the detection of both the voluntary Jaime Ib´a˜nez Pereda 79 Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor Patient Recall Prediction Continuous (%) period (s) Specificity (%) 01 88 0.75±0.98 96 02 - - - 03 92 1.84±1.52 95 04 67 0.41±0.37 96 05 80 1.43±1.39 86 Average 82 ±11 1.11 ±0.52 93 ±5 Table 5.4: Classification results of the ODIM in Experiment 2. movements and the tremor was observed, which highlights the benefit extracted from using the prediction of movements derived from EEG to drive the system. 5.4 Discussion This chapter presented and EEG-based system to predict online voluntary movements with the arm. The robustness of the system against false detections was demonstrated validating its continuous function with a protocol with non-action intervals between movements lasting over 15 s on average (1.4±0.3 false activations per minute were generated in Experiment 1). With most subjects, more than 50 % of the movements could be anticipated by the system. With two patients small recall results were obtained, although the late detection of the movements was achieved, suggesting a delayed appearence of the ERD pattern. In addition, it has been proposed for the first time a design of a multimodal interface taking advantage of the EEG information regarding motor intentions. The EEG system was aimed to give anticipatory information to other systems tracking voluntary movements and tremors and doing so, detection latencies of voluntary and tremulous movement onsets were significantly lower than in the case where no EEG technology was used. The ODIM represents a step forward in the development and validation of BCI technology for patients with tremor. The proposed interaction between EEG and other sensor modalities is also original. The ODIM is conceived to give advanced information on voluntary movements to other sensors, such as EMG and gyroscopes. Information from these sensors are in turn expected to trigger the electrical stimulation that assists tremor patients. As was shown in the comparative Table 5.2, EMGand gyroscopes-based systems require muscle contraction or actual movements to assess that an action is being performed, increasing the latency of the response of a system aimed to assist or compensate the voluntary movement. This is critical with tremor patients, since the tremors are superimposed to the voluntary movement and the precise detection of the movement onset 80 5.4 Discussion becomes more complicated. In this terms, the EEG activity becomes a valuable source of information to improve the response time of neurorobotic or neuroprosthetic devices. In fact, a synchronised operation of an active device and the user’s commands governing it is desired to improve the interface between man and machine [Gomez-Rodriguez et al., 2010]. This depends on how accurately the user’s intentions are estimated. Moreover, after anticipating information on future volitional movements it is then also interesting to start characterizing the patient’s tremor before each movement starts. In such case, the tremor cancellation can be tackled already before the start of the voluntary movement [Kinoshita et al., 2010]. ODIM performance has been tested with ET patients. As it was described in Chapter 4, ET seems to be due to abnormal oscillations within the thalamocortical and olivocerebelar pathways [Elble, 2006], and this may cause variations in the characteristics of the ERD patterns in patients with tremor as observed in previous studies [Tam´as et al., 2006; Lu et al., 2010]. Besides, the proprioception of hand movements while the tremor is present can also influence the ERD patterns during the intervals of intended basal (resting) activity in patients with rest tremor. The ERD single trial detection system must hence be tested with these kind of patients. Here, several differences were observed in Experiment 1 in the results with the ET group compared to the ones obtained with the control group. The feature selection showed that the frequencies at which the tremor patients exhibited ERD corresponded to the lower alphaand beta-bands (7-10 Hz and 13-19 Hz), while with the controls, most features were at frequencies in the 10-13 Hz range. The channels selected in both groups differed slightly, and the C3 position (covering the right hand cortical area) was more frequently selected in the control group than in the patients’ group. Pathological oscillations of cerebellothalamocortical pathways causing ET [BenitoLe´on and Louis, 2006] could be causative of such differences in the spatial and frequencial distribution of the ERD, although other factors, mainly the age of the patients, are also likely to play a role in this regard, in agreement with previous studies [Derambure et al., 1993]. No statistically significant differences were found in the Recall and FPMR results obtained here in Experiment 1, although two patients (P01 and P03) showed the worst performances. These results could be caused by the pathology of these patients, although it may also be due to differences in the task involvement (fatigue, concentration, motivation) of these patients as compared to the rest of the subjects measured. As no studies of ERD in ET patients have been documented to date, further research may be done in this area. Nevertheless, the performance of the ODIM with P02 and P04 is encouraging to consider the ODIM as a valid interface for patients with tremor. Also in Experiment 1, an adaptive design for the ODIM was proposed to face the expected inter-subject variability caused by changes in the subjects’ fatigue, concentration Jaime Ib´a˜nez Pereda 81 Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor and degree of involvement, among others [Blankertz et al., 2006; Shenoy et al., 2006]. Previous studies have demonstrated the benefits of adaptive BCIs based on sensorimotor rhythms [McFarland et al., 2011]. In the present study, no feedback was given, so no learning was expected. The ODIM worked using a training dataset acquired on a different day (in 4 control subjects) or with a small amount of training examples (all patients and 2 control subjects). In both cases the ODIM can benefit online from synchronised movement tracking with the gyroscopes, by enriching the training dataset each time new examples are accomplished. The results obtained with the adaptive design have been compared with non-adaptive alternatives. Using a fixed threshold worked worse with 4 subjects because it was too restrictive (C01 and C02) or too tolerant (P03 and P04). These differences were probably due to aforementioned changes in the subjects’ brain processes, which made the training dataset unrepresentative to choose a threshold for the validation dataset. Comparing the adaptive design with a design only adapting the threshold showed similar results. A higher number of movements predicted and of false detections was obtained in 9 out of 10 subjects with the adaptive alternative. For the here proposed application, the minimization of false detections was not so critical as the maximization of true positives, because the final decision for triggering an active strategy with electrical stimulation would rely on the EMG/gyroscopes-based system. Therefore, the results obtained with the adaptive model are more suitable in this case. Comparing the results obtained here in Experiment 1 with other works is difficult, since the experimental protocols used, the subjects measured and the goals addressed vary significantly. Several studies have presented results of EEG-based movement onset detection systems using the BP pattern and showing similar specificity results and significantly higher recall ratios without anticipation of voluntary movements (see [Niazi et al., 2011; Lew et al., 2012; Xu et al., 2014] and Chapter 6 in this thesis). The fact that, in those studies, movements were detected and not anticipated is a crucial aspect of the significant difference in this regard. The important increase in the number of late detections (Recall-Late) achieved in our study supports this idea. The characteristics of the experimental protocol used are also an important factor, since using longer non-action intervals has a direct influence on the specificity of the system (the longer the basal intervals, the more likely it will be that the system generates false activations). On the other hand, Bai et al. in [Bai et al., 2011] presented results of an EEG-based system predicting voluntary movements, but only 50 % of the movements were detected in the best case. In their study the length of the rest intervals preceding the movements was similar to that in [Niazi et al., 2011] and thus shorter than here. Results obtained in Experiment 2 provided a proof of concept (with a reduced number of trials and patients) on how an mHRI may benefit from the anticipated information 82 5.5 Chapter conclusions regarding motor tasks provided by the ODIM to detect and parameterize the concomitant tremor, in order to drive electrical stimulation to compensate it. The integration of the ODIM in the mHRI shortened the reaction time of the system: significantly (P<0.05) lower delays in movement and tremor onset detections were obtained when the ODIM was integrated in the mHRI. This aspect has obvious implications for tremor compensation, since a response of the interface matched in time with the intended actions of the patients becomes possible, giving rise to a more natural interaction. There are, nevertheless, two scenarios in which the mHRI needs to overcome the absence of ERD information. The first of them is those patients that present not classifiable ERD, where the EMG detection algorithms will have to assume larger detection delays (as for Patient 02 in Experiment 2). The second scenario corresponds to the generation of false positives by the EEG classifier, which unnecessarily increases the overlapping of windows of the EMG subsystem, but these misdetections do not propagate to the patient (electrical stimuli are only triggered by the EMG and gyroscopic systems). As a matter of fact, the idea of enhancing the reliability of the neuroprosthesis control by combining recorded data with redundant information (as in the case of EEG, EMG and gyroscopic signals) constitutes the rationale for always running the EMG classifier in parallel. It is worth noting, however, that the number of false negatives of the EEG classifier throughout the experiments here is remarkably low. It is also worth mentioning that the overlapping of the EMG windows could be increased more, which would yield a faster detection of both voluntary muscle activity and tremor. The value selected here was chosen to analyse the interest of the approach, while ensuring low computational burden. 5.5 Chapter conclusions Experiments with 6 healthy subjects and 4 ET patients were conducted to assert the ability of the proposed EEG-based system to anticipate voluntary movements while reducing the number of false activations during long (>15 s) periods of resting activity. On average, 60±10 % and 42±27 % of the movements were anticipated with the control subjects and the patients respectively. The number of false activations generated per minute was kept low in both groups (1.5±0.1 and 1.4±0.5) despite using an experimental protocol in which long non-action intervals were given. Further experiments with 5 additional ET patients were run to validate the interaction between EEGand EMG-based systems in a proposed mHRI for patients with tremors. The movement predictions provided by the EEG system allowed a significant improvement in the detection of movement and tremor onsets when the patients started new tasks. In summary, this chapter has proposed an asynchronous EEG application in which Jaime Ib´a˜nez Pereda 83 Chapter 5. Prediction of voluntary movements using the EEG signal and its application in BCI systems assisting patients with tremor anticipated detections of motor intentions are performed. To rigorously validate the online function of the system, already proposed metrics by other studies and ad hoc metrics defined here have been used. An adaptive configuration of the detector has been proposed, which allows an optimized robustness of the system when dealing with the nonstationarities of the EEG signals recorded along different measurement days. These are the first results of a BCI system in patients with tremors and the first time that a multimodal platform integrating EEG sensors with other movement-related sources of information is proposed and justified. 84 Chapter 6 Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation 6.1 Abstract As it was presented in the first chapters of this thesis, the EEG signal allows the characterization of movement-related cortical processes with high temporal accuracy. Chapter 5 demonstrated the potential use of the EEG signal to anticipate voluntary movements performed with the arm. In this chapter the goal is slightly modified: it is studied how accurately is it possible to decode the onset of voluntary movements with temporal resolution using the EEG signal. Developing online systems able to decode motor intentions at the exact time they occur is of special interest for the neurorehabilitation of stroke patients, since it then becomes possible to develop conditioning paradigms associating cortical and peripheral neural processes with temporal accuracy (in the range of hundreds of milliseconds). This chapter proposes for the first time an EEG-based detector of the onsets of voluntary upper-limb functional movements using information extracted from cortical rhythms and slow cortical potentials. The system is evaluated with data from healthy subjects and chronic stroke patients and a rationale for the combination of oscillatory and slow cortical informations is provided. Additionally, results of a feasibility study using the developed EEG system in a one-month BCI intervention with chronic stroke patients is presented. Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation 6.2 Introduction During the past few years, the development of brain-computer interfaces (BCIs) for the functional rehabilitation of patients with motor disabilities has gained special interest [Daly and Wolpaw, 2008; Buch et al., 2008]. The main purpose of BCIs in such scenarios is to provide a way to promote the neural rehabilitation of the patients. EEG-based systems allow the real-time characterization of the cortical activity over the motor cortex while the subject is performing motor tasks. This way, it becomes possible to detect online when a person is attempting or imaging a movement [Pfurtscheller and Solis-Escalante, 2009; Bai et al., 2011; Niazi et al., 2011], and to predict certain properties of the movement to be performed [Pfurtscheller et al., 2006; Morash et al., 2008; Gu et al., 2009b; Jochumsen et al., 2013]. Such information may in turn be used to close the loop with neuroprosthetic or neurorobotic devices. In this regard, recent studies have proven the importance of the proprioceptive feedback timing to achieve long-term associative neural facilitation effects [Mrachacz-Kersting et al., 2012; Niazi et al., 2012]. In a series of previous studies, it has been proposed the use of the BP (described in Chapter 2) to detect the movement intention [Niazi et al., 2011; Garipelli et al., 2013; Lew et al., 2012; Jochumsen et al., 2013; Xu et al., 2014]. Since the BP presents an identifiable pattern that is decaying until the movement starts, it is suitable to achieve temporal precision in the detection of the onsets of voluntary movements. In fact, previous studies showing results of online systems based on this pattern indicate that average detection latencies of 315 ±165 ms can be obtained [Xu et al., 2014]. Nevertheless, the BP is not detectable in all cases, since some subjects do not present a significant pattern during self-paced movements. In addition, results obtained in previous studies using the BP have not fully validated the use of this cortical pattern alone to detect movement intentions in stroke patients [Niazi et al., 2011]. In fact, altered BP patterns have been observed in previous studies with this type of patients [Daly et al., 2006; Fang et al., 2007]. A possible way of boosting EEG-based systems aimed to detect the onsets of voluntary movements is to combine the BP with other EEG movement-related patterns providing complementary information [Fatourechi et al., 2008]. The ERD (described in Chapter 2) is a well-known cortical pattern related to the execution of voluntary movements. Although a variable anticipation may be observed in the ERD of a specific channel and frequency in a subject during consecutive movements, the spatio-tempo-frequential distribution of the ERD observed when averaging a number of EEG segments preceding voluntary movements shows a desynchronization pattern attached to the movement event [Bai et al., 2005]. 86 6.3 Methods Therefore, the analysis of the ERD also provides certain degree of information regarding the timing of volitional motor actions. Indeed, previous studies have used the ERD pattern to anticipate movement events [Bai et al., 2011; Ib´a˜nez et al., 2013]. As in the analysis of the BP, the ERD pattern of stroke patients presents variations with respect to healthy subjects [Stepien et al., 2010]. Therefore, it is of special relevance to study how strokerelated cortical changes may affect a BCI driven by these cortical patterns. This chapter presents results from two experiments. In the first experiment, an EEGbased system combining the information extracted from the analysis of the BP and ERD cortical processes is proposed to estimate the onsets of voluntary upper-limb reaching movements. The comparison between the proposed classifier and equivalent classifiers using either the BP or the ERD patterns is also performed to justify the fusion of these two sources of information. The second experiment presents preliminary results of a BCI intervention for stroke patients using functional electrical stimulation (FES) and the developed EEG system. The intervention is tested with four chronic stroke patients in eight sessions along one month. Changes in two functional scales are studied to analyze the effects of the BCI intervention on the patients. 6.3 Methods 6.3.1 Participants Healthy subjects and chronic stroke patients were recruited for the two experiments described in this chapter (referred to as Exp1 and Exp2 from nowon here). Six healthy subjects (all males, right-handed and under 35 years old) were measured and considered the control group in Exp1. Nine patients were recruited (three females, age 62 ±14 years, mean ±SD; details are provided in Table 6.1). Patients P1-P6, P8 and P9 were recruited for Exp1. Patients P8 and P9 were discarded for further analysis because they could not comply with the demands of the task performed during the experimental protocol. Patients P2, P3, P5 and P7 participated in Exp2. None of the subjects measured had prior experience with BCI paradigms. 6.3.2 Data Acquisition The movements of the arm were measured with solid-state gyroscopes and EMG electrodes. Two gyroscopes, placed on the distal third of the forearm, and the middle of the arm measured the limb kinematics. The data were sampled at 100 Hz. Surface EMG was recorded using bipolar derivations on the main muscle groups involved in the execution of the reaching task (pectoralis major, anterior deltoids, medium Jaime Ib´a˜nez Pereda 87 Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation Pat. Age Gender Stroke Affected Years since F¨ugl-Meyer Minimental Ashworth Rh sessions code type hemisphere stroke a week P1 52 F Isquemic L 4 126 30 0 1 P2 54 M Isquemic R 4 69 30 2 2 P3 54 M Isquemic L 3 68 30 3 2 P4 75 M Hemorrg L 1 60 30 3 2 P5 69 M Hemorrg R 4 64 29 3 - P6 57 F Isquemic L 1 93 26 1 Discont P7 40 M Hemorrg R 13 81 30 3 2 P8 83 F Isquemic L 5 112 23 1 2 P9 75 M Isquemic L 3 - (mixed aphasia) - 2 2 Table 6.1: Demographic table of the patients participating in the present study. deltoids, biceps, triceps and wrist extensors). The data were amplified and sampled at 2,000 Hz. EEG signals were recorded from 31 positions (AFz, F3, F1, Fz, F2, F4, FC3, FC1, FCz, FC2, FC4, C5, C3, C1, Cz, C2, C4, C6, CP3, CP1, CPz, CP2, CP4, P3, P1, Pz, P2, P4, PO3, PO4 and Oz) using active Ag/AgCl electrodes. The reference was set to the voltage of the earlobe contralateral to the arm moved. AFz was used as ground. The signal was amplified and sampled at 256 Hz. All recorded data were synchronised with a common digital signal. Additionally, Functional Electrical Stimulation (FES) was used in Exp2 to provide proprioceptive feedback to the patients. Stimuli were delivered at the anterior deltoid, triceps and wrist extensors with a multichannel monopolar neurostimulator with charge compensated pulses. The common electrode was located at the oleocranon. A stimulation sequence was applied each time the FES system was triggered: first the anterior deltoid was stimulated during 500 ms alone, then the stimulation of triceps and extensors was also activated. The three muscles were then stimulated during 1 s and after this period of time the stimulation sequence was finished. The currents of the stimuli at each muscle were adjusted in each session to optimise the elicited movements in each patient. Pulse width and frequency were set to 250 µs and 30 pps, respectively. The stimulator was controlled by a stand alone computer (with a real-time operating system) that received activation commands from the computer recording the EEG activity via a digital signal. 6.3.3 Aims and description of the experimental protocol in Exp1 The first one of the two experiments presented in this chapter was intended to validate the EEG-based detector of movement intention. Each participant was measured during one single session. The study was performed in a soundand light-attenuated room. Participants sat in a comfortable chair with their arms supported on a table. During the measurement phase, participants were instructed to remain relaxed with their eyes open 88 6.4 Results healthy subjects, whereas activation patterns presented a central (P1, P2 and P5 in the alpha band and P2, P3 and P5 in the beta band) or bilateral distribution (P3 in the alpha band and P1 in the beta band) in the patients group. −1.6 −1.2 −0.8 −0.4 0 time (s) time (s) Normalized EEG amplitude 0.1 0.3 0.4 0.6 0.8 Normalized EEG amplitude 0.1 0.3 0.4 0.6 0.8 −1.6 −1.2 −0.8 −0.4 0 Figure 6.2: Average BP of all subjects (discontinuous lines), and average BP across subjects (solid line). Averages from healthy subjects and patients are presented in the left and right panels, respectively. 6.4.3 Results of the EEG-based detection of the onsets of movements Fig. 6.4 shows a representative example of the detector function on a single trial performed by participant C2. The different stages in the EEG signal processing to extract information regarding movement intention are represented. The three last curves show the estimations of the onset of the movement based either on the BP pattern, on the ERD pattern or on the combination of both, respectively. In this example, the EEG-based detection is achieved with few hundreds of milliseconds of anticipation. Table 6.2 summarizes the results obtained by the detector based on the ERD and BP patterns. On average 63.3 ±13.8 % and 66.4 ±18.8 % of GT are obtained with the healthy subjects and the patients, respectively. The percentage of true positives achieved with patients is smaller than with healthy subjects, but also the FP/min generated with the patients is higher. These results lead to a similar average performance of the system in terms of detections and false activations in both groups. Nevertheless, more delayed detections are obtained with patients (35.9 ±352.3 ms) than with healthy subjects (-89.9 ±349.2 ms). The features selected by the ERD-based detector of movement onsets in the healthy subjects and patients are summarized in Table 6.3 and Table 6.4, respectively. According Jaime Ib´a˜nez Pereda 95 Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation P1 P2 P3 P4 P5 P6 αERD βERD BP C1 C2 C3 C4 C5 C6 αERD βERD BP Figure 6.3: ERD and BP spatial maps with healthy subjects (left) and patients (right). Left and central columns show the spatial distribution of the α-ERD and β-ERD (normalized power changes) obtained by comparing a window of 1.5 s ending at the movement onset with an equivalent window 4 s before the onset. The third column shows the spatial distribution of the BP peak amplitude. For each column, the same colour scales are used with all subjects. Colour scale normalization is performed representing the lowest value in each column with dark blue and calibrating the level of dark red in order to optimize the patterns representation. to the average ERD patterns observed in section 6.4.2 a predominance of contralateral central features is observed in the first case (healthy subjects), therefore most features correspond to channel C3 and the surrounding positions. In the case of the patients, a more spatially spread distribution of selected features is obtained. Features from the midline (around Cz) become more relevant in this case. The selection of features from the alpha or beta band varies for each subject, although predominance of beta band features 96 6.4 Results −10 0 10 -50 0 50 −10 0 10 −10 0 10 −0.5 0 0.5 0 0.5 1 0 0.5 Gyroscope (rad/s) raw EEG (µV) BP part of EEG (µV) ERD part of EEG (µV) BP-based detection ERD-based estimation Combined estimation 1s Movement onset Detection Threshold Figure 6.4: Simulated online function of the single-trial EEG-based detector of onsets of voluntary movements. The plots show from top to bottom: 1) the gyroscopic data used to locate the actual onset of the movement, 2) the raw EEG signal of a single channel, 3) the virtual channel obtained after spatial and temporal filtering the EEG signal to detect the BP pattern, 4) the EEG signal in one channel after applying a small laplacian filter and a band-pass filter (between 6 Hz and 35 Hz) for the ERD-based detection, 5) the output of the matched filter applied by the BP-based detector, 6) the output of the bayesian classifier applied by the ERD-based detector, and 7) the final estimation of the intention to move and the optimal threshold level used to convert the estimation to a boolean signal. is observed. Finally, the tables show that selected features relative to the alpha-band in the case of the patients present lower frequencies than the ones in the group of healthy subjects. Jaime Ib´a˜nez Pereda 97 Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation Code GoodTr (%) TP (%) FP/min Latency (ms) C1 81.3 82.8 0.47 -48±351 C2 63.8 81.0 1.34 -24±278 C3 39.0 56.1 2.63 -180±476 C4 64.6 70.8 0.38 -198±322 C5 69.8 84.9 1.13 -3±388 C6 61.5 71.2 1.96 -164±290 Average 63.3 ±13.8 74.5 ±10.8 1.32 ±0.87 -89.9 ±349.2 P1 56.5 84.8 1.83 -58±368 P2 75.0 83.3 0.92 123±290 P3 60.3 80.9 1.94 98±386 P4 60.0 70.0 1.08 83±449 P5 100.0 100.0 0.00 -89±147 P6 46.5 74.4 3.21 50±520 Average 66.4 ±18.8 82.2 ±10.4 1.50 ±1.09 35.9 ±352.3 Table 6.2: Detection results obtained with control subjects and patients. C1 C2 C3 C4 C5 C6 C3/21Hz C3/12Hz Pz/12Hz F1/7Hz C3/12Hz FC3/19Hz CP3/21Hz C3/11Hz C3/12Hz F1/8Hz C3/19Hz CP1/19Hz C3/20Hz C3/23Hz C3/13Hz C6/29Hz C3/11Hz FC3/20Hz CP3/20Hz FC1/18Hz FC4/9Hz C3/27Hz CP3/10Hz FC3/18Hz C3/10Hz FC1/17Hz P1/12Hz FC1/23Hz CP3/11Hz F3/19Hz C3/19Hz C3/22Hz P1/11Hz C3/26Hz C3/22Hz CPz/20Hz C3/22Hz C2/17Hz CP1/8Hz C3/24Hz CP3/12Hz C1/19Hz CP3/19Hz FC1/19Hz Pz/10Hz C3/28Hz Pz/11Hz F3/18Hz C3/9Hz FC1/14Hz P1/9Hz FC2/18Hz C3/18Hz CP3/18Hz CP3/22Hz C3/13Hz FC4/10Hz C3/29Hz CP3/13Hz FC3/17Hz Table 6.3: Features selected by the ERD-based detector for the control group. Fig. 6.7 compares the detection results obtained with the combined detector (ERD and BP) with the results obtained by detectors based only on the BP or the ERD. Statistically significant differences between the three detectors are found in GT, TP and FP/min (p = 0.002, p= 0.010 and p= 0.008, respectively). Pos-hoc multiple comparisons show significant differences between the ERD-based detector and the combined detector in GT (p= 0.007) and FP/min (p= 0.015), but not in TP (p= 0.192). In the comparison between the BP-based detector and the combined detector, significant differences are found in GT (p= 0.003) and TP (p= 0.003), but not in FP/min (p= 0.0.059). Finally, no significant differences are found in GT (p= 0.611), TP (p= 1) and FP/min (p= 0.305) between the detector based on the ERD and the one based on the BP. 98 6.4 Results P1 P2 P3 P4 P5 P6 C1/9Hz C2/9Hz Cz/20Hz F3/8Hz CP2/13Hz C3/14Hz Cz/13Hz C2/8Hz Cz/21Hz C1/10Hz C2/13Hz P2/18Hz FC1/10Hz C2/10Hz Cz/13Hz F3/9Hz C1/22Hz C3/19Hz FC1/13Hz CP2/18Hz Cz/22Hz C2/11Hz C1/21Hz C2/23Hz C1/10Hz C2/7Hz Cz/14Hz F1/8Hz Cz/21Hz CP3/14Hz CP4/18Hz Cz/9Hz Cz/16Hz P1/10Hz C1/20Hz CP1/15Hz FC1/9Hz Cz/10Hz Cz/15Hz C1/9Hz CPz/22Hz FC2/19Hz FC1/11Hz CP2/19Hz Cz/17Hz P3/8Hz CPz/16Hz Pz/22Hz C1/12Hz CP2/17Hz CP1/11Hz F4/20Hz CPz/12Hz CP4/21Hz C1/13Hz Cz/8Hz Cz/19Hz FC3/8Hz C1/23Hz CP3/11Hz Table 6.4: Features selected by the ERD-based detector for the patients. For healthy subjects, the detector combining ERD and BP information achieves 6.5 ± 5.2 % more GT than the BP-based detector and 22.4 ±10.0 % more GT than the ERDbased detector (see Table 6.5). For patients, the percentage of GT also increases when using the combined detector (13.3 ±10.9 % and 12.6 ±16.3 % increase as compared to the BPand ERD-based detectors, respectively). 0 50 100 GT (%) 0 50 100 TP (%) P1 P2 P3 P4 P5 P6 0 1 2 3 4 FP/min 0 50 100 GT (%) 0 50 100 TP (%) C1 C2 C3 C4 C5 C6 0 1 2 3 4 FP/min BP ERD ERD+BP Figure 6.5: Performances of the three compared detectors (BP-based, ERD-based and combined detector) in the healthy subjects group (left) and in the patients (right) in terms of GT, TP and FP/min Finally, the latencies in the detections of the movement onsets are represented by means of histograms in Fig. 6.6. The latencies obtained when using the detectors based only on the BP or the ERD information are superimposed in the figure. The histograms Jaime Ib´a˜nez Pereda 99 Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation Code Combined vs BP Combined vs ERD C1 4.7 32.8 C2 10.3 12.1 C3 14.6 12.2 C4 0.0 35.4 C5 5.7 18.9 C6 3.8 23.1 Average 6.5 ±5.2 22.4 ±10.0 P1 4.3 10.9 P2 20.8 -6.9 P3 30.9 -4.4 P4 10.0 36.0 P5 1.9 19.2 P6 11.6 20.9 Average 13.3 ±10.9 12.6 ±16.3 Table 6.5: Gain in the performance of the detector (GT in %) when using the combined information of the ERD and BP compared to the use of either of these patterns alone. shown depend on how much the ERD and BP patterns vary across trials with respect to the onsets of the movements, and also on the detection threshold applied to each one of the three detectors. The figure shows a more delayed distribution of the detections with the group of patients. Nonetheless, around 85 % of these BP detections are located earlier than +375 ms. Given that the window used for the BP detector are 1.5 s long, this result supports the absence of movement artefacts in the activity analysed. The ERD-based detector appears to be the less precise in terms of latencies of the detections, while the BP-based detector presents distributions clearly centred at t = 0 s. Also noticeably, the ERD-based detector shows a certain degree of anticipation in the detections of movement onsets in the group of healthy subjects, although it generates delayed detections in the case of the patients. 6.4.4 Results of Exp2 Overall, patients could reliably control the EEG-based interface by performing the selfpaced movements and low detection latencies were obtained in most cases. Fig. 6.7 shows, for each intervention session and patient, the percentages of GT obtained and the latencies for the correct detections of the movements. The GT percentages increased across sessions in P2, P3 and P7, suggesting learning mechanisms in the interaction with the BCI platform. The best GT results were 80.9 %, 64.4 %, 91.7 % and 81.2 % for patients P2, P3, P5 and P7, respectively (green bars in the left panels of Fig. 6.7). Results were similar to those obtained in Exp1. The detection latencies were stable across sessions and in some 100 6.4 Results 20 18 16 14 12 10 8 6 4 2 000 +750-750 +750-750-375 +375 -375 +375 % of detections % of detections Latency (ms) Latency (ms) 16 18 16 14 12 10 8 6 4 2 0 ERD-BP ERD BP ERD-BP ERD BP Figure 6.6: Histograms of the distances between the movement detections and the actual movement onsets for healthy subjects (left panel) and stroke patients (right panel). The histograms of the detectors based only on the ERD or the BP are superimposed in the graphs. cases (P3 and P7) they slightly improved as the intervention evolved. Average detection latencies (considering all sessions) for P2, P3, P5 and P7 were 202 ±266 ms, 130 ±316ms, 3±190 ms and 103 ±254 ms, respectively. Unsuccessful results of the BCI system were only observed in one session (sixth session with patient P3). In this case the BCI-based intervention was cancelled since the patient reported an uncomfortable interaction with the FES system. As for the EEG-based system performance with motor imagery (blue bars in the left panels of Fig. 6.7), results varied among patients and they provided in all cases reliable estimations. GT results over 50 % were obtained in three out of four patients, which means that in more than 50 % of the trials the BCI system was able to successfully detect the onset of the movement imagination without generating any false activation in the resting period preceding it. Bad trials, on the other hand, were those presenting a false activation in the resting period preceding the movement, or those in which movement imagery was not detected or it was detected too late according to the patients’ reports. Table 6.6 shows the changes obtained in the two evaluated functional scales. Average increases of 10.5 ±8.7 and 15.7 ±11.9 points in the SIS and the FMI were obtained with the intervention. Patients P2, P5 and P7 showed changes in the quantified FMI over the minimal detectable change (which is 5.2 points for upper-extremity assessments). Changes of FMI in P3 were slightly below this threshold, despite the positive results observed in the self-report test. Interestingly, P3 was also the patient showing the worst detection results Jaime Ib´a˜nez Pereda 101 Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation 0 50 100 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 −500 0 500 1 2 3 4 5 6 7 8 0 50 100 1 2 3 4 5 6 7 8 −500 0 500 0 50 100 1 2 3 4 5 6 7 8 −500 0 500 0 50 100 Session Nr. GT (%) 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 −500 0 500 Session Nr. Latency (ms) P2 P3 P5 P7 P7 P5 P3 P2 1 2 3 4 5 6 7 8 Imag Imag Imag Imag Lost data Lost data Lost data Lost data Bad BCI performance 82% TP 13 FP/session 64% TP 10 FP/session 92% TP 7 FP/session 81% TP 14 FP/session Figure 6.7: Summary of the EEG-based detector performance during the intervention with the patients. Left panels: GT (%) results along sessions for each patient. The best session in each patient is represented with a green bar and for this session the percentage of TP and the number of FP is shown. The GT results for the imagined movements performed by the patients is represented with blue bars. Right panels: Detection latencies (Mean ± SD) along the sessions for each patient. The real onsets of the movements are represented with dashed red lines. of the EEG-based system, both in terms of GT and detection latencies across intervention sessions. Code SIS-pre SIS-post FMI-pre FMI-post P2 64 74 61 93 P3 66 79 83 88 P5 44 64 65 82 P7 73 72 81 90 Table 6.6: ISI and FMI scales of the patients before and after the BCI intervention. 102 6.5 Discussion 6.5 Discussion The accuracy with which movements can be detected online using EEG activity (both in terms of temporal precision and ratio between true and false activations) represents an important criterion to decide whether BCI technology can be brought to clinical practice in neurorehabilitation environments. This study shows the results of an EEG-based detector of voluntary movement onsets combining information extracted from the processing of cortical rhythms and slow cortical potentials. This is the first time that both sources of information are combined to this end. It is also the first time in which the benefits of a detector combining information from the ERD and BP patterns in patients with stroke are demonstrated. Moreover, the EEG system has been tested in a BCI intervention lasting one month with chronic stroke patients. The observed changes in the FMI of three out of four patients were over the minimal detectable change in F¨ugl-Meyer assessments of the upper-extremity function. This is, to the author’s knowledge, the first study of a BCI intervention for upper-limb movements focusing on the idea of inducing associative cortico-muscular facilitation by means of an accurate (in terms of temporal precision) characterization of motor intentions. Previous studies have described several aspects on the characterization of the BP to locate onsets of voluntary movements. On the one hand, Garipelli et al. studied the relevance of choosing appropriate spatial and temporal filters to extract the BP pattern [Garipelli et al., 2013], without showing results regarding temporal precision in the detections. In a study by Lew et al., average results of BP detection were presented for healthy subjects and stroke patients, although no single trial validation was carried out [Lew et al., 2012]. Up to date there are, to the authors’ knowledge, no studies regarding the detection of upper-limb voluntary movements based on the detection of the BP and using an online feasible design. In a recent study, Xu et al. presented a system using a manifold method (Locality Preserving Projection) with a LDA classifier to optimize the classification of the BP. The algorithm was tested on healthy subjects performing ankle dorsiflexions. The TP and FP/min results obtained in that study (79 ±12 % and 1.04 ±0.8, respectively) were similar to the ones obtained here with the healthy subjects and upper-limb movements. Nevertheless, the average latencies presented in their study (315 ±165 ms) were higher than the ones obtained here. This differences could be due to variations in the way subjects performed the task in each experiment (differences between upperand lower limb cortical patterns, length of the resting intervals between movements and speed of movements among others). The observed differences could also be due to the combined use of the ERD and BP features proposed here, which allows to reduce the rate of FP and, as a consequence, allows the selection of less restrictive (more anticipative) Jaime Ib´a˜nez Pereda 103 Chapter 6. Detection of the onsets of upper-limb reaching movements using ERD and BP patterns to elicit associative facilitation detection thresholds. While several previous studies have made use of the cortical rhythms to either detect movement events [Townsend et al., 2004; M¨uller-Putz et al., 2005] or to anticipate movement intentions (see [Bai et al., 2011] and Chapter 5 in this thesis), no studies so far have tried to use ERD information to locate onsets of voluntary movement with time precision. In a previous study by Fatourechi et al., the combined use of cortical rhythms and slow cortical potentials was proposed for an asynchronous BCI, although in that case the device was not intended to detect the onset of voluntary movements [Fatourechi et al., 2008]. The Na¨ıve Bayes classifier described here has demonstrated that the ERD supplies valuable information in this sense. Indeed, it has been shown here the benefits of the combined use of the information about the ERD and BP as compared to detectors relying solely on either the BP or the ERD. Significantly better performances could be achieved with the combined detector in all metrics analysed: a higher number of GT and TP was achieved with lower rates of false activations during the resting intervals. Previous studies have demonstrated that different neural mechanisms are involved in the generation of the ERD and the BP, and therefore may justify their complementarity. On the one hand, the BP is assumed to originate in the presupplementary and supplementary motor areas [Babiloni et al., 1999; Shibasaki and Hallett, 2006], which are associated with the movement planning and with the process of focusing on the intention to move [Lau et al., 2004]. On the other hand, the ERD is first visible over the contralateral motor cortex [Pfurtscheller and da Silva, 1999], and it is associated with the formation of more specific neural assemblies synchronized at higher frequencies in order to generate the desired descending motor commands [Pfurtscheller and da Silva, 1999; Buzs´aki and Draguhn, 2004]. The spatial distribution of both phenomena in the here presented data also points to different cortical sources. Given these evidences, it seems reasonable to point to an improved outcome in the combination of both sources of information to estimate certain aspects regarding the motor planning. Differences in the average ERD and BP patterns between patients and healthy subjects were found in Exp1. On the one hand, a delayed peak of the BP was observed in the patients group, likely associated with the higher cognitive motor planning time and the slower speed with which stroke patients perform voluntary movements [Daly et al., 2006; Jochumsen et al., 2013]. On the other hand, differences in the spatial distribution of both ERD and BP patterns were also observed (see Fig. 6.3), reflecting altered cortical activation patterns in stroke patients, also described in previous studies [Wiese et al., 2004; Daly et al., 2006; Platz, 2000; Stepien et al., 2010]. Regarding the single-trial detection results, previous offline studies [Niazi et al., 2011] showed differences in the BPbased detection performance with healthy subjects and stroke patients (significantly worse 104 7.1 General overview of the work presented in this thesis than others) and the more separated cortical representation of the different joints of a single-limb. In order to fully validate the hypothesis that EEG carries information that allows the classification of the proposed tasks, a set of tests was carried out, all of them aimed at proving that other explanations for the obtained results (e.g. the fact that different initial positions affect the classification) could be discarded. The obtained results open a door to integrating advanced EEG classification techniques in BCI interventions for rehabilitation, enriching the capacities of the BCI systems. The second study in the thesis has presented a neurophysiological characterization of the effects of a clinically used drug (alprazolam) in the tremors and the cortical activity of patients with ET, which has its origin in the pathological behaviour of the central nervous system. The presented study shows the temporal dynamics (due to the drug effects) of neurophysiological variables related to tremor manifestation in ET. This study is in line with other previous works, by a number of research groups, in which the main purpose is to characterize the mechanisms through which tremor is generated and altered in patients with ET, a neurological disease whose origin and action mechanisms are nowadays still not well understood. In this kind of studies, electrophysiological techniques such as the EEG and the EMG are of great interest, since they allow the characterization of electrical processes with high temporal resolutions (in the order of ˜ 1 ms), which is a critical factor to detect the changes in the neurophysiological function intended to be characterized. Additionally, studies of connectivity between distant neural networks provides highly informative data regarding interacting structures and the way in which this interaction changes along time. In the concrete case of the study included in this thesis, the main contribution resides in the characterization of the interplay between the beta and tremor-related oscillations at the cortical level, a phenomenon expected to be shared by other subcortical structures. The interaction between these brain oscillations results in changes of the apparent tremor, and therefore, further understanding them will improve tremor management in patients with ET. Finally, the presented results constitute the first objective quantification of tremor reduction in ET as a result of the administration of a drug. Such numerical description of the effect of a drug in a given pathology are currently demanded by clinical environments, so that a precise and objective characterization of the patients status and the outcomes of applied treatments can be obtained. In the third of the four studies included in this thesis, the design of an EEG-based system to anticipate voluntary movements and its integration in a BNCI to compensate pathological tremors have been presented. The proposed system is conceived as a proof of concept of multimodal systems to be used on patients with pathological tremors. The experimental sessions carried out allowed the evaluation of the EEG-based system as well as the whole acquisition system on patients with ET. The experimental paradigms used Jaime Ib´a˜nez Pereda 111 Chapter 7. Conclusions and future work represented a simplified scenario of the real one in which the patient would be using the platform: long periods of inactivity followed by self-initiated simple movements were used in order to test the ability of the system to provide reliable and anticipated estimations about motor planning when the subjects were about to move. Results achieved demonstrate the potential of the EEG signal to be used to describe periods of movement preparation and they also show that, under optimal conditions (subjects concentrated in the task and reduced electromagnetic interferences), estimations on movement intentions may be achieved reliably (with high percentages of true positives and reduced number of false activations). Nevertheless, the proposed study presents a set of technological and methodological limitations that reduce the impact of these results in daily living conditions: wearable technologies working reliably at home are nowadays not available and the analysis of the EEG signal with currently available techniques is still not able to avoid its contamination (reducing the signal-to-noise ratio), produced while users perform daily living tasks. A major prerequisite of BCI systems for motor compensation is that the benefits provided by the technology outbalance the disadvantages of using it. These disadvantages may be caused, among others, by aesthetic or ergonomic factors regarding the use of the wearable technologies, economic costs associated with the development of the used technology or the cognitive requirements that the use of human-machine interfaces demand from the potential users. From the author’s point of view, there are currently no commercial BCI devices for motor compensation that meet the aforementioned requisites. There are, nevertheless, cases in which the application of the BCI systems results in a clear improvement of the patients’ capabilities, such as spelling interfaces for complete locked-in patients, giving them the only possible way of communication with the outside world. In these cases, the benefits provided by the BCI systems will more likely justify the efforts of using the available technology. The system proposed in this thesis must be considered a proof of concept of the advantages that can be derived from the use of multimodal systems for the precise neurophysiological characterization of the movement generation chain originated in the brain and manifested in the peripheral limbs. While the here proposed mHRI system integrating EEG technology aims at meeting some of the aforementioned requirements for motor compensation technologies (a small number of electrodes is used, adaptive algorithms reduce training demands of the system, the processing techniques described require a relatively small computational load and the system is focused on ecologically characterizing the natural cortical activity observed when a subject performs a voluntary movement), future works in this line will need to focus on ways to solve the mentioned limitations of the technology. Finally, the fourth proposed study has presented an EEG-based technique to detect mental states associated with the initiation of voluntary functional movements. Inspired 112 7.1 General overview of the work presented in this thesis by some previous works regarding the online single-trial decoding of the BP [Niazi et al., 2011; Jochumsen et al., 2013], this study has proposed a way to optimize the detection of the movement onset-related mental states in patients with stroke, and this has been carried out by combining two different types of information: oscillatory changes (related to the ERD) and low-frequency cortical components (giving rise to the BP). The obtained results have demonstrated that it is possible to generate a reliable control signal about the onset of voluntary actions with temporal precision, high recall ratios and almost no false detections in experimental paradigms that could be easily transferable to clinical environments in case minor adjustments were performed (optimize the number of channels used for the detection, reduce training periods of time, increase the detector efficacy with subjects not showing identifiable ERD/BP patterns, etcetera). Importantly, results obtained with chronic stroke patients were similar than those with the control subjects despite the patients’ altered cortical activity [Daly et al., 2006; Stepien et al., 2010]. Consequently, these patients are suitable for BCI systems using electrical stimulation aimed to provide associative facilitation between the cortex and the muscles of affected limbs. In addition, a BCI intervention for patients with a stroke has also been proposed and preliminary results of a clinical validation with four chronic stroke patients have been presented. Although further research must be carried out to fully understand the effects of EEG-based conditioning paradigms in the motor function of stroke patients, results here provide evidences of a possible improvement in the motor condition of the patients after a whole month intervention. In summary, this thesis has proposed a set of novel scientific studies framed in the main research lines that are being currently explored with EEG technology. According to the results and conclusions reached in the proposed studies, it is considered that EEG systems represent a powerful way to characterise the neurophysiological mechanisms of neurologic diseases, and that the acquired information from this sort of studies represents a non-invasive and efficient opportunity to look for the cerebral regions originating certain motor-related pathologies. On the other hand, experiments carried out here with BCI systems using the EEG signal have demonstrated to be reliable and of special interest for rehabilitation scenarios, while the BCI application in daily-living conditions represents a challenging objective that still need to be further explored and requires dramatic technological improvements in order to find more robust and wearable technologies that can lead to an actual benefit of patients using the BCI systems for assistive/compensatory purposes. In conclusion, despite the several and well known limitations of EEG technologies for the analysis of the brain activity, it can be established that these systems allow the acquisition of highly relevant cortical information regarding motor-related mental processes, which makes this kind of technology a valuable tool for the research and conditioning of Jaime Ib´a˜nez Pereda 113 Chapter 7. Conclusions and future work the human neurological system. 7.2 Contributions This thesis has yielded advances from both the technological and the scientific points of view in all studies proposed. The main contributions from the technological point of view are: •The design of an integrated upper-limb platform working in real-time. The platform was designed to acquire information from different types of noninvasive sensors (EEG, EMG and gyroscopic sensors) characterising the planning and execution of voluntary movements. The platform was also capable of processing online the acquired data and generating an adequate feedback. •The development of signal processing and classifying techniques adapted to the kind of signal recorded in the two kinds of patients considered in this thesis and to the requirements of online processing and real-time single-trial function desired for BCI applications. Especially in this regard, an original methodology to detect onsets of voluntary movements using slow cortical potentials and cortical rhythms has been presented. •The design and validation in real-time of asynchronous BCI systems using motor planning EEG segments to anticipate or detect when patients begin a voluntary movement with the upper-limb. •The proof of concept of the use of the EEG activity in a mHRI architecture that constitutes the first multimodal interface taking advantage of the combined acquisition of EEG, EMG and gyroscopic data, which allows the concurrent characterization of different parts of the body associated with the execution of a movement. The main scientific contributions of this thesis are: •It has been proposed for the first time an experiment to inspect whether the EEG signal carries enough information to classify up to seven different tasks performed with a single limb. Both the methodology applied and the validation procedure are also innovative in this sort of studies. •It has been presented the first neurophysiological study using EEG and EMG data to analyse the effects of a drug on cortical activity and tremors of patients with ET. 114 7.3 Scientific dissemination In addition, the obtained results have shown for the first time that a significant correlation exists between the dynamics of specific cortical oscillations and pathological tremor manifestation as a consequence of the drug effects. •The study of the EEG-based anticipation of voluntary movements presented in Chapter 5 was the first demonstration (to the author’s knowledge) of the capacity of the EEG signal to provide reliable movement predictions based on single-trial classification of online data of healthy subjects and ET patients. This study also provides, for the first time, the results of a BCI system tested in ET patients and it represents an original approach to BCI applications for this group of patients. •It has been demonstrated for the first time the relevance of combining different cortical sources of information (such as BP and ERD) to estimate the initiation of voluntary movements with the upper-limb. In this line, special relevance may be given to the positive results achieved with stroke patients, improving the results presented by similar previous EEG-based studies by other research groups. It has also been proposed for the first time an upper-limb intervention protocol for stroke patients using BP and ERD patterns to provide proprioceptive feedback tightly associated with the patients’ expectations of movement. The effects of the proposed intervention have been studied with a small group of patients. 7.3 Scientific dissemination The work performed to carry out this thesis has given rise to a number of contributions in scientific journals, conferences and book chapters in the neurorehabilitation framework. The following lines summarize these contributions: Publications in journals: •J.A. Gallego, J.L. Dideriksen, A. Holobar, J. Ib´a˜nez, J.P. Romero, J.L. Pons, E. Rocon, D. Farina. Properties and determinants of the relative phase between neural drives to antagonist muscles in essential tremor. Brain. To be submitted. •E. Monge, F. Molina, F.M. Rivas, J. Ib´a˜nez, J.I. Serrano, I. Alguacil, J.C. Miangolarra. Electroencefalograf´ıa como m´etodo de evaluaci´on tras un ictus. Una revisi´on actualizada. Neurolog´ıa. In Press. •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo, J. M´ınguez, J.L. Pons. Predictive classification of self-paced upper-limb analytical movements with EEG. Medical & Biological Engineering & Computing. (second revision). Jaime Ib´a˜nez Pereda 115 Chapter 7. Conclusions and future work •J.A. Gallego, J.L. Dideriksen, A. Holobar, J. Ib´a˜nez, E. Rocon, J.L. Pons, D. Farina. Neural drive to muscle and common synaptic inputs to the motor neuron pool in essential tremor. Journal of Neurophysiology. Submitted. •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo, E. Monge, F. Molina, I. Alguacil, J.L. Pons. Detection of the onset of upper-limb movements based on the combined analysis of changes in the sensorimotor rhythms and slow cortical potentials. Journal of Neural Engineering, 11(5):056009, 2014. •J. Ib´a˜nez, J. Gonz´alez de la Aleja, J.A. Gallego, J.P. Romero, R.A. Sa´ız-D´ıaz, J. Benito-Le´on, E. Rocon. Effects of Alprazolam on Cortical Activity and Tremors in Patients with Essential Tremor. PLoS ONE. 9(3): e93159, 2014 •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo, J.A. Gallego, E. Rocon. Online detector of movement intention based on EEG – Application in tremor patients. Biomedical Signal Processing and Control, 8(6):822-9, 2013. •J.A. Gallego, J. Ib´a˜nez, J.L. Dideriksen, J.I. Serrano, M.D. del Castillo, D. Farina, E. Rocon. A multimodal human-robot interface to drive a neuroprosthesis for tremor management. IEEE Transactions on Systems, Man and Cybernetics, Part C: Applications and Reviews, 42(6):1159-68, 2012. •M.D. del Castillo, J.I. Serrano, J. Ib´a˜nez. Metodolog´ıa para la creaci´on de una interfaz cerebro-computador aplicada a la identificaci´on de la intenci´on de movimiento. Revista Iberoamericana de Autom´atica e Inform´atica Industrial (RIAI). 8(2):93-102. 2011. Book chapters: •S. Cremoux, J. Ib´a˜nez, S. Ates, A. Dess´ı. Neuromodulation on Cerebral Activities in Emerging therapies in neurorehabilitation, J.L. Pons and D. Torricelli (Eds.), Springer Verlag, 2014. Selected publications in conference proceedings: •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo, E. Monge, F. Molina, F.M. Rivas, I. Alguacil, J.C. Miangolarra, J.L. Pons. Upper-Limb Muscular Electrical Stimulation Driven by EEG-Based Detections of the Intentions to Move: A Proposed Intervention for Patients with Stroke. 2014 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, accepted. 116 7.3 Scientific dissemination •J. Ib´a˜nez, F. Molina, J.I. Serrano, M.D. del Castillo, E. Monge, F.M. Rivas, M. Carratal´a, J. Iglesias, I. Alguacil, A. Cuesta, R.Cano, J.C. Miangolarra, J.L. Pons. A BCI intervention for upper-limb functional movements of chronic stroke patients. 2014 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, accepted abstract. •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo, E. Monge, F. Molina, F.M. Rivas, I. Alguacil, J.C. Miangolarra, J.L. Pons. Detection of the Onset of Voluntary Movements Based on the Combination of ERD and BP Cortical Patterns. Replace, Repair, Restore, RelieveBridging Clinical and Engineering Solutions in Neurorehabilitation. Springer International Publishing, 437-46; 2014. •J. Ib´a˜nez, E. Monge, J.I. Serrano, M.D. del Castillo, F. Molina, J.L. Pons. Erdand bp-based movement onset detectors in stroke patients. XX Congress of the International Society of Electrophysiology and Kinesiology ISEK 2014, accepted abstract. •I. Alguacil, E. Monge, F. Molina, F.M. Rivas, R. Cano, J. Ib´a˜nez. Entrenamiento de los ritmos motores corticales en sujetos con ictus intervenidos con Brain-Computer Inetrface. 52 Congreso Nacional de la Sociedad Espa˜nola de Rehabilitaci´on y Medicina F´ısica SERMEF 2014, accepted presentation. •J. Ib´a˜nez, M.D. del Castillo, J.I. Serrano, F. Molina, E. Monge, F.M. Rivas, J.C. Miangolarra, J.L. Pons. Single-Trial Detection of the Event-Related Desynchronization to Locate with Temporal Precision the Onset of Voluntary Movements in Stroke Patients. XIII Mediterranean Conference on Medical and Biological Engineering and Computing 2013, 1651-1654; 2013. •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo. Asynchronous BCIs for the Early Detection and Classification of Voluntary Movements: Applications in Stroke Rehabilitation. Converging Clinical and Engineering Research on Neurorehabilitation Biosystems & Biorobotics Volume 1, 629-633; 2013. •I. Alguacil, E. Monge, A. Cuesta, J. Ib´a˜nez, F. Molina, M. P´erez de Heredia. Detecci´on de la intenci´on de movimiento mediante Brain-Computer Interface en el ictus. 51 Congreso Nacional de la Sociedad Espa˜nola de Rehabilitaci´on y Medicina F´ısica SERMEF 2013, accepted presentation. •J. Ib´a˜nez, J.I. Serrano, M.D. del Castillo, L. Barrios, J.A. Gallego, E. Rocon. An EEG-based design for the online detection of movement intention. Advances in Computational Intelligence. Lecture Notes in Computer Science. 6691, 370-377, 2011. Jaime Ib´a˜nez Pereda 117 Chapter 7. Conclusions and future work •J.A. Gallego, J. Ib´a˜nez, J.L. Dideriksen, J.I. Serrano, M.D. del Castillo, D. Farina, E. Rocon, J.L. Pons. Simultaneous recordings of the central and peripheral nervous system together with joint biomechanics improve the characterization of tremor. IFBME Proceedings of the World Congress on Medical Physics and Biomedical Engineering 2012 Abstract. •J. Ib´a˜nez, J.I. Serrano, M.D. Del Castillo, L. Barrios. An asynchronous bmi system for online single-trial detection of movement intention. Proceedings of 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 4562-65, 2010. •J.A. Gallego, E. Rocon, J. Ib´a˜nez, J.L. Dideriksen, A.D. Koutsou, R. Paradiso, M.B. Popovic, J.M. Belda-Lois, F. Gianfelici, D. Farina, M. Manto, T. DAlessio, J.L. Pons. A soft wearable robot for tremor assessment and suppression. Proceedings of the 2011 IEEE International Conference on Robotics and Automation, 2249-54, 2011. •E. Rocon, J.A. Gallego, L. Barrios, A.R. Victoria, J. Ib´a˜nez, D. Farina, F. Negro, J.L. Dideriksen, S. Conforto, T. DAlessio, G. Severini, G. Grimaldi, M. Manto, J.L. Pons. Multimodal BCI-mediated FES suppression of tremor. Proceedings of 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 3337-40, 2010. •J.A. Gallego, E. Rocon, A.R. Victoria, J. Ib´a˜nez, L. Barrios, D. Farina, F. Negro, S. Conforto, T. DAlessio, G. Severini, G. Grimaldi, M. Manto, J.L. Pons. Brain Neural Computer Interface for tremor identification, characterization and tracking. Abstracts of the XVIII Congress of the International Society of Electrophysiology and Kinesiology, 2010 Abstract. 7.4 Future work The applied methods in this thesis and the achieved results are expected to serve as a starting point for future projects and research lines. Some of the topics considered for future research are the consequence of the results and conclusions reached in the here presented work, while others constitute a planned continuation in the framework of the addressed research line. Future studies identified here are organized in two groups associated with the two considered pathologies considered in this thesis. According to this division, future studies in line with the EEG classifier presented in Chapter 3 (EEG-based classification of upper-limb analytical movements) are integrated in the block of studies related to stroke, since such an application is expected to be of interest both to characterize 118 7.4 Future work the cortical status of patients with cortical reorganization and to develop advanced BCI systems capable of predicting the kind of movements that the patients are planning to perform during the rehabilitation sessions. Regarding the use of EEG and BCI technologies in the study and treatment of ET, the following future goals are identified, some of which have already been started at the time this document has been written: •Study the possible benefits of EEG-based neuromodulation systems in pathological tremors. In this line, previous experiments have been carried out by other research groups showing promising results when patients with tremor are able to modulate movement-related cortical activity [Fumuro et al., 2013]. •Develop new signal processing techniques for electrophysiological signals to improve the findings achieved with EEG and EMG technologies and with ET patients so far. Of special interest in this regard are the improvement of techniques studying interaction between neural populations and the development of new means of extracting the directionality of this interaction and the estimated delays. •Explore new technologies for the BNCI platform that allow bringing the proposed system into real-life conditions. This involves testing new acquisition technologies (modern dry electrodes, active electrodes with high signal-to-noise ratio...) and developing new signal processing techniques that allow filtering external sources of signal contamination that are present during daily living conditions. •Design new experimental protocols to increase our knowledge regarding tremor generation mechanisms in ET. In this line, studies using vibrotactile stimulation will be carried out in the near future in order to analyze the effects of periodic sensory afferences in tremor manifestation. Combination of EEG measurements with other technologies such as functional magnetic resonance imaging and magnetoencephalography systems are also expected to provide a more detailed description of tremor-related neural structures, by analysing the pathological tremor from different perspectives. •Look for more complex ways of combining the EMG and EEG information to improve the performance of the proposed mHRI. •Increase the number of recruited patients for the validation of the mHRI system and include patients with pathological tremor caused by other tremor-related diseases, such as Parkinson’s disease, or cerebellar tremor, so that the proposed platform can be robustly validated. Jaime Ib´a˜nez Pereda 119 Chapter 7. Conclusions and future work As for possible future research lines derived from the here presented studies with stroke patients, some of them are listed in the next lines: •To test online the EEG classifier of analytical upper-limb movements (presented in Chapter 3) on a large number of patients with stroke. This will allow the analysis of the extent to which the proposed system is able to describe altered cortical activation patterns and of the possibility of integrating the system in a BCI with other classification modules, leading to an improved characterization of the patients’ motor intentions while they perform rehabilitation tasks. •To advance in the development of the EEG-based BCI intervention for stroke patients, including robotic technologies able to cooperate with the neuroprosthetic device to produce a more natural movements during the rehabilitation. Besides, the adaptive control of the assistive forces delivered to the patients’ arms will be addressed in the future, so that an optimal movement generation is achieved. •To improve certain aspects of the EEG signal processing techniques used in order to make the BCI-based intervention suitable for clinical scenarios. In this line, it is identified as a relevant goal to look for ways to make the training data from a patients valid along different intervention sessions, which requires overcoming the inter-sessions variability of the EEG signal properties (due to changes in the electrode impedances, specific electrode locations on the scalp or patients’ vigilance). It is an additional goal to look for variations in the proposed system so that a reduced number of electrodes can still allow a robust detection of movement intentions. Future experiments will also seek to develop online artefact filtering techniques, so that patients can make use of this technology in a less restrictive way. All these advances will have as a final goal to adapt EEG-based BCI technology to the clinical scenario by reducing the time required for the interventions, allowing cost-effective EEG systems and allowing a proper function of the proposed system in electrically contaminated rooms (typically found in clinical environments). •To explore EEG sources of information that allow a fine characterization of the status of a patient with stroke and an accurate prognosis about the patient’s evolution both in terms of functional motor capacity and cortical activation patterns. It is expected that gaining knowledge in this regard will serve to develop interventions tailored to patients’ needs and to evaluate the efficacy of current interventions using longitudinal studies. •To explore the benefits that the BCI intervention presented here may provide to stroke patients in acute or subacute states. This kind of studies are of great interest, 120 BIBLIOGRAPHY [Cunnington et al., 1995] Cunnington, R., Iansek, R., Bradshaw, J. L., and Phillips, J. G. (1995). Movement-related potentials in Parkinson’s disease. Brain, 118(4):935–950. [Daly et al., 2006] Daly, J. J., Fang, Y., Perepezko, E. M., Siemionow, V., and Yue, G. H. (2006). Prolonged Cognitive Planning Time, Elevated Cognitive Effort , and Relationship to Coordination and Motor. Rehabilitation, 14(2):168–171. [Daly and Wolpaw, 2008] Daly, J. J. and Wolpaw, J. R. (2008). Brain-computer interfaces in neurological rehabilitation. The Lancet Neurology, 7(11):1032–1043. [Delgado Saa and Cetin, 2013] Delgado Saa, J. and Cetin, M. (2013). Discriminative Methods for Classification of Asynchronous Imaginary Motor Tasks From EEG Data. IEEE transactions on neural systems and rehabilitation engineering : a publication of the IEEE Engineering in Medicine and Biology Society, pages 1–8. [Deng et al., 2005] Deng, J., Yao, J., and Dewald, J. P. a. (2005). Classification of the intention to generate a shoulder versus elbow torque by means of a time-frequency synthesized spatial patterns BCI algorithm. Journal of neural engineering, 2(4):131–8. [Derambure et al., 1993] Derambure, P., Defebvre, L., Dujardin, K., Bourriez, J. L., Jacquesson, J. M., Destee, A., and Guieu, J. D. (1993). Effect of aging on the spatiotemporal pattern of event-related desynchronization during a voluntary movement. Electroencephalography and clinical neurophysiology, 89(3):197–203. [Desmurget et al., 2009] Desmurget, M., Sirigu, A., and Bernard, C. (2009). A parietal-premotor network for movement intention and motor awareness. Trends in cognitive sciences, 13(10):411–9. [Deuschl et al., 1998] Deuschl, G., Bain, P., and Brin, M. (1998). Consensus statement of the Movement Disorder Society on Tremor. Ad Hoc Scientific Committee. Movement disorders, 13 Suppl 3:2–23. [Deuschl et al., 2011] Deuschl, G., Raethjen, J., Hellriegel, H., and Elble, R. (2011). Treatment of patients with essential tremor. Lancet neurology, 10(2):148–61. [Dijksterhuis and Nordgren, 2006] Dijksterhuis, A. and Nordgren, L. F. (2006). A Theory of Unconscious Thought. Perspectives on Psychological Science, 1(2):95–109. [Elble, 2006] Elble, R. J. (2006). Report from a U.S. conference on essential tremor. Movement disorders, 21(12):2052–61. [Elble and Deuschl, 2009] Elble, R. J. and Deuschl, G. (2009). An update on essential tremor. Current neurology and neuroscience reports, 9(4):273–7. Jaime Ib´a˜nez Pereda 127 BIBLIOGRAPHY [Engel and Fries, 2010] Engel, A. K. and Fries, P. (2010). Beta-band oscillations– signalling the status quo? Current opinion in neurobiology, 20(2):156–65. [Fang et al., 2009] Fang, Y., Daly, J. J., Sun, J., Hvorat, K., Fredrickson, E., Pundik, S., Sahgal, V., and Yue, G. H. (2009). Functional corticomuscular connection during reaching is weakened following stroke. Clinical neurophysiology, 120(5):994–1002. [Fang et al., 2007] Fang, Y., Yue, G. H., Hrovat, K., Sahgal, V., and Daly, J. J. (2007). Abnormal cognitive planning and movement smoothness control for a complex shoulder/elbow motor task in stroke survivors. Journal of the neurological sciences, 256(12):21–9. [Farina et al., 2013] Farina, D., Negro, F., and Jiang, N. (2013). Identification of common synaptic inputs to motor neurons from the rectified electromyogram. The Journal of physiology, 591(Pt 10):2403–18. [Fatourechi et al., 2008] Fatourechi, M., Ward, R. K., and Birch, G. E. (2008). A selfpaced brain-computer interface system with a low false positive rate. Journal of neural engineering, 5(1):9–23. [Feigin et al., 2003] Feigin, V. L., Lawes, C. M., Bennett, D. A., and Anderson, C. S. (2003). Stroke epidemiology: a review of population-based studies of incidence, prevalence, and case-fatality in the late 20th century. The Lancet Neurology, 2(1):43–53. [Fumuro et al., 2013] Fumuro, T., Matsuhashi, M., Mitsueda, T., Inouchi, M., Hitomi, T., Nakagawa, T., Matsumoto, R., Kawamata, J., Inoue, H., Mima, T., Takahashi, R., and Ikeda, A. (2013). Bereitschaftspotential augmentation by neuro-feedback training in Parkinson’s disease. Clinical neurophysiology, 124(7):1398–405. [Gallego et al., 2010] Gallego, J. A., Rocon, E., Roa, J. O., Moreno, J. C., and Pons, J. L. (2010). Real-time estimation of pathological tremor parameters from gyroscope data. Sensors (Basel, Switzerland), 10(3):2129–49. [Garipelli et al., 2013] Garipelli, G., Chavarriaga, R., and del R Mill´an, J. (2013). Single trial analysis of slow cortical potentials: a study on anticipation related potentials. Journal of neural engineering, 10(3):036014. [Gerloff et al., 1998] Gerloff, C., Richard, J., Hadley, J., Schulman, a. E., Honda, M., and Hallett, M. (1998). Functional coupling and regional activation of human cortical motor areas during simple, internally paced and externally paced finger movements. Brain, 121 ( Pt 8:1513–31. 128 BIBLIOGRAPHY [Gomez-Rodriguez et al., 2010] Gomez-Rodriguez, M., Peters, J., Hill, J., Scho lkopf, B., Gharabaghi, A., and Grosse-Wentrup, M. (2010). Closing the sensorimotor loop: Haptic feedback facilitates decoding of arm movement imagery. In Systems Man and Cybernetics (SMC), 2010 IEEE International Conference on, pages 121–126. IEEE. [Graimann et al., 2002] Graimann, B., Huggins, J. E., Levine, S. P., and Pfurtscheller, G. (2002). Visualization of significant ERD/ERS patterns in multichannel EEG and ECoG data. Clinical Neurophysiology, 113(1):43–7. [Grosse-Wentrup et al., 2011] Grosse-Wentrup, M., Mattia, D., and Oweiss, K. (2011). Using brain-computer interfaces to induce neural plasticity and restore function. Journal of neural engineering, 8(2):025004. [Growdon et al., 1975] Growdon, J. H., Shahani, B. T., and Young, R. R. (1975). The effect of alcohol on essential tremor. Neurology, 25(3):259–62. [Gu et al., 2009a] Gu, Y., do Nascimento, O. F., Lucas, M.-F., and Farina, D. (2009a). Identification of task parameters from movement-related cortical potentials. Medical & biological engineering & computing, 47(12):1257–64. [Gu et al., 2009b] Gu, Y., Dremstrup, K., and Farina, D. (2009b). Single-trial discrimination of type and speed of wrist movements from EEG recordings. Clinical neurophysiology, 120(8):1596–600. [Gunal et al., 2000] Gunal, D. I., Afsar, N., Bekiroglu, N., and Aktan, S. (2000). New alternative agents in essential tremor therapy: double-blind placebo-controlled study of alprazolam and acetazolamide. Neurol Sci, 21(5):315–7. [Hall et al., 2010] Hall, S. D., Barnes, G. R., Furlong, P. L., Seri, S., and Hillebrand, A. (2010). Neuronal network pharmacodynamics of GABAergic modulation in the human cortex determined using pharmaco-magnetoencephalography. Human brain mapping, 31(4):581–94. [Halliday et al., 1998] Halliday, D. M., Conway, B. a., Farmer, S. F., and Rosenberg, J. R. (1998). Using electroencephalography to study functional coupling between cortical activity and electromyograms during voluntary contractions in humans. Neuroscience letters, 241(1):5–8. [Halliday et al., 1995] Halliday, D. M., Rosenberg, J. R., Amjad, a. M., Breeze, P., Conway, B. a., and Farmer, S. F. (1995). A framework for the analysis of mixed time series/point process data–theory and application to the study of physiological tremor, Jaime Ib´a˜nez Pereda 129 BIBLIOGRAPHY single motor unit discharges and electromyograms. Progress in biophysics and molecular biology, 64(2-3):237–78. [Hammon et al., 2008] Hammon, P., Makeig, S., Poizner, H., Todorov, E., and De Sa, V. (2008). Predicting Reaching Targets from Human EEG. IEEE Signal Processing Magazine, 25(1):69–77. [Hellwig et al., 2001] Hellwig, B., H¨auß ler, S., Schelter, B., Lauk, M., Guschlbauer, B., Timmer, J., and L¨ucking, C. (2001). Tremor-correlated cortical activity in essential tremor. The Lancet, 357:519–523. [Hellwig et al., 2000] Hellwig, B., H¨aussler, S., Lauk, M., Guschlbauer, B., K¨oster, B., Kristeva-Feige, R., Timmer, J., and L¨ucking, C. H. (2000). Tremor-correlated cortical activity detected by electroencephalography. Clinical neurophysiology, 111(5):806–9. [Hellwig et al., 2003] Hellwig, B., Schelter, B., Guschlbauer, B., Timmer, J., and L¨ucking, C. (2003). Dynamic synchronisation of central oscillators in essential tremor. Clinical Neurophysiology, 114(8):1462–1467. [Hinterberger et al., 2003] Hinterberger, T., K¨ubler, A., Kaiser, J., Neumann, N., and Birbaumer, N. (2003). A brain-computer interface (BCI) for the locked-in: comparison of different EEG classifications for the thought translation device. Clinical neurophysiology, 114(3):416–25. [Hjorth, 1975] Hjorth, B. (1975). An on-line transformation of EEG scalp potentials into orthogonal source derivations. Electroencephalogr Clin Neurophysiol, 39(5):526–30. [Hua et al., 1998] Hua, S. E., Lenz, F. a., Zirh, T. a., Reich, S. G., and Dougherty, P. M. (1998). Thalamic neuronal activity correlated with essential tremor. Journal of neurology, neurosurgery, and psychiatry, 64(2):273–6. [Huber and Paulson, 1988] Huber, S. J. and Paulson, G. W. (1988). Efficacy of alprazolam for essential tremor. Neurology, 38(2):241–3. [Hummel and Gerloff, 2005] Hummel, F. and Gerloff, C. (2005). Larger interregional synchrony is associated with greater behavioral success in a complex sensory integration task in humans. Cerebral Cortex, 15(5):670–8. [Ib´a˜nez et al., 2013] Ib´a˜nez, J., Serrano, J., del Castillo, M., Gallego, J. A., and Rocon, E. (2013). Online detector of movement intention based on EEG. Application in tremor patients. Biomedical Signal Processing and Control, 8(6):822–829. 130 BIBLIOGRAPHY [Ilan and Gevins, 2001] Ilan, A. B. and Gevins, A. (2001). Prolonged neurophysiological effects of cumulative wine drinking. Alcohol (Fayetteville, N.Y.), 25(3):137–52. [Jenkins et al., 1993] Jenkins, I. H., Bain, P. G., Colebatch, J. G., Thompson, P. D., Findley, L. J., Frackowiak, R. S., Marsden, C. D., and Brooks, D. J. (1993). A positron emission tomography study of essential tremor: evidence for overactivity of cerebellar connections. Annals of neurology, 34(1):82–90. [Jensen et al., 2005] Jensen, O., Goel, P., Kopell, N., Pohja, M., Hari, R., and Ermentrout, B. (2005). On the human sensorimotor-cortex beta rhythm: sources and modeling. NeuroImage, 26(2):347–55. [Jochumsen et al., 2013] Jochumsen, M., Niazi, I. K., Mrachacz-Kersting, N., Farina, D., and Dremstrup, K. (2013). Detection and classification of movement-related cortical potentials associated with task force and speed. Journal of neural engineering, 10(5):056015. [Kandel et al., 2000] Kandel, E., Schwartz, J., and Jessell, T. (2000). Principles of neural science, volume 3. McGraw-Hill, 4 edition. [Kaplan et al., 1998] Kaplan, G. B., Greenblatt, D. J., Ehrenberg, B. L., Goddard, J. E., Harmatz, J. S., and Shader, R. I. (1998). Single-dose pharmacokinetics and pharmacodynamics of alprazolam in elderly and young subjects. Journal of clinical pharmacology, 38(1):14–21. [Kilner et al., 2000] Kilner, J. M., Baker, S. N., Salenius, S., Hari, R., and Lemon, R. N. (2000). Human Cortical Muscle Coherence Is Directly Related to Specific Motor Parameters. J. Neurosci., 20(23):8838–8845. [Kinoshita et al., 2010] Kinoshita, M., Hitomi, T., Matsuhashi, M., Nakagawa, T., Nagamine, T., Sawada, H., Saiki, H., Shibasaki, H., Takahashi, R., and Ikeda, A. (2010). How does voluntary movement stop resting tremor? Clinical Neurophysiology, 121(6):983–5. [Kornhuber and Deecke, 1965] Kornhuber, H. H. and Deecke, L. (1965). Hirnpotential¨anderungen bei Willk¨urbewegungen und passiven Bewegungen des Menschen: Bereitschaftspotential und reafferente Potentiale. Pfl¨ugers Archiv f¨ur die Gesamte Physiologie des Menschen und der Tiere, 284(1):1–17. [Lau et al., 2004] Lau, H. C., Rogers, R. D., Haggard, P., and Passingham, R. E. (2004). Attention to intention. Science (New York, N.Y.), 303(5661):1208–10. Jaime Ib´a˜nez Pereda 131 BIBLIOGRAPHY [Lew et al., 2012] Lew, E., Chavarriaga, R., Silvoni, S., and Mill´an, J. R. (2012). Detection of self-paced reaching movement intention from EEG signals. Frontiers in neuroengineering, 5(July):13. [Liao et al., 2012] Liao, L.-D., Chen, C.-Y., Wang, I.-J., Chen, S.-F., Li, S.-Y., Chen, B.-W., Chang, J.-Y., and Lin, C.-T. (2012). Gaming control using a wearable and wireless EEG-based brain-computer interface device with novel dry foam-based sensors. Journal of neuroengineering and rehabilitation, 9:5. [Libet et al., 1982] Libet, B., Wright, E., and Gleason, C. (1982). Readiness-potentials preceding unrestricted ’spontaneous’ vs. pre-planned voluntary acts. Electroencephalography and Clinical Neurophysiology, 54(3):322–335. [Lindhardt et al., 2001] Lindhardt, K., Gizurarson, S., Stef´ansson, S. B., Olafsson, D. R., and Bechgaard, E. (2001). Electroencephalographic effects and serum concentrations after intranasal and intravenous administration of diazepam to healthy volunteers. British journal of clinical pharmacology, 52(5):521–7. [Louis et al., 2013] Louis, E. D., Babij, R., Cort´es, E., Vonsattel, J.-P. G., and Faust, P. L. (2013). The inferior olivary nucleus: A postmortem study of essential tremor cases versus controls. Movement disorders, 00(00):1–8. [Louis et al., 2007] Louis, E. D., Faust, P. L., Vonsattel, J.-P. G., Honig, L. S., Rajput, A., Robinson, C. a., Rajput, A., Pahwa, R., Lyons, K. E., Ross, G. W., Borden, S., Moskowitz, C. B., Lawton, A., and Hernandez, N. (2007). Neuropathological changes in essential tremor: 33 cases compared with 21 controls. Brain, 130(Pt 12):3297–307. [Louis et al., 1998] Louis, E. D., Ottman, R., and Hauser, W. A. (1998). How common is the most common adult movement disorder? estimates of the prevalence of essential tremor throughout the world. Movement disorders, 13(1):5–10. [Lu et al., 2010] Lu, M.-K., Jung, P., Bliem, B., Shih, H.-T., Hseu, Y.-T., Yang, Y.-W., Ziemann, U., and Tsai, C.-H. (2010). The Bereitschaftspotential in essential tremor. Clinical Neurophysiology, 121(4):622–30. [Magnani et al., 2002] Magnani, G., Cursi, M., Leocani, L., Volont´e, M. A., and Comi, G. (2002). Acute effects of L-dopa on event-related desynchronization in Parkinson’s disease. Neurological Sciences : official journal of the Italian Neurological Society and of the Italian Society of Clinical Neurophysiology, 23(3):91–7. [Magnani et al., 1998] Magnani, G., Cursi, M., Leocani, M. D. P. L., Volont´e, M. D. M. A., Locatelli, M. D. T., Elia, A., Comi, M. D. G., Leocani, L., Locatelli, T., and 132 BIBLIOGRAPHY Comi, G. (1998). Event-related Desynchronization to Contingent Negative Variation and Self-Paced Movement Paradigms in Parkinson’s Disease. Movement Disorders, 13(4):653–660. [Mason and Birch, 2000] Mason, S. G. and Birch, G. E. (2000). A Brain-Controlled Switch for Asynchronous Control Applications. IEEE Transactions on Biomedical Engineering, 47(10):1297–1307. [Mason et al., 2006] Mason, S. G., Kronegg, J., Huggins, J., Fatourechi, M., and Schl¨ogl, A. (2006). Evaluating the performance of self-paced brain computer interface technology. Neil Squire Soc., Vancouver, BC, Canada, Tech. Rep, 0. [McFarland et al., 2011] McFarland, D. J., Sarnacki, W. a., and Wolpaw, J. R. (2011). Should the parameters of a BCI translation algorithm be continually adapted? Journal of neuroscience methods, 199(1):103–7. [Minc et al., 2010] Minc, D., Machado, S., Bastos, V. H., Machado, D., Cunha, M., Cagy, M., Budde, H., Basile, L., Piedade, R., and Ribeiro, P. (2010). Gamma band oscillations under influence of bromazepam during a sensorimotor integration task: an EEG coherence study. Neuroscience letters, 469(1):145–9. [Moazami-Goudarzi et al., 2008] Moazami-Goudarzi, M., Sarnthein, J., Michels, L., Moukhtieva, R., and Jeanmonod, D. (2008). Enhanced frontal low and high frequency power and synchronization in the resting EEG of parkinsonian patients. NeuroImage, 41(3):985–97. [Morash et al., 2008] Morash, V., Bai, O., Furlani, S., Lin, P., and Hallett, M. (2008). Classifying EEG signals preceding right hand, left hand, tongue, and right foot movements and motor imageries. Clinical Neurophysiology, 119(11):2570–8. [Mrachacz-Kersting et al., 2013] Mrachacz-Kersting, N., Jiang, N., Dremstrup, K., and Farina, D. (2013). IS 22. Coupling of motor imagination and nervous system stimulation to induce cortical plasticity. Clinical Neurophysiology, 124(10):e46–e47. [Mrachacz-Kersting et al., 2012] Mrachacz-Kersting, N., Kristensen, S. R., Niazi, I. K., and Farina, D. (2012). Precise temporal association between cortical potentials evoked by motor imagination and afference induces cortical plasticity. The Journal of physiology, 590(Pt 7):1669–82. [M¨uller-Putz et al., 2005] M¨uller-Putz, G., Scherer, R., Pfurtscheller, G., and Rupp, R. (2005). EEG-based neuroprosthesis control: a step towards clinical practice. Neuroscience letters, 382(1-2):169–74. Jaime Ib´a˜nez Pereda 133 BIBLIOGRAPHY [M¨uller-Putz et al., 2007] M¨uller-Putz, G., Zimmermann, D., Graimann, B., Nestinger, K., Korisek, G., and Pfurtscheller, G. (2007). Event-related beta EEGchanges during passive and attempted foot movements in paraplegic patients. Brain research, 1137(1):84–91. [Muthuraman et al., 2012] Muthuraman, M., Heute, U., Arning, K., Anwar, A. R., Elble, R., Deuschl, G., and Raethjen, J. (2012). Oscillating central motor networks in pathological tremors and voluntary movements. What makes the difference? NeuroImage, 60(2):1331–9. [Nascimento, 2008] Nascimento, O. (2008). Movement-Related Cortical Potentials Allow Discrimination of Rate of Torque Development in Imaginary Isometric Plantar Flexion. Biomedical Engineering, IEEE, 55(11):2675–2678. [Negro and Farina, 2011] Negro, F. and Farina, D. (2011). Linear transmission of cortical oscillations to the neural drive to muscles is mediated by common projections to populations of motoneurons in humans. The Journal of physiology, 589(Pt 3):629–37. [Neuper et al., 2006] Neuper, C., W¨ortz, M., and Pfurtscheller, G. (2006). ERD/ERS patterns reflecting sensorimotor activation and deactivation. Progress in brain research, 159:211–22. [Niazi et al., 2013] Niazi, I. K., Jiang, N., Jochumsen, M., Nielsen, J. r. F. k., Dremstrup, K., and Farina, D. (2013). Detection of movement-related cortical potentials based on subject-independent training. Medical & biological engineering & computing, 51(5):507–12. [Niazi et al., 2011] Niazi, I. K., Jiang, N., Tiberghien, O., Nielsen, J. r. F. k., Dremstrup, K., and Farina, D. (2011). Detection of movement intention from single-trial movement-related cortical potentials. Journal of neural engineering, 8(6):066009. [Niazi et al., 2012] Niazi, I. K., Mrachacz-Kersting, N., Jiang, N., Dremstrup, K., and Farina, D. (2012). Peripheral electrical stimulation triggered by self-paced detection of motor intention enhances motor evoked potentials. IEEE transactions on neural systems and rehabilitation engineering, 20(4):595–604. [Niedermeyer, 2005] Niedermeyer, E. (2005). The normal EEG of the waking adult. In Niedermeyer, E. and da Silva, F. H. L., editors, Electroencephalography: Basic Principles, Clinical Applications and Related Fields, fifth edition, pages 167–192. Lippincott Williams and Wilkins, Philadelphia. 134 BIBLIOGRAPHY [Niedermeyer and da Silva, 2005] Niedermeyer, E. and da Silva, F. H. L. (2005). Electroencephalography: Basic Principles, Clinical Applications, and Related Fields. Lippincott Williams & Wilkins, 5th edition. [Nijboer et al., 2008] Nijboer, F., Sellers, E. W., Mellinger, J., Jordan, M. a., Matuz, T., Furdea, A., Halder, S., Mochty, U., Krusienski, D. J., Vaughan, T. M., Wolpaw, J. R., Birbaumer, N., and K¨ubler, A. (2008). A P300-based brain-computer interface for people with amyotrophic lateral sclerosis. Clinical neurophysiology, 119(8):1909–16. [Norman and Komi, 1979] Norman, R. W. and Komi, P. V. (1979). Electromechanical delay in skeletal muscle under normal movement conditions. Acta Physiologica Scandinavica, 106(3):241–248. [Nunez and Srinivasan, 2006] Nunez, P. L. and Srinivasan, R. (2006). Electric Fields of the Brain: The Neurophysics of EEG, volume 4. [Obhi et al., 2009] Obhi, S. S., Planetta, P. J., and Scantlebury, J. (2009). On the signals underlying conscious awareness of action. Cognition, 110(1):65–73. [Oby et al., 2013] Oby, E. R., Ethier, C., and Miller, L. E. (2013). Movement representation in the primary motor cortex and its contribution to generalizable EMG predictions. Journal of neurophysiology, 109(3):666–78. [Onose et al., 2012] Onose, G., Grozea, C., Anghelescu, A., Daia, C., Sinescu, C. J., Ciurea, A. V., Spircu, T., Mirea, A., Andone, I., Spanu, A., Popescu, C., Mihaescu, A.-S., Fazli, S., Danoczy, M., and Popescu, F. (2012). On the feasibility of using motor imagery EEG-based brain-computer interface in chronic tetraplegics for assistive robotic arm control: a clinical test and long-term post-trial follow-up. Spinal cord, 50(8):599– 608. [Pahapill et al., 1999] Pahapill, P. A., Levy, R., Dostrovsky, J. O., Davis, K. D., Rezai, A. R., Tasker, R. R., and Lozano, A. M. (1999). Tremor arrest with thalamic microinjections of muscimol in patients with essential tremor. Annals of neurology, 46(2):249–52. [Park et al., 2010] Park, Y.-G., Park, H.-Y., Lee, C. J., Choi, S., Jo, S., Choi, H., Kim, Y.-H., Shin, H.-S., Llinas, R. R., and Kim, D. (2010). Ca(V)3.1 is a tremor rhythm pacemaker in the inferior olive. Proceedings of the National Academy of Sciences of the United States of America, 107(23):10731–6. [Petersen et al., 2012] Petersen, T. H., Willerslev-Olsen, M., Conway, B. a., and Nielsen, J. B. (2012). The motor cortex drives the muscles during walking in human subjects. The Journal of physiology, 590(Pt 10):2443–52. Jaime Ib´a˜nez Pereda 135 BIBLIOGRAPHY [Pfurtscheller and Berghold, 1989] Pfurtscheller, G. and Berghold, A. (1989). Patterns of cortical activation during planning of voluntary movement. Electroencephalography and Clinical Neurophysiology, 72(3):250–8. [Pfurtscheller et al., 2006] Pfurtscheller, G., Brunner, C., Schl¨ogl, A., da Silva, F. H. L., and Lopes da Silva, F. H. (2006). Mu rhythm (de)synchronization and EEG single-trial classification of different motor imagery tasks. NeuroImage, 31(1):153–9. [Pfurtscheller and da Silva, 1999] Pfurtscheller, G. and da Silva, F. H. L. (1999). Event-related EEG/EMG Synchronization and Desynchronization: Basic Principles. Clinical Neurophysiology, 110:1842–1857. [Pfurtscheller et al., 2003] Pfurtscheller, G., Graimann, B., Huggins, J. E., Levine, S. P., and Schuh, L. A. (2003). Spatiotemporal patterns of beta desynchronization and gamma synchronization in corticographic data during self-paced movement. Clinical neurophysiology, 114(7):1226–36. [Pfurtscheller and Solis-Escalante, 2009] Pfurtscheller, G. and Solis-Escalante, T. (2009). Could the beta rebound in the EEG be suitable to realize a ”brain switch”? Clinical Neurophysiology, 120(1):24–9. [Pichiorri et al., 2011] Pichiorri, F., De Vico Fallani, F., Cincotti, F., Babiloni, F., Molinari, M., Kleih, S. C., Neuper, C., K¨ubler, A., and Mattia, D. (2011). Sensorimotor rhythm-based brain-computer interface training: the impact on motor cortical responsiveness. Journal of neural engineering, 8(2):025020. [Pinker and Jackendoff, 2005] Pinker, S. and Jackendoff, R. (2005). The faculty of language: what’s special about it? Cognition, 95(2):201–36. [Platz, 2000] Platz, T. (2000). Multimodal EEG analysis in man suggests impairmentspecific changes in movement-related electric brain activity after stroke. Brain, 123(12):2475–2490. [Popescu et al., 2007] Popescu, F., Fazli, S., Badower, Y., Blankertz, B., and M¨uller, K. (2007). Single trial classification of motor imagination using 6 dry EEG electrodes. PloS one, 2(7):e637. [Raethjen et al., 2008] Raethjen, J., Austermann, K., Witt, K., Zeuner, K. E., Papengut, F., and Deuschl, G. (2008). Provocation of Parkinsonian tremor. Movement disorders, 23(7):1019–23. 136