Assist-as-needed control for a smart orthotic system aiming a personalized gait rehabilitation

Universidade do Minho Escola de Engenharia Luís Carlos Rodrigues Moreira Assist-As-Needed Control for a Smart Orthotic System Aiming a Personalized Gait Rehabilitation Outubro de 2024 UMinho | 2024 Luís Carlos Rodrigues Moreira Assist-As-Needed Control for a Smart Orthotic System Aiming a Personalized Gait Rehabilitation i i Luís Carlos Rodrigues Moreira Assist-As-Needed Control for a Smart Orthotic System Aiming a Personalized Gait Rehabilitation Tese de Doutoramento Programa Doutoral em Engenharia Biomédica Trabalho efetuado sob a orientação de Professora Doutora Cristina Peixoto dos Santos Professor Doutor João José Fernandes Cardoso de Araújo Cerqueira Doutora Joana Sofia Campos Figueiredo outubro de 2024 ii DIREITOS DE AUTOR E CONDIÇÕES DE UTILIZAÇÃO DO TRABALHO POR TERCEIROS Este é um trabalho académico que pode ser utilizado por terceiros desde que respeitadas as regras e boas práticas internacionalmente aceites, no que concerne aos direitos de autor e direitos conexos. Assim, o presente trabalho pode ser utilizado nos termos previstos na licença abaixo indicada. Caso o utilizador necessite de permissão para poder fazer um uso do trabalho em condições não previstas no licenciamento indicado, deverá contactar o autor, através do RepositóriUM da Universidade do Minho. Atribuição-NãoComercial-SemDerivações CC BY-NC-ND https://creativecommons.org/licenses/by-nc-nd/4.0/ iii AGRADECIMENTOS Após o término deste importante capítulo, é hora de agradecer. Em primeiro lugar, quero expressar a minha profunda gratidão aos meus pais. Eles proporcionaram-me apoio, carinho, amor, conforto, estrutura e a força emocional necessária ao longo de toda esta caminhada. São o meu porto-seguro e todo o sucesso académico, profissional e pessoal deve-se a eles. Agradeço também à minha família e família da minha esposa pelo orgulho demonstrado em cada etapa do meu trabalho e pelas conquistas alcançadas. Um agradecimento aos meus orientadores (Professora Cristina, Doutora Joana Figueiredo e Doutor João Cerqueira) e ao corpo clínico do Hospital de Braga. À Professora Cristina, a minha gratidão por me proporcionar a oportunidade de trabalhar num projeto tão enriquecedor e desafiador, do qual me orgulho profundamente. Agradeço pela orientação, disponibilidade, incentivo constante e por alimentar a minha capacidade de sonhar. À Doutora Joana Figueiredo, agradeço pela empatia, amizade, partilha e crescimento pessoal e profissional. A tua dedicação e amor pela causa são evidentes em tudo o que fazes. Obrigado pelos valores que me transmitiste e por todo o apoio disponibilizado ao longo desta jornada. Aos meus colegas de laboratório, agradeço pelos momentos de partilha, aprendizagem, disponibilidade e pelo companheirismo. Um agradecimento especial aos amigos do Porta Aberta, que tornaram esta jornada doutoral mais descontraída, repleta de alegrias, e temperada com o sabor das terras que nos viram crescer. Agradeço também a todos os participantes dos estudos experimentais que possibilitaram a validação das tecnologias desenvolvidas nesta tese. Em particular, o meu sincero agradecimento ao Sr. João e à Margarida (e correspondentes famílias) pela dedicação, paciência, interesse e disponibilidade demonstrados nas terapias conduzidas. Juntos, fizemos ciência! Um agradecimento ao melhor naipe de clarinetes de Portugal (Palhetas D’Ouro) e à Banda Musical de Rio de Moinhos por todos os momentos musicais e de confraternização. À minha esposa, Mariana Costa, agradeço por tudo e por tanto. A tua presença fez desta jornada algo ainda mais especial e bonito. Ajudaste-me a nunca perder o foco e foste essencial para traçar e concretizar o caminho possível que me levou a atingir este objetivo. Obrigado a todos! Juntos, contribuímos para o avanço da ciência! iv STATEMENT OF INTEGRITY I hereby declare having conducted this academic work with integrity. I confirm that I have not used plagiarism or any form of undue use of information or falsification of results along the process leading to its elaboration. I further declare that I have fully acknowledged the Code of Ethical Conduct of the University of Minho. v Controlo de assistência conforme as necessidades para um sistema ortótico inteligente com vista à reabilitação personalizada da marcha RESUMO O acidente vascular cerebral (AVC) corresponde à terceira principal causa de incapacidade motora em adultos. Através de terapias de reabilitação envolvendo um treino repetitivo e personalizado ao doente, a reabilitação da marcha é essencial para a recuperação da mobilidade. Para tal, as ortóteses ativas (OAs) e exosqueletos têm sido apontados como um meio para potenciar os efeitos das terapias de reabilitação. Contudo, é crucial adaptar a reabilitação da marcha às necessidades e intenções de locomoção de cada doente, através do desenvolvimento de (i) novos sensores e algoritmos inteligentes que permitam uma avaliação objetiva da marcha; e (ii) estratégias de controlo para OAs que promovam uma terapia adaptada às necessidades de cada doente, isto é, assist-as-needed (AAN). Esta tese tem como objetivo expandir o SmartOs, um sistema ortótico ativo, modular e vestível, para oferecer um treino de marcha adaptado às necessidades de doentes com AVC e avaliar a locomoção através de dados cinemáticos e musculares. A tese é estruturada em cinco passos de investigação. Primeiro, foi expandida a estrutura modular do SmartOs para integrar novos sistemas sensoriais, ferramentas de análise da marcha e estratégias de controlo. Segundo, foi criado um sistema sensorial baseado em têxteis condutores, capaz de monitorizar, simultaneamente, a atividade muscular de vários grupos musculares. O benchmarking com sistemas comerciais demonstrou o potencial deste sistema para a avaliação objetiva da contração muscular. Terceiro, foi desenvolvida uma ferramenta de deep learning para a descodificação de modos de locomoção, destacando-se pela sua classificação precisa e antecipativa. Quarto, foram desenvolvidas três estratégias de controlo AAN. A estratégia baseada em modos de locomoção oferece assistência de acordo com as intenções de locomoção do utilizador; a estratégia baseada em eletromiografia contribui para o fortalecimento muscular e promove a participação ativa do utilizador na terapia; e a estratégia baseada em energia visa adaptar o controlo do dispositivo para reduzir o custo metabólico do utilizador. Por fim, foram conduzidos dois casos de estudo para determinar os efeitos do sistema SmartOs como ferramenta de reabilitação em pacientes que sofreram um AVC. Em suma, os resultados indicam que o sistema SmartOs está apto para aplicação em ambiente clínico, tanto como uma solução personalizada de assistência, como uma ferramenta de avaliação da marcha em pacientes que sofreram um AVC. Palavras-Chave: assistência e reabilitação da marcha, descodificação de modos de locomoção, estratégias de controlo, exosqueletos, inteligência artificial, sensores vestíveis. vi ABSTRACT Stroke is the third leading cause of motor disability in adults. Through rehabilitation therapies involving repetitive and user-oriented training, gait rehabilitation is essential to regain mobility. To this end, active orthoses (AOs) and exoskeletons have been identified as a means to enhance the effects of rehabilitation therapies. However, it is crucial to adapt gait rehabilitation to the specific needs and locomotion intentions of each post-stroke patient through the development of (i) new sensors and intelligent algorithms that allow an objective assessment of gait; and (ii) control strategies for AO that promote therapy tailored to the needs of each post-stroke patient, i.e., assist-as-needed (AAN). This Ph.D. thesis aims to extend SmartOs, an active, modular, and wearable orthotic system, to provide gait training adapted to the needs of post-stroke patients and to assess locomotion through kinematic and muscular data. The thesis is divided into five research steps. Firstly, the modular structure of SmartOs was extended to integrate new sensory systems, gait analysis tools, and control strategies. Secondly, a sensory system based on conductive textiles was developed, capable of simultaneously monitoring the muscular activity of several muscle groups. Benchmarking with commercial systems demonstrated the potential of this system for objective assessment of muscle contraction. Thirdly, a deep learning tool was developed to decode locomotion modes, which is characterized by its accurate classification and predictive nature. Fourth, three AAN control strategies were developed. The AAN Locomotion Mode-driven trajectory control provides assistance according to the user's locomotion intentions. The AAN EMG-based control contributes to muscle strengthening and promotes the user's active participation in therapy. The AAN Human-in-the-Loop control aims to adapt the control of the device to reduce the user's energy expenditure. Finally, two case studies were conducted to determine the effects of the SmartOs system as a rehabilitation tool for post-stroke patients. In short, the results indicate that the SmartOs system is suitable for use in a clinical environment, both as a personalized assistance solution and as a gait assessment tool for patients who have suffered a stroke. Keywords: Gait assistance and rehabilitation, locomotion mode decoding, control strategies, exoskeletons, artificial intelligence, wearable sensors vii Table of Contents List of Figures ...................................................................................................................x List of Tables .................................................................................................................. xiv List of abbreviations and acronyms ................................................................................ xvi 1. Introduction .............................................................................................................. 1 1.1 Motivation ................................................................................................................... 2 1.2 Problem statement ....................................................................................................... 4 1.3 Goals ........................................................................................................................... 5 1.4 Research questions ...................................................................................................... 8 1.5 Contributions to knowledge.......................................................................................... 9 1.6 Publications ............................................................................................................... 11 1.7 Manuscript outline ..................................................................................................... 12 2. Literature Research ................................................................................................ 14 2.1 Introductory Insight ................................................................................................... 15 2.2 Review of the Effects of Lower Limb Assistive Devices on Post-Stroke Patients ........... 17 2.2.1 Search Methodology ................................................................................................................. 17 2.2.2 Results ..................................................................................................................................... 18 2.2.3 Discussion ................................................................................................................................ 36 2.3 Conclusions ............................................................................................................... 39 3. SmartOs System Overview ...................................................................................... 40 3.1 Introductory Insight ................................................................................................... 41 3.2 Conceptual Design and Functionalities ....................................................................... 41 3.2.1 Actuators .................................................................................................................................. 43 3.2.2 Wearable Motion Lab ................................................................................................................ 43 3.2.3 Gait Analysis Tools .................................................................................................................... 45 3.2.4 Hierarchical Control Architecture ............................................................................................... 46 3.2.5 Graphical Application ................................................................................................................ 47 3.3 Conclusions ............................................................................................................... 49 4. MuscLab System ..................................................................................................... 50 4.1 Introductory Insight ................................................................................................... 51 4.2 Critical Analysis of Related Work ................................................................................ 52 4.3 Methodology .............................................................................................................. 54 xiv List of Tables Table 2.1 – Technical details extracted from the selected studies (A, P, and NI means active, passive, and not indicated, respectively) ......................................................................................................... 21 Table 2.2 – Clinical details extracted from the selected studies .......................................................... 30 Table 3.1 – Different types of data collected by each sensor system .................................................. 44 Table 3.2 – Different control strategies available in the SmartOs system ............................................ 47 Table 4.1 – MuscLab requirements................................................................................................... 54 Table 4.2 – Average delay (standard deviation) between the MuscLab and EMG Delsys signals, and MuscLab and Xsens Awinda signals, in milliseconds, for each motion cadence (40, 70, 105 bpm) .... 66 Table 4.3 – Spearman Correlation coefficient ( r ) between the MuscLab and EMG systems. A p -value below 0.05 was verified in all correlations ................................................................................................... 67 Table 5.1 – Identification of the type of data used to decode LMs. The joint angles, magnitude of 3D raw accelerometer and gyroscope, segment angles, and EMG data are shaded in green, yellow, gray, and red, respectively ...................................................................................................................................... 78 Table 5.2 – Identification of the conditions for real-time tests performed by the able-bodied participants and a post-stroke patient .................................................................................................................. 83 Table 5.3 – LOSOCV metrics for all input combinations and DL models. The best and the worst results are colored in blue and red, respectively. The highest performance achieved is shaded in green color 85 Table 5.4 – Offline test metrics ......................................................................................................... 88 Table 5.5 – Online test metrics for able-bodied participants ............................................................... 89 Table 5.6 – Online test metrics for the stroke patient ........................................................................ 89 Table 5.7 – Success rate (%) per LM ................................................................................................. 90 Table 5.8 – Average prediction time in milliseconds (orange) and as a percentage of the gait cycle (blue) ........................................................................................................................................................ 91 Table 6.1 – Online test metrics for able-bodied participants ............................................................. 104 Table 6.2 – Average update time in advance in milliseconds (orange) and as a percentage of the gait cycle (blue) ..................................................................................................................................... 105 Table 6.3 – DL models performance when EMG signals were/were not included as inputs .............. 116 Table 6.4 – RMSE, NMSE, and r metrics for LOSOCV procedure ..................................................... 118 Table 6.5 – RMSE, NMSE, and r during the LOSOCV and model test procedures ............................. 119 Table 6.6 – CNN’s computational load ............................................................................................ 121 xv Table 6.7 – Variation of the EMG signals from tibialis anterior and gastrocnemius lateralis , range of motion (ROM) of the hip joint, and human ankle joint torque between unconditioned and conditioned tasks during AAN EMG-based control strategy (a negative and a positive sign means a reduction and an increase, respectively, when comparing the variable measured at the conditioned task to the unconditioned task) ...................................................................................................................................................... 124 Table 6.8 – Variation of the dorsiflexion and plantar flexion motor torques between unconditioned and conditioned tasks (a negative and a positive sign means a reduction and an increase, respectively, when comparing the variable measured at the conditioned task to the unconditioned task). Maximum dorsiflexion, and plantar flexion angles measured at the conditioned task ........................................ 125 Table 6.9 – Variation of the EMG signals from tibialis anterior and gastrocnemius lateralis , and dorsiflexion and plantar flexion motor torques between unconditioned and conditioned tasks (a negative and a positive sign means a reduction and an increase, respectively, when comparing the variable measured at the conditioned task to the unconditioned task). Maximum dorsiflexion, and plantar flexion angles measured at the conditioned task ................................................................................................................... 126 Table 7.1 – Characteristics of the recruited post-stroke patients ...................................................... 145 Table 7.2 – Procedure adopted during pre-, post-training, and follow-up sessions............................. 147 Table 7.3 – Procedure adopted during the familiarization session with the SmartOs system ............. 147 Table 7.4 – Procedure adopted during intervention sessions with the SmartOs system .................... 151 xvi List of abbreviations and acronyms A AAN Assist-As-Needed ACC Accuracy ADC Analog-to-Digital Converter AI Artificial Intelligence AO Active Orthosis API Application Programming Interface B BBS Berg Balance Scale BI Bartel Index C CAN Control Area Network CCU Central Control Unit CNN Convolutional Neural Network CPG Control Pattern Generator D DL Deep Learning E EGPR Exponential Gaussian Process Regression EMG Electromyography F FAC Functional Ambulation Classification FEL Feedback-Error Learning FIM Functional Independence Measure 5MWT 5-Meter Walk Test FMA-LE Fugl-Meyer Assessment – Lower Extremity FSR Force Sensing Resistor H HAI Hauser Ambulation Index HIL Human-in-the-Loop xvii I IMU Inertial Measurement Unit K KPI Key Performance Indicator L LGW Level-Ground Walking LM Locomotion Mode LOSOCV Leave-One-Subject-Out Cross-Validation LSTM Long Short-Term Memory M MAS Modified Ashworth Scale MCC Matthew’s Correlation Coefficient MI Motricity Index MMG Mechanomyography MMSE Mini-Mental State Examination MRCS Medical Research Council Scale MSD Musculoskeletal Disorder MSE Mean Square Error MVC Maximum Voluntary Contraction N NIHSS National Institute of Health Stroke Scale O OB Objective P PAFO Powered Ankle-Foot Orthosis PID Proportional Integral Derivative PKO Powered Knee Orthosis R r Spearman Correlation coefficient R2 Coefficient of Determination RA Ramp Ascent xviii RD Ramp Descent RMI Rivermead Mobility Index RMS Root Mean Square RMSE Root Mean Square Error ROM Range of Motion RQ Research Question S SA Stair Ascent SD Stair Descent SIAS Stroke Impairment Assessment Set Sit Sitting 6MWT 6-Minute Walk Test SmartOs SMARt control of a sTand-alone active Orthotic System St Standing STD Standard Deviation T 10MWT 10-Meter Walk Test TCP/IP Transmission Control Protocol/Internet Protocol TCT Trunk Control Test TRL Technology Readiness Level TUG Timed Up and Go Test 2MWT 2-Minute Walk Test U UART Universal Asynchronous Receiver/Transmitter 1 1. INTRODUCTION Chapter 1 2 This Ph.D. thesis presents the research developed in the last four years in the scope of the doctoral program in Biomedical Engineering. The investigation was developed in the Biomedical Robotic Devices Laboratory (BiRDLAB) at the Center for MicroElectroMechanical Systems (CMEMS) established at the University of Minho, together with the Centro Clínico Académico de Braga (2CA – Braga), a partnership between the University of Minho and Braga Hospital, which provides the expertise, equipment, facilities, and access to patients and clinicians. The developed activities have been included in the project SmartOs – SMARt control of a sTandalone active Orthotic System (POCI-01-0247-FEDER-039868). In addition, this research extends beyond a research grant (UMINHO/BI/83/2020), a Master’s thesis [1], and a Ph.D. thesis [2] by proposing Assist-As-Needed (AAN) control strategies for a wearable orthotic system in an attempt to restore the functional lower limb motor abilities of stroke survivors. This Ph.D. thesis addresses the nexus of assistive control strategies, wearable assistive devices, artificial intelligence (AI), and human gait analysis to personalize the assistance of the wearable orthotic system according to the user’s needs and locomotion intentions to promote post-stroke recovery. 1.1 MOTIVATION Walking is one of the most daily performed human motor tasks. However, human gait can be compromised due to neurological diseases, such as stroke. According to the World Health Organization, stroke events correspond to the second leading cause of death and the third leading cause of disability globally [3]. Studies have reported around 13.7 million new incident cases of stroke in the world in 2016 (between 121 and 151 per 100.000 people only in Portugal) [4]. The annual economic cost of stroke in Europe is around 60 billion euros, representing 0.36% of the European Gross Domestic Product (nearly 0.40% for Portugal) [5], [6]. Based on the study [7], more than 80% of stroke survivors present gait dysfunction due to muscle weakness, asymmetrical gait pattern, pain, spasticity, or a loss of motor control of the lower limbs. Consequently, the patient’s quality of life is affected since they cannot perform their daily locomotion activities (e.g., walking, running, standing, sitting, turning, and climbing stairs/ramps), or their accomplishment is difficult. These limitations commonly result in social and work exclusion, costly medical assistance, and early retirement [8]. There is an extreme necessity for improving the quality of life of post-stroke patients. Conventional physical rehabilitation provided by physiotherapists has been widely applied to face long- Chapter 1 3 term motor disabilities of these neurologically injured patients [9]. Nonetheless, lower limb rehabilitation in the last years has acknowledged the need to deal with (i) the disadvantages associated with the interand intra-therapist variances; (ii) the dependency on the malleability of the patient’s joint (commonly affected by spasticity); and, (iii) the absence of precise, user-oriented, and personalized assistance during therapy [10]. In addition, recent studies have identified the costs of rehabilitating patients who have survived a stroke [11], [12]. These costs are divided into inpatient and outpatient rehabilitation costs. Inpatients are patients who stay in the hospital and receive medical treatment, as well as food and accommodation in a hospital. Outpatients are patients who do not require hospitalization. An outpatient visits a hospital, clinic, or similar facility for a diagnosis, treatment, or procedure and may leave. According to the study [11], the cost of rehabilitation therapies for outpatients is typically three times lower than for inpatients. While the largest contributor to inpatient rehabilitation costs is the cost of stay (34%), the second largest contributor is the cost of physical rehabilitation (16%). In the case of outpatients, physical rehabilitation corresponds to the largest contributor to the rehabilitation costs (26%). Considering the mentioned topics, in the rehabilitation area, robotic assistance driven by wearable assistive devices, such as exoskeletons and active orthoses (AOs), has steadily gained importance [10]. According to the study [2], rehabilitation therapies driven by exoskeletons may improve the patient’s muscular strength, movement coordination, and balance control, fostering the patient’s ambulation and performance for successful locomotion. Accordingly, the use of wearable assistive devices as a complementary rehabilitation tool for conventional physical rehabilitation strategies may empower the long-term functional motor recovery of patients with lower limb impairments, fostering their confidence and independence by providing daily assistance [13]– [18]. Furthermore, when it comes to comparing the cost of rehabilitation between conventional therapy and robotics-guided therapy, a recent study concluded that in 80% of the cases analyzed, roboticsguided therapy was more cost-effective than conventional therapy [12]. Although some wearable assistive devices can cost several hundred euros, their use in rehabilitation therapies can increase the efficiency of the therapist's work, meaning that more patients can be treated, leading to an overall reduction in the cost of treatment per patient. Chapter 1 4 1.2 PROBLEM STATEMENT Despite becoming a prominent intervention to tackle the necessity of physical rehabilitation therapies, most of the available exoskeletons present poor usability and do not deliver personalized assistance according to the patient’s needs and intentions to move [19]–[26]. Additional imperative developments include exploring variants of AAN control strategies to face the limitations of the typically adopted trajectory-tracking control strategies [27]–[31]. Despite the repetitive nature of gait training imposed by trajectory-tracking control strategies, they tend to ignore the human-robot interaction and the needs of each user. As a result, these strategies may limit motor relearning and end up being abandoned [27], [28]. The development of adaptive and compliant control strategies is fundamental to timely personalize assistive trajectories according to the user’s motor needs, and consequently, assist the users as much and when needed. At this level, AAN control strategies are needed to dynamically adjust the level of assistance based on the user's real-time muscular performance and participation ability, ensuring that the wearable assistive device assists movement without overcompensating. The goal is to encourage the user's active participation and muscle engagement, thereby promoting improvements in strength and motor function over time [29]–[31]. In order to capture the user’s muscular performance and participation ability, electromyography (EMG) signals have been employed in AAN EMG-based control strategies. Although promising, most of the AAN EMG-based control strategies were developed for upper limbs [32]–[34]. And those developed for the lower limbs were not designed to assist the entire gait cycle, but rather focus on isolated flexion and extension movements [28], [35]–[38]. In addition, current challenges in personalized robotics-based assistance are related to decoding different locomotion modes (LMs) with a non-intrusive sensor setup to timely trigger the assistance delivered by wearable assistive devices according to the user’s locomotion intentions (i.e., AAN LM-driven controls). Despite recent advancements, most of the current AAN LM-driven controls integrated into wearable assistive devices (i) address a limited number of daily LMs (are non-generic tools) [19], [39]; (ii) present high recognition delays, classifying and assisting the LM only after the user has already transitioned to the new LM [20], [21], [40]–[44]; (iii) do not account for the typically slower gait speeds of individuals with lower limb disabilities, since studies addressed self-selected and/or fixed speeds of healthy participants (typically above 2.7 km/h) [20], [21], [45], [46]; and (iv) do not present clinical evidence [19]–[22]. It is of the utmost importance that wearable assistive devices tackle these limitations by including algorithms and AAN LM-driven controls capable of accurately and timely decoding Chapter 1 5 different LMs to provide personalized assistance. If the user’s LMs cannot be properly welldetermined, the exoskeleton assistance may be affected [42], [47]. Furthermore, recent variations of AAN approaches include human-in-the-loop (HITL) control strategies, using the user's energy expenditure as a means of adapting the assistance provided by the wearable assistive device [31], [48]. By tailoring the assistance to the user's current physical state and activity level, these control strategies aim to improve overall comfort, performance, and endurance during the wearable assistive device’s use [31], [49]–[51]. However, despite the promising performance, the available method for estimating energy expenditure is indirect calorimetry, which uses non-portable equipment, is time-consuming, produces noisy estimates, and is impractical for real-world applications [52]. In addition, most wearable assistive devices present a poor clinical evidence base, characterized by a lack of studies evaluating the longitudinal and follow-up effects of robotic therapies. There is a high number of studies that do not follow a randomized controlled procedure and do not comprehensively assess the effects of wearable assistive devices at kinematic, physiological, and functional levels [53]– [60]. This Ph.D. thesis intends to tackle the challenges mentioned above using a smart orthotic system – SmartOs system, in an attempt to provide new insights and innovative research directions for post-stroke rehabilitation using wearable assistive devices. 1.3 GOALS This innovative multidisciplinary Ph.D. thesis focuses on developing AAN control strategies and their integration into the SmartOs system, to promote personalized assistance automatically adapted according to the patient’s physiological needs and locomotion intentions. For this purpose and considering the main goal of this thesis, there are six objectives to be pursued: • Objective 1: To review the effects of lower limb assistive devices in poststroke gait rehabilitation. This review aims to identify the wearable assistive devices already applied in the rehabilitation of post-stroke patients, as well as the joints typically assisted, the sensors used, and the control strategies employed. It also focuses on extracting information on the characterization of post-stroke patients, study design, protocols used, and clinical outcomes. The review concludes by highlighting the benefits of using wearable assistive devices in rehabilitation sessions for post-stroke patients. The Chapter 1 12 Conference Papers • Luís Moreira, Roberto Martins Barbosa, Joana Figueiredo, Pedro Fonseca, João Paulo Vilas-Boas, Cristina P. Santos, “Real-Time Torque Estimation Using Human and Sensor Data Fusion for Exoskeleton Assistance”, The Sixth Iberian Robotics Conference (ROBOT2023), Coimbra, 2023. • Luís Moreira, Joana Figueiredo, João Cerqueira, Cristina P. Santos, “A Real-time Kinematic-based Locomotion Mode Prediction Algorithm for an Ankle Orthosis”, 24th IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC), Paredes de Coura, 2024. • Luís Moreira, Joana Figueiredo, João Cerqueira, Cristina P. Santos, “Assist-As-Needed Electromyography-based Control for a Wearable Ankle Robotic Orthosis”, IEEE RAS EMBS 10th International Conference on Biomedical Robotics and Biomechatronics (BioRob 2024), Heidelberg, 2024. 1.7 MANUSCRIPT OUTLINE This Ph.D. thesis is organized into eight chapters, as illustrated in Figure 1.2. Chapter 2 provides a comprehensive review of the clinical evidence of using lower limb assistive devices in post-stroke gait rehabilitation. It covers the joints typically assisted, the control strategies employed, the sensors used, post-stroke patient characterization, study design, protocols, and both sensor-based and clinical outcomes. Chapter 3 details the design of the SmartOs system, highlighting the proposed advances in the architecture developed in the previous study [2]. This includes the integration of a new wearable sensor system, gait analysis tools, and AAN control strategies into the SmartOs framework. Chapter 4 presents the development and validation of a wearable sensor system (the MuscLab system) to monitor muscle contraction, utilizing e-textile sensors embedded in a flexible and elastic wearable band. Chapter 5 focuses on the development, integration, and validation of an LM decoding tool to decode four LMs (St, LGW, SA, and SD), in real-time. It also includes a benchmark analysis to identify the best fusion of sensor features and DL algorithms for effective decoding of daily LMs. Chapter 6 presents the three proposed AAN control strategies, detailing the design, methodology, and validation of each controller developed for the SmartOs system. Chapter 7 introduces the clinical protocol developed and states the physiological, kinematic, and functional effects of using the SmartOs system in rehabilitation therapies for post-stroke Chapter 1 13 patients. Chapter 8 summarizes the main findings and conclusions of the Ph.D. thesis, together with directions for future research and opportunities for technical improvement. Figure 1.2 – Ph.D. thesis organization. OB. and RQ. represent objective and research question, respectively. Chapter : Introduction Chapter : Literature Research Chapter : SmartOs System Overview Chapter : MuscLab System Chapter : Locomotion Mode Decoding Tool Chapter : Assist As Needed Control Strategies Chapter : SmartOs Clinical alidation Chapter : Conclusions 14 2. LITERATURE RESEARCH Chapter 2 15 2.1 INTRODUCTORY INSIGHT Stroke is one of the leading causes of motor disability in adults, and those affected often face profound and lasting neurological consequences [3]. According to the study [68], 50% of post-stroke patients are initially unable to walk, 12% can walk with assistance, and 38% can walk independently. Poststroke patients who are able to walk, experience a gait pattern that is commonly different from that of healthy individuals and that is associated with a higher risk of falling. This phenomenon is highly associated with hemiparesis allied to joint spasticity [69], [70]. Spasticity is a common sequela in stroke survivors, often leading to various deformities and changes in posture and movement patterns due to the abnormal increase in muscle tone and stiffness [69], [70]. Ankle spasticity is among the most prevalent movement disorders following a stroke typically resulting in equinus, varus, or equinovarus deformities [69], [70]. Equinus foot deformity causes the ankle and foot to be in a plantarflexed position, making voluntary dorsiflexion of the ankle joint difficult. This deformity is primarily caused by spasticity in the ankle plantar flexors ( gastrocnemius and soleus muscles). Varus foot deformity is a condition in which the foot is held in an inverted position. This abnormal alignment is primarily caused by spasticity in both the tibialis anterior and tibialis posterior muscles. On the other hand, an equinovarus deformity is characterized by ankle plantar flexion and inversion. It is primarily caused by spasticity in the plantar flexors and tibialis posterior muscle (invertor muscle), with minimal to no contribution from the tibialis anterior (dorsiflexor muscle) [69], [70]. In addition to these different ankle deformities, post-stroke patients also reveal changes in the physiological (EMG), kinematic, and spatiotemporal parameters of the lower limb joints during locomotion [71]–[75]. At the physiological level, there is high heterogeneity in the EMG signals of post-stroke patients. This heterogeneity is found not only between individuals (inter-individual variations) but also in the same individual (intra-individual variations). Inter-individual variation may be explained by the nature of the stroke, which can affect different regions of the brain and result in different motor disabilities [7]. On the other hand, intra-individual variations may occur due to temperature variations [72]. As stated in the study [72], older and female post-stroke patients are more vulnerable to weather conditions. Although there is high variability, there are also common factors at the EMG level, namely a reduction in the magnitude of the EMG signals obtained from the muscles of the paretic limb, early onset of fatigue, and prolonged duration of firing during the gait cycle [71], [76]. Chapter 2 16 Lower limb joint kinematics are affected as a result of irregular EMG patterns. Post-stroke patients are commonly characterized by reduced hip extension during the stance phase [73]. This phenomenon is a consequence of the overactivation of the ankle plantar flexors. The excessive activity of the ankle plantar flexors does not allow the ankle to dorsiflex as much as required. Besides, after a stroke, the length of plantar flexors is commonly decreased, which reduces the ability of these muscles to produce enough force to perform plantar flexion movements at the terminal stance phase [73]. As a result, the forward movement of the upper body is restricted and the hip joint does not extend as expected [71]. Moreover, during the swing phase, hip flexion tends to be reduced as a consequence of the inability to activate hip flexors and/or overactivity of hip extensors [71]. Regarding the knee joint, it typically presents a reduced flexion ability during the early-stance phase. This phenomenon is followed by knee hyperextension used as a compensatory mechanism to achieve stability [74]. At the mid-swing phase, the knee joint typically exhibits a decreased knee flexion due to overactivity of the rectus femoris and/or weakness of the biceps femoris [71]. At the same phase of the gait cycle, the ankle joint presents a reduced dorsiflexion due to overactivity of the plantar flexors. Thus, a neutral position of the ankle joint (around 0º in the sagittal plane) is not commonly verified [75]. In summary, restricted hip and knee flexion, along with decreased ankle dorsiflexion, elongates the leg length during the swing phase. This elongation decreases the foot's clearance from the floor, leading to toe dragging or compensatory leg circumduction [71]. As a consequence of EMG and kinematics being affected, the spatiotemporal performance of poststroke patients is also compromised. At this level, the locomotion of post-stroke patients is typically characterized by a reduced gait speed (below 2.7 km/h [77]) and cadence, increased stride time of the non-paretic limb, and double support time. Moreover, the paretic limb commonly presents a stance phase with a lower duration and a swing phase with a higher duration, compared to the non-paretic limb [71]. To deal with the above-mentioned motor disabilities, physical rehabilitation interventions that promote brain plasticity may include user-oriented, task-oriented, and repetitive gait training that encourages active participation in therapy. Wearable assistive devices, such as exoskeletons and AOs, in conjunction with conventional rehabilitation therapies, may help achieve these goals [9]. These devices have been employed to restore or modify human motor function, enhancing walking ability in individuals with impaired gait due to neurological and/or motor diseases or injuries, such as poststroke patients [78]. Recently, research into lower limb exoskeletons and AOs has surged exponentially, establishing them as prominent physical rehabilitation interventions. These devices reduce the Chapter 2 17 physical burden of therapists and may represent a powerful tool for improving mobility and recovery of post-stroke patients [79]. 2.2 REVIEW OF THE EFFECTS OF LOWER LIMB ASSISTIVE DEVICES ON POST-STROKE PATIENTS As wearable assistive devices have been incorporated into rehabilitation therapies for post-stroke patients, it is notable to consider the benefits of their use. Considering this, this review aims to identify the wearable assistive devices that have been tested in post-stroke patients, as well as the joints typically assisted, the control strategies adopted, and the sensors used. This review also focuses on extracting information regarding the characterization of post-stroke patients, study design, protocols used, and sensor-based and clinical outcomes. The review concludes by identifying the benefits of using wearable assistive devices in rehabilitation sessions for post-stroke patients. 2.2.1 SEARCH METHODOLOGY The literature search was conducted from December 2020 to June 2024 in the Scopus, PubMed, and Cochrane databases using the following keywords: (exoskeleton OR exoskeletons OR orthos?s) AND ("post-stroke") AND ("lower-limb" OR "lower-limbs" OR "lower limb" OR "lower limbs"). This search was limited to titles, keywords, and abstracts. Manuscripts were evaluated based on the following inclusion criteria: (i) original studies; (ii) clinical intervention including post-stroke patients; (iii) lower limb active exoskeletons/orthoses applied for gait rehabilitation. The following exclusion criteria were applied: (i) unused exoskeletons/orthoses (3 papers rejected); (ii) only one training session applied (12 papers rejected); (iii) only protocol presented (6 papers rejected); and (iv) not post-stroke patients (3 papers rejected). Technical and clinical information data were extracted from the selected studies. Technical information includes (i) the exoskeleton used, actuators employed, and joint assisted; (ii) sensor systems used to adapt the assistance; and (iii) assistive control strategies adopted. On the other hand, clinical information includes (i) characteristics of post-stroke patients; (ii) the study design; (iii) the clinical protocol followed; (iv) sensor-based and clinical outcomes; and (v) clinical evidence of wearable assistive devices on post-stroke recovery. Chapter 2 18 2.2.2 RESULTS The literature search identified 123 studies, of which 78, 21, and 19 were found in the Scopus, PubMed, and Cochrane databases and 5 papers were identified manually from the reference sections of other studies. After removing duplicates, 100 studies remained for screening. Based on their titles and abstracts, 58 papers were excluded. Consequently, 42 full-text articles were assessed for eligibility. According to the inclusion and exclusion criteria, 18 studies were finally included. Figure 2.1 depicts the PRISMA flowchart detailing this selection process. Figure 2.1 – PRISMA flowchart. Studies included in review (n = 18) Identification of studies via databases and registers Screening Records screened (n = 100) Records excluded (n = 58) Reports sought for retrieval (n = 42) Reports not retrieved (n = 0) Reports assessed for eligibility (n = 42) Reports excluded: The study does not use exoskeletons/orthoses (n = 3) The study applies a single session (n = 13) The study only presents the protocol (n = 6) The study does not include post-stroke patients (n = 3) Included Identification Records removed before screening : Duplicate records removed (n = 23) Records marked as ineligible by automation tools (n = 0) Records removed for other reasons (n = 0) Records identified from: Scopus (n = 78) PubMed (n = 21) Cochrane (n = 19) Additional Records identified through other sources (n = 5) Chapter 2 19 A. Technical Information Table 2.1 summarizes the technical information extracted from the selected studies, namely, (i) the device, actuator type, and assisted joints; (ii) sensor systems; and (iii) assistive control strategies. Device, Actuator Type, and Assisted Joints According to Table 2.1, the effects of wearable assistive devices in post-stroke patients were studied for different devices. The most explored wearable assistive devices were the EksoNR (six studies (33.3%) [16], [54], [56], [57], [60], [80]) and the HAL (five studies (27.6%) [17], [58], [59], [81], [82]). Two studies explored the ExoAtlet [14] [15]. The effects of FreeWalk [13], H2-Exo [53], Healbot T [83], SMA [18], and a self-made active ankle-foot orthosis [55] were explored in only one study each. Of the eighteen selected studies, seventeen used an electric motor-based actuator with gearbased transmission [13]–[18], [53], [54], [56]–[60], [80]–[83] (94.4%). Information about the actuator of the self-made active ankle-foot orthosis of the study [55] was not detailed. Additionally, the lower limb joint typically addressed by wearable assistive devices was the hip joint (seventeen studies (94.4%) [13]–[18], [53], [54], [56]–[60], [80]–[83]), followed by the knee joint (sixteen studies (88.9%) [13]–[17], [53], [54], [56]–[60], [80]–[83]). These two joints were assisted together in sixteen studies (88.9%) [13]–[17], [53], [54], [56]–[60], [80]–[83]. The ankle joint was (i) actively assisted in two studies (11.1%) [53], [55]; (ii) passively assisted in fourteen studies (77.8%) [13]–[17], [54], [56]–[60], [80]–[82]; and (iii) not assisted in two studies (11.1%) [18], [83]. Sensor Systems Different sensor systems were employed in the control strategy architecture of each study. Seven studies (41.2%) used Force Sensing Resistor (FSR) sensors to detect lateral and forward weight shifts [16], [54], [56]–[58], [60], [80] and to segment gait cycles [81]. Five studies (29.4%) employed EMG systems, acquiring EMG signals from illiopsoas , gluteus maximus , semitendinosus , and vastus lateralis muscles [17], [58], [59], [81], [82]. Three studies (17.6%) used embedded encoders to measure the lower limb joint angles [18], [53], [83]. Single studies have used an Inertial Measurement Unit (IMU) sensor at the trunk to segment gait cycles [57] and an interaction torque sensor [53] to determine the human-robot interaction in each lower limb joint. Chapter 2 20 Assistive Control Strategies Several assistive control strategies were adopted by different wearable assistive devices. Of the eighteen selected studies, eight studies (44.4%) used a therapist-led step-by-step control [13]– [15], [54], [56], [57], [60], [80]. This control strategy is driven by a physical therapist who commands the steps that the user should perform by using a push button in the wearable assistive device. Six studies (33.3%) explored the potentialities of a trajectory-tracking control based on lateral and forward weight shift [16], [54], [56], [57], [60], [80], all applied with the EksoNR. This control strategy acts as a hybrid approach in which each step is triggered by the lateral and forward weight shift of the user. Once triggered, the user follows the trajectory imposed by the wearable assistive device. In five studies (27.8%), this control strategy was adopted after the therapist-led step-by-step control, when the user learns to weight-shift to a stance position [54], [56], [57], [60], [80]. Five studies (27.8%) employed an EMG-based control [17], [58], [59], [81], [82], all applied with the HAL. This control strategy tries to find a correlation between EMG and torque signals. Once found, the controller sets assistive torque commands proportional to the EMG signals of the patient. Two studies (11.1%) explored a trajectory-tracking control [17], [83]. With this control strategy, the wearable assistive device provides fixed assistance levels that are required to complete the desired position trajectories. The user follows the trajectory imposed by the wearable assistive device. One study (5.6%) explored the impedance control for H2 exoskeleton [53]. When using this control type, the aim is to adjust the degree of freedom in the patient's movements by varying the interaction stiffness. Low levels of interaction stiffness allow greater freedom of movement, while high levels result in a more rigid behavior by the wearable assistive device to strictly enforce the reference trajectory. Adjusting the interaction stiffness controls the torque exerted by the wearable assistive device, thereby influencing the patient's interaction and the effort required to maintain the desired gait pattern. One study (5.6%) used a Control Pattern Generator (CPG)-based control. This control system uses neural oscillators in addition to the user's CPG to synchronize with the user's movements. The joint angles are measured in real-time, acting as input in the controller, which assesses the joint angles' symmetry. Based on this analysis, the controller generates assistive torques at specific points of the gait cycle to improve its symmetry. Chapter 2 21 Table 2.1 – Technical details extracted from the selected studies (A, P, and NI means active, passive, and not indicated, respectively) Study Device Actuator Actuated joint Sensor systems used in control strategies Control strategy Hip Knee Ankle Sensor Measurement Lee et al. [13] Free Walk Electric with gear-based transmission A A P Embedded encoder Therapist-led step-by-step exercise Kotov et al. [14] ExoAtlet Electric with gear-based transmission A A P Therapist-led step-by-step exercise Kovalenko et al. [15] ExoAtlet A A P Therapist-led step-by-step exercise Louie et al. [80] EksoNR Electric with gear-based transmission A A P Embedded Foot Pressure Sensor (FSR) Lateral and forward weight shift Therapist-led step-by-step exercise and Trajectorytracking control based on lateral weight shift Calabrò et al. [16] EksoNR A A P Embedded FSR Lateral and forward weight shift Trajectory-tracking control based on lateral weight shift Zhu et al. [60] EksoNR A A P Embedded FSR Lateral and forward weight shift Therapist-led step-by-step exercise and Trajectorytracking control based on lateral weight shift Infarinato et al. [57] EksoNR A A P Embedded FSR Lateral and forward weight shift Therapist-led step-by-step exercise and Trajectory- Chapter 2 28 The post-stroke stage effects were verified by the Fugl-Meyer Assessment – Lower Extremity (FMA-LE) (87.5%) [17], [53], [55], [58], [80]–[82], and the Stroke Impairment Assessment Set (SIAS) (12.5%) [55]. The mental effects were analyzed in single studies by the 36-item [80] and 12-item Short Survey [13], Montreal Cognitive Assessment [80], and Patient Health Questionnaire [80]. To evaluate the mentioned effects, all selected studies carried out a preand post-training evaluation. The studies [18], [54], [59], [81], [82] performed additional evaluations in intermediate time points between the beginning and end of the training. Moreover, the study [18] and the studies [80], [82] performed a 3-month and 6-month follow-up, respectively, to evaluate the retention of the effects. Clinical Evidence of Wearable Assistive Devices in Post-Stroke Patients For randomized controlled and non-randomized controlled studies, the clinical evidence was presented as improvements achieved by the experimental group (the group that performed conventional and robotic-based training sessions) over the control group (the group that carried out conventional therapy), considering the preand post-training time points. In the case of uncontrolled studies, the clinical evidence was evaluated by comparing the outcomes of the experimental group at the preand post-training time points. Gait speed improvements were the outcome mostly referenced (27.5%) [13]–[18], [53]–[56], [58]–[60], [83], followed by improvements in the balance (11.8%) [14]–[17], [54], [59], distance walked (9.8%) [13], [17], [53], [54], [60], gait symmetry (9.8%) [16], [18], [58], [60], [81], muscle strength of the quadriceps (7.8%) [13], [14], [16], [54], and postural stability (5.9%) [14], [15], [54]. Other improvements were reported once (2.0%), such as stride length [18], step length [18], walking time [56], time of verticalization [56], number of steps [56], step cadence [58], foot clearance [60], knee flexion [16], range of motion of the hip, knee, and ankle of the non-paretic limb [58], range of motion of the hip and knee of the paretic limb [58], muscle strength of tibialis anterior, gastrocnemius, and biceps femoris [57]. Further, the study [14] also reported a decrease in the hemiparesis degree and energy expenditure. Additionally, the results of the post-training for two studies [80], [82] did not report any improvements in the experimental group over the control group. Unlike the study [14], the study [60] did not show improvements in the energy expenditure. The post-training results of the study [57] did not reveal improvements in gait speed, spasticity level, and postural stability. Despite the post-training Chapter 2 29 results revealing improvements in the gait speed, balance, gait symmetry, knee flexion, and knee muscle strength, the study [16] did not find significant improvements in the muscle activity of the tibialis anterior and gastrocnemius muscles of the experimental group over the control group. Moreover, the study [58] also reported a decrease in the range of motion of the ankle joint on the paretic side. Chapter 2 30 Table 2.2 – Clinical details extracted from the selected studies Study Sample Size (EG/CG*) Study design Clinical protocol Outcomes Evaluation time points Clinical evidence Sensor-based Clinical Lee et al. [13] 38 (17/21) RCT 3 times/week until 12 sessions Quadriceps isokinetic muscle strength Timed Up and Go Test (TUG); 6-min Walk Test (6MWT); 12-item short form survey Preand post-training • Improvement of knee muscle strength, distance walked, gait speed, and quality of life Kotov et al. [14] 42 (21/21) RCT 30-min 5 times/week until 10 sessions Vertical stability; Energy consumption Medical Research Council (MRC); Modified Rankin Scale; Bartel Index (BI); Hauser Ambulation Index (HAI); Berg Balance Scale (BBS); 10-Meter Walk Test (10MWT) Preand post-training • Improvement of paretic limb muscle strength, balance, postural stability, and gait speed. • Decrease in hemiparesis degree and energy consumption Louie et al. [80] 36 (17/19) RCT 25-min twice/week until 24 sessions Functional Ambulation Categories (FAC); FuglMeyer Assessment – Lower Extremity (FGM-LE); 5Meter Walk Test (5MWT); 6MWT; BBS; Montreal Cognitive Assessment Pre-, posttraining, and 6-month follow-up No improvements were found when comparing the exoskeleton-based therapy with conventional therapy Chapter 2 31 Study Sample Size (EG/CG*) Study design Clinical protocol Outcomes Evaluation time points Clinical evidence Sensor-based Clinical (MCA); 36-item short form survey Kovalenko et al. [15] 62 (31/31) RCT 1-h/day until 10 sessions Tardieu Scale (TS); Modified Ashworth Scale (MAS); MRCS; 10MWT; Rivermead Mobility Index (RMI); BBS; Modified Rankin Scale Preand post-training Improvement of gait speed, balance, and postural stability Calabrò et al. [16] 40 (20/20) RCT 45-min 5 times/week until 40 sessions RMS of tibialis anterior, soleus, rectus femoris, and biceps femoris ; step cadence; stance/swing ratio; gait quality index; gait cycle duration 10MWT; RMI; TUG; Preand post-training • Improvement of gait speed, balance, gait symmetry, knee flexion, and knee muscle strength. • The tibialis anterior and soleus did not show improvements Zhu et al. [60] 21 (10/11) Non-RCT 50-min 3 times/week until 15 sessions Energy expenditure; lower limb joint angles (hips, knees, and ankles); RMS of tibialis anterior; soleus, gastrocnemius medialis, 10MWT; 6MWT, TUG; Preand post-training • Improvement of gait speed, distance walked, foot clearance, and gait symmetry. Chapter 2 32 Study Sample Size (EG/CG*) Study design Clinical protocol Outcomes Evaluation time points Clinical evidence Sensor-based Clinical vastus medialis, rectus femoris, biceps femoris, semitendinosus, and gluteus medius ; walking speed; foot clearance; stance/swing ratio • The energy expenditure did not show improvements Tan et al. [58] 8 (8/0) Uncontrolled 20-min 3 times/week until 9 sessions Walking speed, cadence, step length, stance/swing ratio, range of motion of the lower limb joints (hips, knees, ankles); EMG from vastus medialis, semitendinosus, tibialis anterior, gastrocnemius, adductor longus, and gluteus maximus FAC; Functional Independence Measure (FIM); FMA-LE; Preand post-training Improvement of gait speed, step cadence, gait symmetry, hip and knee range of motion of the paretic side, and hip, knee, and ankle range of motion of the unaffected side Tan et al. [81] 20 (9/11) RCT Stance ratio between paretic and non-paretic limb; EMG from vastus medialis, semitendinosus, tibialis anterior, gastrocnemius, adductor longus, and gluteus maximus FAC; FIM; FMALocomotion; FMA-Motor; FMA-LE Pre-, after session 3, after session 6, and posttraining Improvement of gait symmetry Chapter 2 33 Study Sample Size (EG/CG*) Study design Clinical protocol Outcomes Evaluation time points Clinical evidence Sensor-based Clinical Infarinato et al. [57] 8 (8/0) Uncontrolled 5 times/week until 15 sessions EMG from tibialis anterior, gastrocnemius medialis, rectus femoris, and biceps femoris; knee joint angles MAS; Motricity Index (MI); FAC; Trunk Control Test (TCT); 10MWT Preand post-training • Improvement of muscle strength at tibialis anterior, gastrocnemius medialis, and biceps femoris • No improvements were found in gait speed, spasticity level, and postural stability Szurle et al. [56] 19 (19/0) Uncontrolled 50-min 3 times/week until 12 sessions Number of steps; walking time; time of verticalization TUG Preand post-training Improvement of gait speed, number of steps, walking time, and time of verticalization Lee et al. [83] 43 (33/10) RCT 30min/session until 10 sessions 10MWT; BBS; FAC; TUG; MI Preand post-training Improvement of postural stability, balance, and gait speed Tomioka et al. [55] 27 (27/0) Uncontrolled 40-min 6 times/week FMA-LE; TUG; 10MWT; Stroke Impairment Assessment Set (SIAS) Preand post-training Improvement of gait speed Chapter 2 34 Study Sample Size (EG/CG*) Study design Clinical protocol Outcomes Evaluation time points Clinical evidence Sensor-based Clinical until 24 sessions Molteni et al. [54] 23 (23/0) Uncontrolled 60-min 3 times/week until 12 sessions MAS; MI; TCT; FAC; 10MWT; 6MWT Pre-, after session 6, and posttraining Improvement of muscle strength, postural stability, balance, gait speed, and distance walked Watanabe et al. [17] 22 (11/11) RCT 20-min 3 times/week until 12 sessions Isometric muscle strength at hip and knee flexion and extension movements FAC; 10MWT; 6MWT; FMALE Preand post-training Improvement of gait speed, balance, and distance walked Yoshimoto et al. [59] 18 (9/9) Non-RCT 20-min once/week until 8 sessions FAC; 10MWT; TUG; BBS Pre-, after session 4, and posttraining Improvement of gait speed and balance Bortole et al. [53] 3 (3/0) Uncontrolled 3 times/week until 12 sessions BBS; Barthel Index (BI); FAC; FMA-LE; TUG; 6MWT Preand post-training Improvement of distance walked and gait speed Chapter 2 35 Study Sample Size (EG/CG*) Study design Clinical protocol Outcomes Evaluation time points Clinical evidence Sensor-based Clinical Buesing et al. [18] 50 (25/25) RCT 45-min 3 times/week until 18 sessions Walking speed; cadence; step time; step length; stride length; stance time; double support time Pre-, after session 9, post-training, and 3-month follow-up Improvement of gait speed, step time, step length, stride length, and gait symmetry Wall et al. [82] 32 (16/16) RCT 60-min 4 times/week until 16 sessions FAC; FMA-LE; 2-minute Walk Test; BI Pre-, posttraining, and 6-month follow-up No improvements were found when comparing the exoskeleton-based therapy with conventional therapy *EG and CG mean experimental and control groups, respectively Chapter 2 36 2.2.3 DISCUSSION A. Technical Information Among various clinical trials, EksoNR and HAL were the most commonly used devices to evaluate the effects of wearable assistive device therapy in post-stroke patients. Although the use of exoskeletons for the rehabilitation of post-stroke patients has been a topic of discussion for some time, these results show that only two wearable assistive devices have featured prominently in studies designed to prove the effectiveness of these devices in post-stroke patients. Future research should address this gap by conducting clinical trials with other devices to investigate the benefits of each device or investigate if there are differences among them. According to the study [53], (i) hydraulic and pneumatic actuators offer high power density but are typically bulky and prone to internal leakage and friction; (ii) series elastic actuators are limited by the fixed spring constant of their elastic elements; and (iii) electric motor-based actuators have been indicated due to its reduced power consumption during gait. These results support the findings of the studies reviewed since electric motor-based actuators were typically adopted for the majority of the studies. In most of the studies, the hip and knee joints were actively assisted by the wearable assistive device, while the ankle joint was passively assisted. According to the study [71], the dorsiflexion of the ankle joint during the stance phase is restricted due to the overactivity of the plantar flexor muscles, specifically the gastrocnemius muscle. This limitation affects the hip joint by preventing it from extending as far as necessary. Another common kinematic issue is reduced ankle dorsiflexion during the swing phase. Normally, the ankle reaches a neutral position at the mid-swing phase and maintains this position until initial contact. Achieving a neutral ankle position at this point is crucial because it brings the foot closer to the ground, facilitating limb movement to prevent toe drag. The lack of dorsiflexion during the swing phase and at the heel-strike phase is likely caused by overactive plantar flexor muscles and insufficient dorsiflexion. As a result, post-stroke patients tend to perform leg circumduction [71]. These phenomena suggest that the ankle joint is often affected in post-stroke patients and that sometimes a kinematic change in another lower limb joint (such as the hip) is a consequence of the ankle joint being affected. Therefore, future research should address these issues by developing wearable assistive devices that invoke the ankle joint to play an active role in rehabilitation therapies. Chapter 2 37 Additionally, weight shift detections measured by FSRs, EMG signals from, gluteus maximus , semitendinosus , and vastus lateralis muscles, and lower limb joint angles measured by embedded encoders were the most used type of data in the control strategies of the reviewed studies. Although EMG sensors were widely used, their application was limited to the HAL device. Given the importance of these sensors in discriminating the locomotion intentions and needs of post-stroke patients, future research should address the inclusion of these sensors in the control architecture. The therapist-led step-by-step control was the most used control when using wearable assistive devices, followed by the trajectory-tracking control based on lateral and forward weight shift. This last control strategy was typically adopted after the therapist-led step-by-step control when the post-stroke patient learned to weight-shift to a stance position. While in the therapistled step-by-step control, post-stroke patients are commanded by the physical therapist, the trajectorytracking control based on lateral weight shift invokes the patients' participation to continuously perform right and left steps. Nonetheless, the user’s participation is only required at the beginning of each step and not for the complete gait cycle. In addition, the effect of EMG-based control strategies was also explored. For the wearable assistive device to move, the post-stroke patient must actively participate throughout the gait cycle, as the device provides assistance that is proportional to the level of muscle activation of the monitored muscles. Therefore, non-participation by post-stroke patients means that the wearable assistive device does not assist [37]. Despite being a valuable contribution to muscle strengthening, these EMG-based torque controls do not consider different levels of motor disabilities, nor assist the user when and as much as needed. Future research should address this limitation by developing AAN control strategies to provide the robotic assistance needed for post-stroke patients to complete a movement, taking into account their motor impairments and invoking their participation. B. Clinical Information Clinical trials investigating the effects of wearable assistive devices on post-stroke recovery had, on average, a diverse sample of participants, ranging from 3 to 62, which provides a benchmark for future research. Participants in these trials were primarily characterized by age (from 18 to 80 years old), gender (balanced gender distribution), time post-stroke (varying between sub-acute (from 25 days to 6 months) and chronic (from 6 months to 9.5 years) stages), hemiplegic side, absence of cardiopulmonary disease, and stroke etiology. To increase the reliability of future Chapter 3 44 sensor systems were developed in the study [2]. However, in this Ph.D. thesis, a new version of the MuscLab system was projected and concretized (detailed in Chapter 4). The FootLab consists of instrumented shoes with two IMUs (one in the upper of the shoe (LSM6DS3) and the other in the insole (LSMDSOX)) and an insole instrumented with 8 FSRs. Both sensors enable the acquisition of the foot kinematics and the creation of a pressure map on the surface of the foot, respectively. The system presents Bluetooth wireless technology and a rechargeable battery. The InertialLab is composed of seven IMUs (MPU6050) positioned at the pelvis, right/left thighs, shanks, and feet, measuring the kinematics of lower limb segments. The MuscLab (detailed in Chapter 4) is an e-textile designed to monitor the kinematics and the muscle contraction of muscles in a human segment. The system is made of an IMU (LSM6DS3) and piezoresistive textile strips (Shieldex® Technik-tex P130+B), integrating Bluetooth wireless technology and a rechargeable battery. The EMG Trigno Avanti system integrates eight sensors that enable the acquisition of EMG muscle activity during different tasks. Each EMG sensor has a built-in triaxial IMU, composed of an accelerometer and gyroscope with a transmission range of 40 meters and a rechargeable battery. Table 3.1 summarizes the data collected by each sensor system. Table 3.1 – Different types of data collected by each sensor system Sensor system Sensor data FootLab ➢ Feet kinematics (angular speed and acceleration) ➢ Plantar pressure (pressure map) InertialLab ➢ Kinematics (angular speed, acceleration) of lower limb segments (right and left foot, shank, thigh, trunk) MuscLab ➢ Kinematics (angular speed, acceleration) of the shank segment ➢ Muscle contraction of the shank muscles EMG Trigno Avanti ➢ Root Mean Square up to 8 EMG signals In this Ph.D. thesis, the EMG Trigno Avanti system was integrated into SmartOs system by using the Trigno SDK, to enable the real-time acquisition of EMG signals during the orthosis’ use [86]. For that, a Transmission Control Protocol/Internet Protocol (TCP/IP) was implemented to communicate between both systems (CCU of SmartOs system and the Base Station of Trigno Avanti system). To decrease the amount of data to be sent via TCP/IP and to work with cleaner EMG signals, the Root Mean Square (RMS) mode of the Trigno Avanti system was configured (Avanti-Only Modes: 83). This mode enables the Chapter 3 45 acquisition of rectified EMG signals with a frequency of 148 Hz, by applying the RMS method. By default, the RMS mode sends 27 EMG samples at every 0.0135 s (74 Hz). These samples are reprocessed in SmartOs system by using the RMS at 74 Hz [88]. Figure 3.2 depicts the EMG sample transmission between 1 EMG sensor and the CCU of SmartOs. As mentioned, for each EMG sensor, it is applied the RMS method to the EMG raw data collected from a specific muscle, returning a single EMG sample with a frequency of 148 Hz. These rectified EMG samples are sent in packets of 2 samples to the Trigno Avanti Base Station via a Bluetooth protocol integrated into the Trigno Avanti system, being then sent via UART to an external computer running the Trigno SDK. It is noteworthy that this external computer is only required because the Trigno SDK runs in the Windows operating system, and not in the Ubuntu operating system, which is the operating system running in the SmartOs’ CCU. Subsequently, the two received EMG samples are resampled to 27 EMG samples inside the Trigno SDK, being then sent by Wi-Fi to the CCU that applies again the RMS method to retain one EMG sample with a frequency of 74 Hz. Figure 3.2 – Flowchart for acquiring RMS EMG samples with the EMG Trigno Avanti system. 3.2.3 GAIT ANALYSIS TOOLS In the SmartOs framework, all signals measured with the wearable motion lab can be subsequently used in gait analysis tools. In the study [2], several gait analysis tools were developed, namely, the estimation of (i) spatiotemporal parameters (walking speed, step length, step width, stride length, distance); (ii) force parameters (center of pressure); (iii) lower limb joint angles (ankle, knee, hip) and Chapter 3 46 segment angles (foot, shank, thigh, trunk). This Ph.D. thesis advances by presenting more three gait analysis tools, i.e., the ability to decode LMs (Chapter 5), and to estimate ankle joint torques (Chapter 6.3), and energy expenditure (Chapter 6.4), in real-time. 3.2.4 HIERARCHICAL CONTROL ARCHITECTURE The SmartOs system endows a hierarchical control architecture organized into three levels, as suggested in the study [9]. Inspired by the human motion control system, this architecture integrates both structural and functional aspects of control and sensor feedback systems. At the highlevel, known as the perception layer, the system generates user-oriented trajectories. The mid-level acts as the translation layer, being responsible for transforming the user-oriented trajectories into reference trajectories for the AO in accordance with walking speed. At last, the low-level generates assistive commands, ensuring that the state of the AO effectively tracks the desired assistance trajectory in a timely manner. Currently, the control architecture includes low-level position-based and torque-based tracking controllers using Proportional Integral Derivative (PID) and Feedback-Error Learning (FEL) control, respectively [89]. In terms of control frequency, the highand mid-level frequencies were set at 100 Hz, a sufficient rate for human-machine gait analysis. Conversely, the low-level operates at 1 kHz to allow high-frequency operation, which is conducive to effective human-machine tracking control loop dynamics. The software routines controlling the lowand mid-level controllers were coded in C and implemented on the STM32F407VGT microcontroller. In contrast, the high-level controllers, coded in C++, are executed within the CCU, which is housed within a UDOO X86. This architectural framework embodies a modular design, facilitating scalability to incorporate additional assistive control strategies as required to expand SmartOs into a versatile robot-based gait training solution. Currently, the SmartOs system includes seven user-centered, closed-loop assistive control strategies, as presented in Table 3.2. These strategies have been carefully designed to address the therapeutic purposes identified for the SmartOs system in post-stroke gait training. Taken together, these strategies make SmartOs adaptable to different therapeutic approaches and address both immediate and permanent changes in motor ability. Each strategy allows for gait speed adjustments in the range of 0.5 to 1.6 km/h, taking into account the mechanical limitations of the AOs, thereby facilitating gait training across the different challenges inherent in each strategy. This Ph.D. thesis advances by proposing the AAN LM-driven trajectory (Chapter 6.2), the AAN EMG-based (Chapter 6.3), and the AAN HITL (Chapter 6.4) controls. Chapter 3 47 Table 3.2 – Different control strategies available in the SmartOs system Control Strategy Therapeutic Purpose Benefits Zero-torque Control Rehabilitation ➢ Familiarization period ➢ Muscle strengthening Trajectorytracking Position Control Assistance and/or Rehabilitation ➢ User-oriented repetitive gait training ➢ Recovery of the user’s gait pattern (range of motion (ROM) and symmetry) Adaptive Impedance Control Rehabilitation ➢ Recovery of the user’s gait pattern (ROM and symmetry) ➢ Invoke the user’s active participation ➢ Muscle strengthening ➢ Manual assistance level adjustment ➢ Long-term recovery of functional motor abilities EMG-based Control Rehabilitation ➢ Invoke the user’s active participation ➢ Muscle strengthening ➢ Long-term recovery of functional motor abilities AAN LM-driven Trajectory Control Assistance and/or Rehabilitation ➢ User-oriented repetitive gait training in different LMs ➢ Invoke the user’s active participation AAN EMG-based Control Rehabilitation ➢ Recovery of the user’s gait pattern (ROM and symmetry) ➢ Invoke the user’s active participation ➢ Muscle strengthening ➢ Automatic assistance level adjustment ➢ Long-term recovery of functional motor abilities AAN HITL Control Assistance ➢ Energetic-efficient motor recovery 3.2.5 GRAPHICAL APPLICATION The mobile graphical application facilitates the intuitive configuration of all SmartOs modules, allowing the system to be set up for monitoring and/or assistance in training sessions. This application meets requirements such as (i) simplified and guided interaction for fast, natural, and easy navigation, and (ii) the use of explicit graphical components for better user understanding, as initially defined in the study [2]. Developed for the Android operating system, all messages are transmitted via the Bluetooth Chapter 3 48 protocol to the SmartOs CCU. In this Ph.D. thesis, the mobile graphical application has been extended to allow the selection of new sensors (MuscLab, and EMG Trigno Avanti system), algorithms (ankle joint torque estimation, LM decoding, and energy expenditure estimation), and control strategies (AAN LMdriven trajectory, AAN EMG-based, and AAN HITL controls), as shown in Figure 3.3. Figure 3.3 – Mobile graphical application. Chapter 3 49 3.3 CONCLUSIONS For delivering personalized and user-oriented assistance for post-stroke patients, the modular and hierarchical architecture of the SmartOs system was extended. The MuscLab system (Chapter 4) was developed and integrated into the SmartOs system to enable the monitorization of multiple lower limb muscles simultaneously, in real-time. Future research on this sensor may allow (i) muscle fatigue estimation; (ii) LMs decoding; and (iii) lower limb joint and segment angle estimation. Additionally, the integration of an LM decoding tool (Chapter 5) and the development of an AAN LM-driven trajectory control (Chapter 6.2) enables the adaptation of the system dynamics according to the user’s locomotion intentions. The integration of the EMG Trigno Avanti system and the development of a joint torque estimation algorithm facilitated the design of an AAN EMG-based control strategy (Chapter 6.3) to assist the user as much and when needed. Finally, the development of AAN HITL control (Chapter 6.4), based on real-time estimation of energy expenditure, makes it possible to optimize the control parameters of the AOs to help users reduce their metabolic costs. The mobile graphical application was updated with these new functionalities to enable the application of these strategies during rehabilitation therapies. 50 4. MUSCLAB SYSTEM Chapter 4 51 This chapter describes the MuscLab system, an elastic and flexible textile band that simultaneously monitors muscle contraction in extensor and flexor muscles and the kinematics of the segment in which it is located. The chapter begins with an introductory overview of the MuscLab system, followed by the conceptual design and functionalities of the system. In addition, this chapter presents the technological solutions implemented, taking into account the hardware and software interfaces. The chapter ends with the system validation against gold-standard tracking technologies. 4.1 INTRODUCTORY INSIGHT According to the World Health Organization, musculoskeletal disorders (MSDs) are the world's leading cause of human motor disability. An estimated 1.71 billion people worldwide live with MSDs. Most MSDs are associated with the work context, sedentary lifestyle, and practice of sports, at a professional or amateur level [90]. As in the case of stroke, MSDs make it difficult to carry out daily tasks at home and work and are the main reason for absenteeism and early retirement [91]. This raises the need for developing solutions to promote objective, non-invasive, and real-time monitoring of muscle contraction and relaxation. Ideally, this solution should be versatile to use in different contexts, such as (i) health, to support the clinical diagnosis of the evolution of muscle contraction or to serve as a tool for gait disability level assessment in post-stroke patients; (ii) sports, for analysis of muscle performance and prediction of the risk of MSDs; (iii) work, to support the ergonomics assessment of muscle condition and prediction of the risk of MSDs. Typically, surface muscle contraction monitoring is conducted by expensive but high-precision sensing EMG equipment. The gold-standard surface EMG solutions include the non-invasive EMG electrodes from Ultium EMG (Noraxon, Toronto, Canada) [92] and Pico EMG (Cometa Systems, Newburg, USA) [93], or the dry EMG sensors such as Trigno Avanti (Delsys Incorporated, Natick, USA) [87]. However, the long-term use of these sensors can be affected by (i) sweating; (ii) temperature variations; and, (iii) movements between the skin and the electrodes [42]. Furthermore, if several agonist and antagonist muscles need to be monitored simultaneously, it will need as many sensors as the number of muscles to monitor. This may result in a more intrusive and time-consuming solution for donning and doffing, narrowing its practical use. Current challenges center on the development of easily wearable and cost-effective sensors that enable the real-time monitoring of several muscles simultaneously in a practical way. In this connection, this chapter presents a second version of the MuscLab system (advancing [94]), which corresponds to a Chapter 4 52 low-cost, non-intrusive, self-calibrated, and stand-alone prototype to monitor the muscle contraction in several muscles of a human body segment, simultaneously. 4.2 CRITICAL ANALYSIS OF RELATED WORK Different devices have been proposed for the non-invasive evaluation of muscle contraction. The studies [95], [96] proposed the development of garments covered with dry EMG electrodes. In the study [95], two garments (trousers and a shirt) monitored the electrical activity of the muscles present in the lower limbs, upper limbs, and torso, making a maximum of 19 muscles. In the study [96], a flexible and elastic textile band with three electrodes sewn parallel to each other was responsible for measuring EMG signals in fingers, wrists, elbows, shoulders, back, hips, knees, ankles, and neck. Considering the two patent documents presented above [95], [96], it was verified that the inclusion of EMG electrodes in textiles with flexible character facilitates the monitoring of several muscle groups simultaneously. Nonetheless, dry EMG electrodes need to be always in contact with the user’s skin. As an alternative to overcome the limitations of using EMG, MMG has emerged for muscular analysis [97]–[103]. MMG sensors, although they do not measure electrical signals from muscles, can detect whether muscles are contracting or relaxing, without the need to be in direct contact with the user's skin. The principle underlying the MMG sensors is that muscle contraction is typically associated with muscle shortening, causing an increase in muscle cross-sectional area, stiffness, tension, and mechanical vibration. In this connection, an MMG sensor (e.g., force, capacitive, and inertial sensor) positioned above the muscle may detect its contraction [97]. The studies [97], [98] used a piezoresistive force sensor placed on arm muscles to monitor their contraction. Due to its piezoresistive property, there is a shape variation of the force sensor as the muscle under the sensor is contracted; thus, varying the sensor resistance. Both studies verified that the MMG signals have a similar pattern to the envelopes of the EMG signals. Moreover, a patent document [99] described another MMG-based device for monitoring muscle activity, capturing muscle mechanical vibrations through an inertial sensor. The patent document [100] used a capacitive sensor over the muscle to measure capacitance variations according to muscle contraction and relaxation. From these studies [97], [98] and patent documents [99], [100], it was found that the use of force, inertial, and capacitive sensors allows the monitoring of muscle activity of the muscles where the sensors are located. However, these solutions may become non-practical for daily use when it is necessary to monitor multi- Chapter 4 53 muscle groups of a body segment simultaneously since the number of MMG sensors increases according to the number of muscles to monitor. Studies [101], [102], [103] proposed a device that can monitor the muscle contraction of several muscles or muscle groups simultaneously. The study [101] presented a cord-shaped sensor with piezoresistive properties to measure muscle contraction in the forearm muscles. To do this, the cord was placed around the forearm segment. The contraction of agonist and antagonist muscles changes the length of the cord; thus, varying its resistance. Despite simultaneously monitoring the contraction of agonist and antagonist muscles, this solution cannot discriminate contractions of both muscles, since both contractions cause an increase in the length of the cord. On the other hand, the study [102] disclosed a flexible textile band (72% nylon and 28% spandex) for monitoring the muscle contraction of the muscles in the shank segment, while discriminating which muscle is contracted. A flexible, piezoresistive, and e-textile (conductive textile) has been integrated into an external textile band and sewn into a matrix shape. The external textile band was designed for the shank segment and, due to its matrix configuration, it is possible to discriminate which muscles of the shank region are contracted/relaxed through the pressure they make on the different zones of the matrix. Similarly, in the patent document [103], a circular band with pressure sensors is described for the simultaneous monitoring of muscle contraction/relaxation of muscles responsible for hand, wrist, or arm movements. Although these two approaches [102], [103] can monitor the muscle activity of several muscles simultaneously, the textile bands do not have elastic characteristics. As such, these solutions require the development of a user-specific monitoring device to ensure the proper sensor placement according to the user’s anthropometry. Considering all the studies and patent documents presented above, it becomes imperative to develop a solution based on MMG sensors sewn onto a flexible and elastic textile band, so that the same device can be used for different anthropometries and on different parts of the human body. For this purpose, this Ph.D. thesis advances with the MuscLab system – a stand-alone MMG solution based on e-textile (piezoresistive textile) sensors sewn onto a flexible and elastic textile band. The MuscLab system was designed as a proof of concept to simultaneously monitor and discriminate the muscle contraction of the shank segment muscles (mainly, the tibialis anterior and gastrocnemius lateralis ) in individuals with different anthropometries. Chapter 4 60 resistance of each piezoresistive textile strip. During this calibration step, the user must remain in a sitting position (with the knee joints performing 90º) and static for 10 seconds, while the software routine finds the optimal position among 256 possible programmable positions of the digital potentiometer until the output of the ADC reaches 0.75 V for each piezoresistive textile strip. Once calibrated, the algorithm moves on to a calibration routine for the second digital potentiometer (AD8400 50 kΩ). This potentiometer is responsible for amplifying the signal read for each strip, allowing gains between 1 and 250. To calibrate it, the user was asked to perform dynamic contractions (maximum voluntary contractions are also valid) for 10 seconds. This calibration aimed to determine the gain necessary to ensure that the maximum voltage is 3.3 V, the maximum value allowed by the Arduino. After calibrating the digital potentiometers, the Arduino’s program waits to receive a command message from the MuscLab API. These commands are received by the Nina processor and sent via UART serial communication to the Cortex-M0 32-bit SAMD21 of the Arduino Nano 33 IoT. If the command sent by the API is a start command, it will initiate the collection of inertial data, piezoresistive textile data, and battery voltage level. This is done by reading the battery data every 5 minutes and the inertial and piezoresistive data every 10 ms. The data are sent to the MuscLab API at 100 Hz through a data message, organized as follows: • byte to indicate the start or header of the packet (indicated with the value of the ‘h’ character); • 1 byte comprising the message number and the battery value (first 4 bits for the message number, 0 to 15; and the last 4 bits for the battery value, (0, 10, 20, ..., 100 %) divided by 10; • The sensor data bytes (variable size according to the number of sensors requested by the API, these can be inertial data and/or data from the piezoresistive textile); o Inertial sensor: 2 bytes per signal (accelerometer and gyroscope in X, Y, and Z components) - a total of 12 bytes; o Piezoresistive textile sensor: 2 bytes per strip - a total of 10 bytes. • 2 bytes indicating the end of the packet or the tail of the packet (indicated with the value of the ‘t’ character). Once the data has been received by the API, the data are parsed into package ID, battery level, time stamp, 3-axis accelerometer and gyroscope, and voltage read of each piezoresistive textile strip. The piezoresistive data is filtered in the API by using a low-pass exponential filter with a cutoff frequency of 5 Chapter 4 61 Hz. The API then saves all data into a text file. The code is continuously running until the user stops it through a stop command sent by the API. Figure 4.4 – Flowchart of the MuscLab acquisition system. DP1 and DP2 represent the Digital Potentiometer of the Wheatstone bridge and summing inverting amplifier, respectively. The blue and green blocks represent the flow diagram of the Arduino and API codes, respectively. Start Switch on the system Calibrate DP Calibrate DP DP calibrated DP calibrated Wait for a Bluetooth connection Bluetooth connected Wait for Start Command Start Command received Collect inertial, pie oresistive, and battery data Send collected data Stop Command received Stop data collection Start Run API Input directory Devices found Search devices to connect Wait for a Start Command by the user Start Command introduced Send Start Command to client Wait for new data New data received Save collected data Wait for a Stop Command by the user Stop Command introduced Stop data collection Connect to client yes yes yes yes yes yes yes yes yes no no no no no no no no no Chapter 4 62 4.3.5 EXPERIMENTAL VALIDATION Two experimental tests were carried out to infer the operability of the MuscLab system. First, bench tests were conducted to check if the requirements defined for MuscLab were met. Moreover, human tests were performed to evaluate the reliability of the MuscLab signals against EMG and kinematic signals during dorsiflexion and plantar flexion movements of the ankle joint. A. Bench Tests A benching protocol was carried out to evaluate the MuscLab system requirements, stated in Table 4.1, regarding the (i) acquisition frequency (100 Hz); (ii) percentage of packet loss of wireless communication (below 5%); and (iii) autonomy (> 2h). The MuscLab mass and dimensions were also measured. The protocol consisted of two trials, where the MuscLab system acquired and transmitted the inertial and piezoresistive textile data at 100 Hz. The protocol started with a battery level of 100%. The battery level was measured with a multimeter at the beginning of the protocol and in 20-minute intervals until the system switched off due to lack of battery power. During the data collection, the time stamp of each received sample at the API was registered. The trial duration was recorded by a chronometer to compute the battery autonomy. After the data collection, the time between consecutive samples was computed. To infer the acquisition frequency, the average time between consecutive samples was determined. Further, all samples received with a time between consecutive samples of less than 9.9 ms or more than 10.1 ms were considered lost samples. The percentage of lost packets was calculated taking into account the received and lost samples. B. Human tests Ten healthy participants (5 males and 5 females) were involved in the validation of the MuscLab system (average age of 26.5 ± 2.8 years old, an average body mass of 65.4 ± 11.2 kg, and an average body height of 169.8 ± 10.9 cm). The lowest and highest shank perimeters measured were 35.0 and 42.1 cm, respectively, achieving an average value of 37.6 ± 1.9 cm. Prior to the experiments, all participants gave written informed consent in accordance with the ethical guidelines established by the Ethics Committee of the University of Minho (CEICVS 006/2020). The protocol started by equipping the participants, as illustrated in Figure 4.5. First, the participants were instrumented with the EMG Trigno Avanti sensors [Trigno Avanti (Delsys Incorporated, Natick, USA)] to capture EMG signals from two lower limb muscles: the tibialis anterior Chapter 4 63 and the gastrocnemius lateralis of the right leg. The sensors were positioned above the referred muscles, following SENIAM recommendations [105]. Further, the system was configured to compute the RMS of the EMG signals at 100 Hz. Moreover, participants were equipped with 7 IMUs (torso, thighs, shanks, and feet) from the Xsens Awinda system (Movella – Henderson, USA) to collect lower limb joint angles at 100 Hz. Figure 4.5 – A participant instrumented with the (i) Trigno Avanti and Xsens Awinda systems (top view); (ii) Trigno Avanti, Xsens Awinda, and MuscLab (without outer band) systems (middle view); and (iii) Trigno Avanti, Xsens Awinda, and MuscLab (with outer band) systems (bottom view). Chapter 4 64 Then, the participants were instrumented with the MuscLab system at the right shank. The MuscLab system was positioned above the sensors from the Trigno Avanti and Xsens Awinda systems, as depicted in Figure 4.5. Participants were then asked to perform the N-pose +Walk calibration procedure from the Xsens Awinda system. Participants stood in the N-pose for 3 seconds, then walked 5 meters forward and backward in a straight line. Afterward, participants were asked to sit to perform the MuscLab calibration: remain static for 10 seconds and then perform repetitive dorsiflexion and plantar flexion movements for 5 seconds. Once calibrated, participants completed three independent tests per each movement type: dorsiflexion and plantar flexion movements. In all tests, the time was controlled by an external researcher. The first test consisted of the following sequence: 1) started in a relaxed position for 5 seconds (Figure 4.6 – a)); 2) carried out five dorsiflexion repetitions (Figure 4.6 – b)) at three cadences imposed by a metronome, with each cadence separated by a 5-second relaxed position. The chosen cadences were 40, 75, and 105 beats/minute, which may correspond to walking speeds of 1.0, 2.5, and 4.0 km/h [106]. This first test aimed to analyze the system's responsiveness to different motion cadences. For that, the delay between the MuscLab signals and the signals collected by the Trigno Avanti and Xsens Awinda systems was computed, for each cadence. The second test included two constant dorsiflexion movements for 3 seconds each, separated by a 3-second relaxed position. This test aims to evaluate the MuscLab’s measuring repeatability. In this test, for both the MuscLab and Trigno Avanti systems, the percentage of the average magnitude variation between the first and the second dorsiflexion movements was computed. In the third test, the participants were instructed to execute three constant dorsiflexion movements for 2 seconds each, at different ankle dorsiflexion angles, i.e., (i) a small dorsiflexion movement; (ii) a dorsiflexion movement at an intermediate position between the relaxed and the maximum positions; and (iii) a maximum dorsiflexion movement. This magnitude test was performed to verify if the system can discriminate different muscle contraction levels at different ankle joint angles. In this connection, the Coefficient of Determination ( R2 ) was computed between the MuscLab and the Xsens Awinda signals. Subsequently, the participants were asked to perform the same three tests but for plantar flexion movements (Figure 4.6 – c)). In addition, the correlation of the MuscLab system with the EMG signals of the tibialis anterior and gastrocnemius lateralis muscles was assessed under the described dorsiflexion and plantar Chapter 4 65 flexion movements, respectively. For that, the Spearman Correlation coefficient ( r ) was used to evaluate the strength and direction of the monotonic relationship between MuscLab and EMG signals. A p -value of 0.05 was used to determine whether the observed correlation between the two systems was statistically significant. According to the study [107], the relationship can be categorized as (i) negligible if 0.0 < r < 0.19; (ii) weak if 0.20 < r < 0.29; (iii) moderate if 0.30 < r < 0.39; (iv) strong if 0.40 < r < 0.69; and (v) very strong if r ≥ . . Figure 4.6 – A participant performing a) a relaxed movement; b) a dorsiflexion movement; and c) a plantar flexion movement. 4.4 RESULTS 4.4.1 BENCH TESTS The MuscLab system showed an average (i) acquisition frequency of 100.0 ± 0.007 Hz (≅ requirement of 100 Hz); (ii) percentage of packet loss of 0.0972 ± 0.0318% (< 5%); and (iii) battery life of 2 hours and 9 minutes (> 2h). Regarding the MuscLab dimensions, the final length, width, and height of the case containing the electronic board were 5.4 cm (< 5.5 cm), 4.3 cm (< 4.5 cm), and 2.2 cm (< 2.5 cm), respectively. The flexible textile band presented a length and a width of 33.5 cm and 18.0 cm, respectively, allowing the monitoring of shank segments with perimeters ranging from 33.5 to 48.7 cm. The mass of the case and flexible textile band were 40 g and 49 g, respectively, comprising a total of 89 g (< 100 g). a) Relaxed b) Dorsiflexion c) Plantar Flexion Chapter 4 66 4.4.2 HUMAN TESTS The MuscLab system was able to monitor the muscle contraction of individuals with a shank perimeter ranging from 35.0 and 42.1 cm. Table 4.2 presents the results of the MuscLab system's responsiveness to different motion cadences. According to Table 4.2, all MuscLab signals were, on average, (i) delayed regarding the EMG Trigno Avanti signals (135.8 ± 78.0 ms); and (ii) anticipated regarding the Xsens Awinda signals (-36.8 ± 112.2 ms). Moreover, the average delay of the MuscLab signals measured by each strip was consistent across different cadences, since it varied from (i) 135.5 and 136.2 ms, regarding the EMG Trigno Avanti signals; and (ii) 34.5 and 38.8 ms regarding the Xsens Awinda signals. Table 4.2 – Average delay (standard deviation) between the MuscLab and EMG Delsys signals, and MuscLab and Xsens Awinda signals, in milliseconds, for each motion cadence (40, 70, 105 bpm) MuscLab Strip Trigno Avanti System Xsens Awinda System 40 bpm 75 bpm 105 bpm 40 bpm 75 bpm 105 bpm S1 138.0 (82.3) 125.0 (84.1) 134.0 (73.2) - 5.0 (128.4) - 4.5 (127.1) - 8.5 (121.3) S2 136.7 (71.6) 124.0 (79.3) 134.0 (94.3) - 8.0 (121.4) - 3.5 (126.7) - 4.0 (122.1) S3 129.0 (66.2) 134.0 (90.7) 120.0 (93.4) - 66.1 (72.1) - 69.4 (83.5) - 60.0 (89.4) S4 127.0 (92.5) 128.0 (95.0) 113.0 (93.2) - 76.7 (90.1) - 61.1 (104.4) - 67.8 (90.8) S5 146.7 (68.0) 170.0 (43.2) 178.3 (43.7) - 38.0 (118.2) - 46.7 (167.4) - 32.0 (120.4) Average ± STD* 135.5 (76.1) 136.2 (78.4) 135.9 (79.5) - 38.8 (106.0) - 37.0 (121.8) - 34.5 (108.8) Total Average ± STD 135.8 (78.0) -36.8 (112.2) *STD means standard deviation The results of repeatability tests focus on the comparison of the piezoresistive textile strips number 1 and 2 (S1 and S2, respectively, of Figure 4.1) with tibialis anterior signals for dorsiflexion movements. On the other hand, for plantar flexion movements, the piezoresistive textile strips number 3, 4, and 5 (S3, Chapter 4 67 S4, and S5, respectively, of Figure 4.1) were compared to the gastrocnemius lateralis signals. In this context, the tibialis anterior EMG, S1, and S2 signals revealed close average variations of muscle contraction between two performed dorsiflexion movements, namely 3.17 ± 11.7%, 2.97 ± 4.32%, and 2.41 ± 3.77%. Moreover, the gastrocnemius lateralis EMG, S3, S4, and S5 signals revealed an average variation of 4.88 ± 13.7%, 5.30 ± 8.98%, 3.47 ± 8.68%, and -20.2 ± 94.9%. In relation to the magnitude test results, an R2 of (i) 0.91 ± 0.12 and 0.92 ± 0.08 was found between the dorsiflexion ankle angle and the S1 and S2 strips, respectively; and (ii) 0.92 ± 0.15, 0.85 ± 0.18, and 0.69 ± 0.22 was found between the plantar flexion angle and the S3, S4, and S5 strips, respectively. Furthermore, Table 4.3 presents the r values obtained between the MuscLab and EMG Trigno Avanti signals, per participant. The results presented in Table 4.3 indicated very strong correlations between the MuscLab signals at S1, S2, S3, and S4 strips and EMG Trigno Avanti signals, reporting average r values above 0.78 ± 0.08 ( p -value < 0.05). In detail, very strong correlations were achieved between (i) tibialis anterior signals and both S1 and S2 signals ( r = 0.78 ± 0.07); and (ii) gastrocnemius lateralis signals and both S3 and S4 ( r = 0.78 ± 0.09 and 0.77 ± 0.07, respectively). The S5 signals were moderately correlated with the gastrocnemius lateralis signals ( r = 0.38 ± 0.23). Moreover, low standard deviation values were verified for S1, S2, S3, and S4 (< 0.09), while the S5 presented the highest standard deviation value (0.23). Additionally, Figure 4.7 depicts the EMG Trigno Avanti, Xsens Awinda, and MuscLab signals for a random participant executing the validation protocol, i.e., responsiveness, repeatability, and magnitude tests. Table 4.3 – Spearman Correlation coefficient ( r ) between the MuscLab and EMG systems. A p -value below 0.05 was verified in all correlations Participant ID Shank Perimeter (cm) Shank Length (cm) r S1 S2 S3 S4 S5 1 42.1 33.0 0.84 0.83 0.80 0.76 0.33 2 38.2 33.0 0.70 0.74 0.77 0.84 0.48 3 38.8 35.8 0.86 0.88 0.58 0.71 0.49 4 37.9 31.5 0.74 0.71 0.66 0.74 0.75 5 35.0 41.5 0.84 0.87 0.87 0.63 0.75 6 36.5 36.4 0.67 0.70 0.81 0.77 0.22 Chapter 4 68 Participant ID Shank Perimeter (cm) Shank Length (cm) r S1 S2 S3 S4 S5 7 37.4 35.0 0.74 0.75 0.79 0.83 0.17 8 37.9 35.0 0.70 0.72 0.81 0.75 0.41 9 35.3 34.4 0.88 0.85 0.89 0.84 0.07 10 37.2 33.2 0.83 0.71 0.77 0.86 0.15 Average ± STD 37.6 ± 1.9 34.9 ± 2.6 0.78 ± 0.07 0.78 ± 0.07 0.78 ± 0.09 0.77 ± 0.07 0.38 ± 0.23 Figure 4.7 – A participant performing dorsiflexion (top view) and plantar flexion (bottom view) movements. Boxes 1, 2, and 3 represent the speed, repeatability, and magnitude tests, respectively. For easier visualization, the magnitude of EMG signals was scaled to MuscLab signal magnitudes using a factor of x10,000. Chapter 4 69 4.5 DISCUSSION This study proposes a flexible and elastic textile band instrumented with piezoresistive textile sensors to monitor the muscle contraction of individuals with different anthropometries. In an attempt to address the state-of-the-art limitations, the proposed system advances (i) studies [95]–[100] by monitoring muscle contraction of different muscle groups, simultaneously; (ii) study [101] by discriminating the muscle contraction of the agonist and antagonist muscles with a single sensor; and (iii) studies [102], [103] by monitoring the muscle contraction of individuals with different anthropometries (ranging from 35.0 to 42.1 cm) due to the elastic nature of the proposed solution. In terms of responsiveness performance, the MuscLab signals were delayed regarding EMG signals (average delay of 135.8 ± 78.0 ms). According to the study [97], there is a time delay between the onset of EMG and the onset of force generation, defined as the electromechanical delay. This delay typically varies between 9 and 130 ms [108]. Since an MMG, when positioned in a muscle, measures the response of a force generation, its measured signals are expected to be delayed with respect to the EMG signals. In this context, the average delays reported by the MuscLab system (135.8 ± 78.0 ms) are in line with what would be expected because the MuscLab system is an MMG sensor; it, therefore, detects muscle contraction only after the muscle electrical activation (signal measured by EMG). Conversely, an average delay of -36.8 ± 112.2 ms was found between the MuscLab and Xsens Awinda signals. Despite the high standard deviation value, these results suggest that the MuscLab signals may be anticipated regarding ankle joint kinematics. These results may be justified by the fact that muscle contractions occur from 20 to ms before the user’s lower limb kinematic motion [108]. Furthermore, results indicated that the average delay of the MuscLab signals remained consistent across the tested cadences. These results are aligned with those presented in the study [104] since from a comparison between sixteen e-textiles, the Shieldex® Technik-tex P130+B (used in the MuscLab system) showed the quickest responses between relaxing and stretching movements. Thus, the MuscLab system can be employed to monitor muscle contractions performed at cadences between 40 and 105 bpm. In addition, MuscLab demonstrated sensing repeatability when measuring similar muscle contractions. Considering the repeatability tests, it was found that (i) the variation in the magnitude of the signals measured in strips S1 and S2 (2.97 ± 4.32% and 2.41 ± 3.77%.) was, on average, close to the variation in the magnitude of the tibialis anterior signals (3.17 ± 11.7%); and (ii) the variation in the Chapter 5 76 Figure 5.1 – SmartOs’ mobile graphical application (left view) and a male participant instrumented with the SmartOs, Trigno Avanti, and InertialLab systems (right view). C. Experimental Protocol Once instrumented with the eight EMG and seven IMU sensors, participants performed two Maximum Voluntary Contractions (MVCs) for each muscle to normalize all EMG signals, following the protocol described in the study [121]. Subsequently, participants were equipped with the SmartOs system and provided a 10 min-familiarization session with the device operating under zero-torque control. The participants then completed a continuous circuit of nine task sequences while the ankle orthosis operated in zero-torque control (Figure 5.2). The sequences, presented in Figure 5.2, consisted of: 1. standing for 10 seconds (St); 2. walking in a straight line for 2.5 meters (LGW); 3. descending stairs with 13 steps (SD); 4. walking again in a straight line for 3.5 meters (LGW); 5. stopping in the standing position for 5 seconds (St); 6. walking in a straight line for 3.5 meters (LGW); 7. ascending stairs with 13 steps (SA); 8. walking in a straight line for an additional 2.5 meters (LGW); and 9. stopping in the standing position for 5 seconds (St). This circuit was repeated five times per participant. During dynamics tasks, participants maintained a fixed cadence dictated by a metronome at 40 bpm. Depending on the participant's height and subsequent step length, this cadence corresponded to gait speeds ranging from 1.0 to 1.5 km/h, as validated in the study [106]. This range of values falls in the slow speeds typically adopted by individuals with motor disabilities [77]. Additionally, all participants were able to perform each transition with the self-selected limb, except in the SD task. At this task, participants were instructed to initiate the SD task with their right foot, making it the leading limb to advance to the next step. Following the advancement with the right Chapter 5 77 foot, the left foot was then moved to the same step. This SD procedure was followed due to two reasons. First, it aligns with the advice for patients with lower limb impairments to descend stairs with their impaired leg as the leading limb [122]. Second, the ankle orthosis is mechanically limited to a maximum dorsiflexion angle value of 20º (a value easily exceeded when descending stairs with the non-instrumented leg (left leg) being the first to advance). Figure 5.2 – A random male participant performing the four LMs (St, LGW, SD, SA) in the first scenario. Positions (1) and (9) denote the starting and ending positions, respectively. 5.3.2 DATA PREPARATION During the protocol, a researcher annotated the instant of transitioning between two consecutive LMs. This instant corresponded to a moment between the participant having lifted their leading foot off the ground and the critical moment (the instant when the participant placed the foot on the new LM [44]). This instant was recorded using the SmartOs’ mobile graphical application (Figure 5.1) to be synchronous with the remaining collected data. After data collection, all trials underwent manual inspection by two Chapter 5 78 additional researchers to ensure reliability in labeling the transition instants. At this level, four classes were defined, namely, 0 – St, 1 – LGW, 2 – SD, and 3 – SA. Considering the reviewed studies [41], [43]–[46], [62], [114]–[117], several types of data were employed in decoding LMs, namely features from EMG signals [114], [115], and segment [43]–[46], [116] and joint angles [41], [117]. There is still no consensus on what type of data yielded more accurate LM decoding. Thus, the LM decoding performance was compared using four types of sensor data: (i) RMS EMG normalized by MVC (one of the most used EMG features [114], [115]); (ii) the segment angles (torso, thighs, shanks, and/or feet) and their first and second-order derivatives, corresponding to the angular velocity and angular acceleration; (iii) the lower limb joint angles (hips, knees, and/or ankles), and their first and second-order derivatives; and, (iv) the magnitude of 3D acceleration and angular velocity of the torso, thigh, shank, and/or foot segments. In the segment or joint angle data types, their first and second-order derivatives were computed, since these variables improved the model’s performance in empirical tests. The fourth data type aims to explore the potential of directly using IMUs’ data, which was not yet explored in the state-of-the-art for LM decoding. Equation 5.1 presents an example of computing the magnitude of 3D acceleration data (‖ a ‖), in which ax, ay, a represent the acceleration measured in the x-axis, y-axis, and z-axis, respectively. ‖ a ‖ = √ ax ay a (5.1) Overall, a total of 38 input combinations were created and compared. They are identified by an ID in Table 5.1. The data was organized into sequences made up of X columns and Y rows. While X represents the number of samples organized sequentially by time and participants, the Y rows represent the number of inputs. Table 5.1 – Identification of the type of data used to decode LMs. The joint angles, magnitude of 3D raw accelerometer and gyroscope, segment angles, and EMG data are shaded in green, yellow, gray, and red, respectively ID Inputs ID Inputs 1 Ankles JAD* 20 MAG (pelvis and thighs) 2 Hips JAD 21 MAG (pelvis) 3 Ankles and hips JAD 22 Feet SAD*** 4 Knees JAD 23 Feet and shanks SAD 5 Ankles and knees JAD 24 Shanks SAD 6 Knees and hips JAD 25 Feet and thighs SAD 7 Ankles, knees, and hips JAD 26 Shanks and thighs SAD Chapter 5 79 ID Inputs ID Inputs 8 MAG** (feet) 27 Thighs SAD 9 MAG (feet and shanks) 28 Pelvis and feet SAD 10 MAG (shank) 29 Pelvis, shanks, and feet SAD 11 MAG (feet and thighs) 30 Pelvis and shanks SAD 12 MAG (shanks and thighs) 31 Pelvis, thighs, and feet SAD 13 MAG (thighs) 32 Pelvis, thighs, shanks, and feet SAD 14 MAG (pelvis and feet) 33 Pelvis, thighs, and shanks SAD 15 MAG (pelvis, shanks, and feet) 34 Pelvis and thighs SAD 16 MAG (pelvis and shanks) 35 Pelvis SAD 17 MAG (pelvis, thighs, and feet) 36 EMG (TA, GL, RF, and BF) 18 MAG (pelvis, thighs, shanks, and feet) 37 EMG (RF, and BF) 19 MAG (pelvis, thighs, and shanks) 38 EMG (TA, and GL) *JAD means Joint Angles and Derivatives (angular velocity and acceleration). **MAG means magnitude of 3D raw accelerometer and gyroscope. ***SAD means Segment Angles and Derivatives (angular velocity and acceleration). 5.3.3 CLASSIFICATION MODELS In the scope of this thesis, Long Short-Term Memory (LSTM), CNN, and Transformers were explored since these models revealed high generalization ability and high performances in previous studies and are adequate for time series [123]–[125]. All models were implemented and evaluated in Matlab® (2023b, The Mathworks, MA, USA), running in a Hewlett-Packard computer (Intel® Core™ i74710MQ CPU @ 2.50 GHz processor and 16.0 GB random access memory). Several empirical analyses were conducted to find the best LM decoding tool. Regarding LSTM, it was studied (i) the number of neurons (5, 10, 15, 20, 25, 30, 35, 40); and (ii) the number of LSTM layers (1, 2, 3). In the case of CNN, the hyperparameters studied were (i) the number of filters (8, 16, 32, 64, 128); (ii) the kernel size (10 to 100 with increments of 5); (iii) the number of convolutional layers (1, 2, 3); and (iv) the pooling layer (average and max pooling layer). The hyperparameters explored in the Transformer were (i) the number of encoder layers (1 and 2); (ii) the number of heads (2, 4, 8, 16, 32); and (iii) the embedding size (128, 256). For three DL models, the following parameters were explored: (i) dropout rate (between 0% and 80% with increments of 10%); (ii) the batch size (8, 16, 32, 64, 128, and 254); (iii) the sequence length (80, 100, 120, 140, 160, 180, and 200); and (iv) the normalization method (robust, max-min, and z- Chapter 5 80 score). Furthermore, it was utilized (i) the adaptive moment estimation optimization algorithm to optimize the parameters of the neural networks, including weights and biases [126]; (ii) the focal cross-entropy loss to address potential class imbalance [127]; and (iii) various data augmentation techniques such as jittering, scaling, magnitude warping, and a combination of scaling and magnitude warping. These data augmentation techniques were employed to increase the diversity of the available data while preserving accurate labels [128], [129]. To determine the optimal hyperparameters and the best LM decoding tool, a leave-one-subject-out cross-validation (LOSOCV) was performed. Two participants, selected randomly from the fifteen participants in the study (a 25-year-old able-bodied female with a body mass of 68.3 kg and a height of 1.65 m, and a 27-year-old able-bodied male with a body mass of 83.0 kg and a height of 1.70 m), were chosen to validate and test the model, offline, respectively. Subsequently, the model was trained using data from the remaining thirteen participants. 5.3.4 EVALUATION METRICS Six evaluation metrics were computed to assess the performance of the LM decoding tool, namely, the ACC, Matthew’s Correlation Coefficient (MCC), F1-score, success rate per class, computational load , and prediction time . The computational load (in ms) denotes the time necessary for the LM decoding tool to classify the LM upon receiving new data. The prediction time (in ms and as a percentage of the gait cycle) is determined by Equation 5.2, in which (i) Tcritical represents the critical instant; and (ii) Tprediction is instant in which the LM decoding tool performs a classification. Thus, prediction time was computed and averaged for six transitions: St-LGW, LGW-St, LGW-SD, SD-LGW, LGW-SA, and SA-LGW. The prediction time was also assessed as a percentage of the gait cycle to understand at what phase of the gait cycle the new LM is decoded. Prediction time = Tcritical Tprediction (5.2) 5.3.5 EXPERIMENTAL VALIDATION Once trained, the model was converted to the ONNX format, in order to be integrated into the highlevel of the SmartOs system. This model conversion was done since the ONNX platform increases the Chapter 5 81 portability and interoperability of AI models [130]. After integrating the trained model into the SmartOs system, the LM decoding tool was evaluated under real-time procedures, as follows. A. Participants To validate the LM decoding tool in real-time, three participants were included: two able-bodied participants (one male and one female with an average age, body mass, and body height of 26.0 ± 1.00 years old, 76.5 ± 6.5 kg, and 167.5 ± 2.5 cm) and one female participant that suffered a stroke (age: 22 years old; body mass: 51.0 kg; body height: 160.0 cm; paretic side: right; stroke type: ischemic; stroke time: 9 months; FMA-LE: 26). It is noteworthy that the model training process did not involve these participants. The proposed control strategy was validated with the two able-bodied participants. All participants provided written and informed consent, considering the University of Minho Ethics Committee (CEICVS 006/2020). B. Experimental Protocol In the beginning, participants were instrumented with the sensor configuration that provided the best classification performance, along with the ankle orthosis. To ensure familiarity with the device, a 10-minute walking trial was performed wearing the SmartOs system operating under zero-torque control. Once familiarized with the device, experiments were conducted with the SmartOs system operating under zero-torque control. In these experiments, able-bodied participants were asked to walk at different gait speeds, while the post-stroke patient was asked to walk at a self-selected speed, as summarized in Table 5.2. Firstly, participants walked at self-selected speeds. Then, participants were asked to walk at four different cadences, starting with 35 beats/minute until 50 beats/minute, with increments of 5 beats/minute. The values of the chosen cadences were higher and lower than the values chosen during the data collection performed to train the DL models. This was done to study how the LM decoding tool performs in speed conditions different from those the tool was trained. During these experiments, the able-bodied participants walked in three different scenarios and, for each scenario, walked at different gait speeds. This validation aims to explore the robustness of the developed LM decoding tool in different environments. The first scenario (Figure 5.2) consists of the same environment where data were collected to train the DL model. It consists of an indoor scenario composed of 13 steps with dimensions of 18.0 cm (height), 110 cm (width), and 31.0 cm (depth). The second scenario (Figure 5.3 – left view) represents another indoor scenario composed of Chapter 5 82 8 steps with the same dimensions as scenario 1. The third scenario (Figure 5.3 – right view) is an outdoor scenario consisting of 13 steps with dimensions of 16.0 cm (height), 500 cm (width), and 30.0 cm (depth). The protocol adopted for real-time experiments in the first scenario was the same performed during the data collection for training the DL models. For the second and third scenarios, participants completed a continuous circuit of nine task sequences, as depicted in Figure 5.3, namely: 1. standing for 10 seconds (St); 2. walking in a straight line for 3.5 meters (LGW); 3. ascending stairs (SA); 4. walking again in a straight line for 2.5 meters (LGW); 5. stopping in the standing position for 5 seconds (St), turning 180º, and stopping again in the standing position for more 5 seconds (St); 6. walking in a straight line for 2.5 meters (LGW); 7. descending stairs (SD); 8. walking in a straight line for an additional 3.5 meters (LGW); and 9. stopping in the standing position for 5 seconds (St). At task sequence number 5, the data corresponding to the turning task were posteriorly deleted, since the LM decoding was not trained under turning tasks. During this protocol, a researcher annotated the instant of transitioning between two consecutive LMs using the SmartOs’ mobile graphical application (Figure 5.1). These events were then inspected by two additional researchers to ensure accuracy in labeling the transition instants. Figure 5.3 – A healthy test participant performing the four LMs (St, LGW, SD, SA) in the second scenario (left view) and the third scenario (right view). Positions (1) and (9) denote the starting and ending positions, respectively. Chapter 5 83 Table 5.2 – Identification of the conditions for real-time tests performed by the able-bodied participants and a poststroke patient Participant Control Mode Scenario Cadence Able-bodied Zero-torque mode 1 Self-selected 35 beats/minute 40 beats/minute 45 beats/minute 50 beats/minute 2 3 Post-stroke Zero-torque mode 2 Self-selected 5.4 RESULTS 5.4.1 DATA BALANCING A total of 954,298 samples were collected to train the DL models, distributed as follows: 211,381 for the St task, 290,710 for the LGW, 221,928 for the SD, and 230,279 for the SA. This analysis is complemented by Figure 5.4, which shows the mean and standard deviation for each class, considering the number of participants. Figure 5.4 shows that the St and SD tasks represent the minority classes, while the LGW is the majority class. Nevertheless, it was observed that the mean number of samples per class is relatively consistent, ranging from 22% to 30%. Furthermore, the amount of each class per participant shows minimal variability, as indicated by low standard deviation values ranging from 1.4% to 2.3%. Based on these findings, the dataset is considered to be balanced. Figure 5.4 – Mean and standard deviation values of the sample distribution per class across subjects. St, LGW, SD, and SA mean Standing, Level-Ground Walking, Stair Descent, and Stair Ascent, respectively. Chapter 5 84 5.4.2 INPUT DATA ANALYSIS Table 5.3 shows the results obtained by the LOSOCV method for each DL model and for each set of inputs. The achieved results indicated that the lower performances for all DL models (LSTM, CNN, and Transformers) were verified when using the RMS EMG as input (ACC, MCC, and F1-score below 0.794 ± 0.078, 0.561 ± 0.131, and 0.664 ± 0.098, respectively). Among these data, it was also observed that using RMS EMG signals from the tibialis anterior and gastrocnemius lateralis muscles provided similar performances to the ones achieved when combining these muscles with rectus , and biceps femoris muscles. Within the joint angle data type, all DL models showed their weakest performance (average ACC, MCC, and F1-score between 0.915 and 0.954, 0.796 and 0.883, and 0.846 and 0.912, respectively) when using the angles of both ankles together with their derivatives (angular velocities and accelerations) as input. Conversely, the models performed best when the angles, angular velocities, and angular accelerations of the knees and hips were used as input (average ACC, MCC, and F1-score between 0.976 and 0.980, 0.937 and 0.950, and 0.953 and 0.962, respectively). Concerning the dataset composed of the magnitude of 3D data derived from accelerometers and gyroscopes, the three DL models showed their worst performance when using these input data from pelvic IMU only (average ACC, MCC, and F1-score between 0.482 and 0.900, 0.306 and 0.765, and 0.318 and 0.820, respectively). Conversely, the best performance was observed with the magnitude of 3D acceleration and angular velocity from the pelvic, thigh and shank IMUs for the LSTM and Transformer models (average ACC, MCC, and F1-score between 0.968 and 0.979, 0.917 and 0.947, and 0.938 and 0.960, respectively), and from the thigh and shank IMUs specifically for the CNN (average ACC, MCC, and F1-score of 0.975 ± 0.007, 0.935 ± 0.019, and 0.952 ± 0.014, respectively). Finally, when considering the data type combining the segment angles along with their derivatives, for all DL models, it was verified that using (i) the pelvic segment angles and their derivatives reduced the models' performance (average ACC, MCC, and F1-score between 0.732 and 0.794, 0.471 and 0.571, and 0.581 and 0.662, respectively); and (ii) the thigh and shank segment angles and their derivatives improved the model performance (average ACC, MCC, and F1-score between 0.979 and 0.982, 0.946 and 0.953, and 0.960 and 0.965, respectively). Overall, the type of data that provided the best performance across all DL models was achieved using the thigh and shank segment angles along with their derivatives as input data. Chapter 5 85 Table 5.3 – LOSOCV metrics for all input combinations and DL models. The best and the worst results are colored in blue and red, respectively. The highest performance achieved is shaded in green color Input ID LSTM CNN Transformer ACC* MCC* F1score* ACC* MCC* F1score* ACC* MCC* F1score* 1 0.954 (0.023) 0.883 (0.051) 0.912 (0.040) 0.940 (0.023) 0.853 (0.051) 0.889 (0.040) 0.915 (0.030) 0.796 (0.064) 0.846 (0.051) 2 0.967 (0.023) 0.917 (0.054) 0.937 (0.042) 0.967 (0.019) 0.915 (0.046) 0.936 (0.035) 0.931 (0.039) 0.837 (0.079) 0.874 (0.066) 3 0.978 (0.013) 0.945 (0.032) 0.958 (0.025) 0.971 (0.024) 0.928 (0.055) 0.946 (0.043) 0.969 (0.019) 0.921 (0.046) 0.941 (0.036) 4 0.973 (0.010) 0.931 (0.026) 0.949 (0.019) 0.966 (0.014) 0.913 (0.034) 0.935 (0.026) 0.950 (0.024) 0.877 (0.052) 0.907 (0.042) 5 0.974 (0.016) 0.934 (0.039) 0.950 (0.030) 0.968 (0.024) 0.920 (0.055) 0.940 (0.043) 0.965 (0.001) 0.913 (0.059) 0.934 (0.046) 6 0.980 (0.010) 0.950 (0.024) 0.962 (0.019) 0.979 (0.013) 0.945 (0.031) 0.959 (0.024) 0.976 (0.010) 0.937 (0.027) 0.953 (0.020) 7 0.978 (0.011) 0.944 (0.028) 0.958 (0.022) 0.975 (0.015) 0.935 (0.036) 0.951 (0.028) 0.974 (0.012) 0.933 (0.031) 0.950 (0.024) 8 0.959 (0.011) 0.895 (0.028) 0.921 (0.021) 0.951 (0.017) 0.875 (0.039) 0.907 (0.030) 0.898 (0.025) 0.757 (0.052) 0.816 (0.041) 9 0.975 (0.011) 0.936 (0.029) 0.952 (0.022) 0.973 (0.010) 0.930 (0.026) 0.948 (0.019) 0.954 (0.011) 0.883 (0.026) 0.909 (0.020) 10 0.974 (0.011) 0.933 (0.028) 0.950 (0.021) 0.969 (0.011) 0.922 (0.027) 0.941 (0.021) 0.917 (0.021) 0.799 (0.045) 0.849 (0.036) 11 0.971 (0.009) 0.926 (0.024) 0.945 (0.018) 0.971 (0.009) 0.926 (0.023) 0.944 (0.018) 0.965 (0.009) 0.910 (0.023) 0.932 (0.017) 12 0.978 (0.007) 0.944 (0.019) 0.958 (0.015) 0.975 (0.007) 0.935 (0.019) 0.952 (0.014) 0.964 (0.013) 0.910 (0.032) 0.932 (0.025) 13 0.970 (0.007) 0.924 (0.019) 0.943 (0.014) 0.962 (0.010) 0.904 (0.025) 0.928 (0.019) 0.951 (0.018) 0.878 (0.042) 0.908 (0.032) 14 0.961 (0.014) 0.901 (0.035) 0.926 (0.027) 0.954 (0.022) 0.884 (0.053) 0.913 (0.040) 0.930 (0.021) 0.827 (0.047) 0.869 (0.037) 15 0.974 (0.014) 0.933 (0.034) 0.950 (0.026) 0.970 (0.013) 0.924 (0.032) 0.943 (0.025) 0.921 (0.032) 0.810 (0.065) 0.855 (0.054) 16 0.970 (0.015) 0.923 (0.037) 0.942 (0.029) 0.974 (0.007) 0.933 (0.018) 0.950 (0.014) 0.939 (0.016) 0.848 (0.037) 0.885 (0.029) 17 0.971 (0.012) 0.925 (0.031) 0.944 (0.023) 0.967 (0.016) 0.916 (0.038) 0.937 (0.030) 0.963 (0.014) 0.906 (0.034) 0.929 (0.026) 18 0.973 (0.014) 0.931 (0.033) 0.948 (0.027) 0.971 (0.011) 0.924 (0.027) 0.944 (0.020) 0.966 (0.016) 0.915 (0.038) 0.936 (0.029) 19 0.979 (0.008) 0.947 (0.020) 0.960 (0.015) 0.970 (0.014) 0.924 (0.035) 0.943 (0.027) 0.968 (0.013) 0.917 (0.031) 0.938 (0.024) Chapter 5 92 derivatives for LM decoding revealed ACCs of 0.976. The use of ankle angles and their derivatives provided the worst performances (0.915 < ACC < 0.954). This result may be explained by the fact that the flexion/extension of the knee and hip, and consequently the range of motion of these joints, present a higher variation between the different LMs, whereas the angle of the ankle is more identical across LMs. With respect to the dataset composed of the magnitude of 3D data derived from accelerometers and gyroscopes, the use of pelvic, thigh, and shank IMU sensors revealed the best results (0.970 < ACC < 0.979). Despite this type of data being not used in the reviewed studies, the IMU combinations that provided the best results (ACC ranging from 0.932 to 0.984) were those positioned in the same segments (pelvis, thighs, and shanks) used in studies [41], [43]–[46], [116], [117]. Moreover, the results indicated the worst LM decoding performance when using the magnitude of the 3D accelerometer and gyroscope data from the pelvic IMU only (0.482 < ACC < 0.900). This can be explained by the fact that the pelvic segment is the most stable segment during locomotion. Thus, it may not present significant changes in such a way that it is easy to distinguish which LM the user is in. Considering segment angle data type, the use of pelvic segment angles and their derivatives as input showed the lowest performances (0.732 < ACC < 0.794), as verified when using the magnitude of 3D accelerometer and gyroscope data. On the other hand, the highest performance was verified when using the thigh and shank segment angles and their derivatives as input (0.979 < ACC < 0.982). These results are aligned with those achieved in studies [43], [44], since the use of thigh and shank segment angles and derivatives revealed ACCs ranging from 0.930 and 0.983, respectively. Overall, there seems to be a tendency to achieve high LM decoding performances when using kinematic information extracted from IMU sensors placed at the shank and thigh segments (i.e., hip and knee joint angles, thigh and shank segment angles, or magnitude of accelerometer and gyroscope data at these segments). This tendency could be attributed to the high variation in knee and hip flexion and extension, as well as the range of motion of these joints, across different LMs. 5.5.2 OFFLINE PERFORMANCE EVALUATION From the benchmark analysis between types of data and DL algorithms, the best LOSOCV results were achieved for an LSTM fed by thigh and shank segment angles and derivatives Chapter 5 93 (average ACC of 0.982 ± 0.009). From the literature analysis, the LSTM model was not employed by studies [41], [43]–[46], [62], [114]–[117] in decoding LMs when using wearable assistive devices. A direct comparison cannot be made between the results of this study and the literature, as the protocols and the classified LMs differ among them. Nonetheless, the obtained performances are close to the ones obtained in the study [45] (ACC = 0.984). The study [45] used a Multilayer Feedforward Neural Network to decode ST, LGW, SA, SD, RA, and RD tasks. Despite decoding RA and RD tasks, the study [45] required a training step of 18 minutes for each user. On the other hand, the achieved average ACC was promising since it was closer to the highest ACC found in the literature (99.7%) [62]. In addition, both the LOSOCV and the model offline test performances indicate that the proposed LM decoding tool demonstrates strong generalization capabilities across participants. 5.5.3 ONLINE PERFORMANCE EVALUATION Overall, the results obtained during real-time test conditions for both health and stroke subjects are similar to the ones achieved during the LOSOCV and offline test procedures. Despite being slightly inferior, the results obtained for the stroke patient (ACC = 0.972) are comparable to those achieved for the able-bodied participants (ACC = 0.986). This slight reduction in the evaluation metrics was also verified in the study [43] and it may be related to the asymmetrical and abnormal gait patterns typically performed by stroke patients [7]. The obtained model showed potential for decoding LMs with a stroke patient, although it was trained on data from able-bodied participants. On the other hand, none of the existing tools [41], [43]–[46], [116], [117] evaluated the model performance at the preferred speeds of neurologically impaired users (below 2.7 km/h). At this level, the proposed LM decoding tool was demonstrated to be robust to different cadences, corresponding to gait speeds ranging from 1.0 to 1.6 km/h (0.988 < ACC < 0.991). Additionally, in an attempt to advance studies [41], [43]–[46], [116], [117], the robustness of the model was analyzed in relation to different scenarios. The DL decoding tool was tested in different environments (indoor and outdoor), with a different number of steps (8 and 13), with different step dimensions (height: 16.0 - 18.0 cm; width: 110 – 500 cm; and depth: 30.0 - 31.0 cm), and with a different order of task execution. The proposed LM decoding tool proved to perform consistently in scenarios other than those adopted during tool training, revealing ACCs ranging from 0.978 to 0.991. Chapter 5 94 Furthermore, the SD and SA tasks provided the highest success rate for both able-bodied and stroke subjects, with values ranging from 95.7% to 99.1%. These results are in line with those pointed out by studies [44]–[46], [117]. The St task presented the lowest success rate (88.9% - 93.8%). This result may be explained by the fact that in the transition between St and LGW, a small intention to start walking (e.g., by bending the knee) is immediately detected by the LM decoding tool as the LGW task, although the user still has both feet on the ground (the ground truth is still St task). This phenomenon may reduce the success rate in the St task. The proposed LM decoding tool presented a low computational load (1.58 ± 0.42 ms) to perform classification upon receiving new sensor data. This computational load is within the timing requirements of the high-level of the SmartOs system (10 ms, operating at 100 Hz). Considering that the natural frequency of human motion during walking tasks is less than 2 Hz (i.e., 500 ms), the proposed LM decoding tool works fast enough to decode human LMs [132]. It is noteworthy that the assertiveness of LM decoding tools is critical since misclassifications can cause the wearable assistive device to adopt inappropriate gait patterns, potentially causing discomfort to the users [14]. However, the adoption of inappropriate gait patterns can result not only from misclassifications but also from delays in decoding the new LMs [119]. Despite the high accuracies reported, most of the reviewed studies recognize the new LM after the user has already entered on it [41], [45], [46], [62], [114]–[117]. The decoding delays typically range from 50.0 to 1897.9 ms. Only two studies ([43], [44]) demonstrated the ability to predict specific LMs. The study [44] successfully predicted the transition from (i) SA to LGW in 10.7 ± 9.88 ms; (ii) SD to LGW in 78.7 ± 130.0 ms; and (iii) LGW to SA in 185.3 ± 56.9 ms. However, the detection of the LGW-SD transition occurred 40.0 ± 107.5 ms after the critical instant. In the study [43], the following transitions were predicted in advance: (i) SD-LGW at 0.70 ± 58.7 ms, (ii) LGW-SA at 5.40 ± 74.6 ms, (iii) SA-LGW at 46.7 ± 46.6 ms, and (iv) LGW-SD at 78.5 ± 25.0 ms. Nevertheless, the study [43] required offline training to create a userspecific decoding model for each participant prior to real-time experiments. In light of these findings, the proposed LM decoding tool outperforms the aforementioned studies ([41], [43]–[46], [62], [114]–[117]) by predicting upcoming LMs with an average prediction time of 482 ± 227 ms. Further, the average prediction time per transition was always positive (ranging from 241 to 860 ms), which means that, across the six possible transitions (St-LGW, LGW-SA, SA-LGW, LGW-SD, SD-LGW, and LGW-St), the upcoming LM was decoded in advance, on average. Translating these results as a percentage of the gait cycle (from 72.3% to 93.1%), it can be assumed that the next LM was decoded, on average, between the midand terminal swing of the gait cycle preceding the upcoming LM [133]. Chapter 5 95 In addition, the St-LGW transition was decoded with a higher prediction time (860 ± 178 ms). This result could be explained by the possibility that during the St task, subtle movements could indicate the imminent onset of the LGW task. Conversely, the LGW-St transition presented the lowest prediction time (241 ± 294 ms). The high standard deviation value suggests that, at this transition, the participants may be not timely assisted by the SmartOs system in some cases, i.e., the assistance could change after the critical instant. This can be related to the fact that the LGW-St transition may involve a more abrupt shift by the user, making it difficult to predict the St task. 5.6 CONCLUSIONS This work offers two contributions to the state of the art: (i) to investigate the best fusion of sensor data and classification algorithms to build a DL decoding tool capable of predicting in advance four LMs (St, LGW, SD, and SA) and transitions between them; and (ii) to develop an LM decoding tool the timely decode four LMs in advance. This research concludes that, for decoding LMs, the use of (i) RMS EMG features is not sufficient; (ii) thigh and shank segment angles as input to an LSTM provide the best performances. Furthermore, the proposed tool was found to be accurate in both offline and online testing with healthy and stroke participants, presenting a low computational load. In addition, the tool proved to be robust to different walking cadences and different scenarios. These conclusions highlight the suitability of the developed LM decoding tool to foster the development of LM-driven trajectory control strategies (detailed in Chapter 6) that timely adapt the assistance of wearable assistive devices according to the LM decoded. Although the proposed LM decoding tool showed promising results, there are still opportunities for improvement. Specifically, there is a need to (i) conduct additional real-time validation procedures with more participants, including both able-bodied individuals and those with varying degrees of neurological impairments; (ii) extend the decoding capabilities to other LMs, particularly RA and RD, and transitions; and, (iii) improve the prediction performance for LGW-St transitions to increase the prediction time . 96 6. ASSIST-AS-NEEDED CONTROL STRATEGIES Chapter 6 97 This chapter begins with an introduction regarding the trends of assistive control strategies applied for robot-based gait training. It then describes the implemented hierarchical control architecture, focusing on the development and validation of the assistive control strategies investigated in this thesis: AAN LMdriven trajectory control, AAN EMG-based control, and AAN HITL control. These strategies were designed, implemented, and validated with healthy participants using the SmartOs system. The potential rehabilitation benefits of each assistance strategy are reviewed and discussed. The chapter concludes with a comprehensive overview of the proposed assistive control strategies. 6.1 INTRODUCTORY INSIGHT Bio-inspired control architectures have begun to emerge, inspired by the principles and organization of the human motion control system [9]. These architectures aim to efficiently implement user-oriented control strategies and facilitate synergistic and continuous interaction between the user and the wearable assistive device, thereby enhancing brain plasticity [134]. For that, these architectures may include low-, mid-, and high-level controls that are hierarchically organized to address the physical interactions between the user, the environment, and the device. The design and development of the hierarchical control architecture of the SmartOs system followed these research directions and exploited the potential of different assistive control strategies, as described in this chapter. According to studies [37], [134], advanced technology should be used rationally. It seems that post-stroke patients may benefit more from training with wearable assistive devices when the active participation of the patient is invoked. These results may be due to the fact that spontaneous neuroplasticity phenomena occur with greater intensity in the initial acute and sub-acute phases of rehabilitation after stroke. In this sense, patients' active participation should be encouraged to achieve rapid neuromuscular recovery during rehabilitation training. Recent trends suggest that user-oriented assistance that encourages user participation may be achieved through the use of AAN control strategies integrated into wearable assistive devices [29]–[31], [33], [37]. AAN control strategies are focused on dynamically adjusting the level of assistance based on the user's real-time performance and locomotion needs, ensuring that the wearable assistive device assists movement only when and as much as required. The goal is to encourage the user's active participation and muscle engagement while correcting the joint pattern during the execution of the walking motion [29]–[31], [33], [37]. Chapter 6 98 This Ph.D. thesis progresses the current literature by proposing the development of three AAN control strategies, as detailed below: 1. Locomotion Mode-driven Trajectory Control: a control responsible for mapping the user’s locomotion intention and generating reference trajectories according to the desired/intended LM; 2. Assist-As-Needed Electromyography-based Control: a control combining position and torque controllers to adjust the orthosis assistance according to the user's muscular needs and joint pattern; 3. Human-in-the-Loop Control: a control designed to optimize the user metabolic cost by modifying control variables, namely, torque peak magnitudes. 6.2 LOCOMOTION MODE-DRIVEN TRAJECTORY CONTROL 6.2.1 CRITICAL ANALYSIS OF RELATED WORK Often, wearable assistive devices have been equipped with non-intrusive sensor configurations and sophisticated algorithms to decode LMs, with the aim of tailoring the robot's assistance to help users perform their daily mobility tasks. From the literature, only five studies have adapted the assistance of the wearable assistive device attending to the user’s decoded LM [40], [42], [43], [116], [117]. In the studies [40], [42], a two-level hierarchical control design was followed. In the study [40], a constant torque of 10 N.m was provided by an ankle orthosis (low-level) when the SD task was recognized (high-level). For that, FSR sensors embedded in the orthosis foot were used to detect the gait cycle phase. Then, the LM was decoded, and the constant torque was provided. The torque was delivered during (i) the first 50% of the gait cycle to support the ankle joint motion when the leg without assistance was in the swing phase; (ii) the last 20% of the gait cycle to guarantee that the foot was performing plantar flexion before touching the next stair step. In the study [42], a two-level hierarchical control was designed to provide a phase-locked assistive torque (low-level) when an LM (among the Sit, St, LGW, SA, and SD tasks and transitions between them) was recognized (high-level). After recognizing the LW, SA, and SD tasks, a gait phase estimation algorithm was employed such that the most suitable phase-locked assistive torque was provided. The provided torque was adopted from public databases with joint trajectories recorded during locomotion-related activities [133], [135]–[137]. During the Sit and St tasks, no torque was provided. Nonetheless, torque Chapter 6 99 patterns modeled according to the body mass of each participant were delivered during Sit-St or St-Sit transitions. These transitions were detected based on the hip joint angles, using a threshold algorithm. On the other hand, the studies [43], [116], [117] followed a three-level hierarchical control architecture. In the study [117], the high-level presented a gait phase estimation algorithm using IMU signals to segment the input data into strides. Based on these strides, the LM (LGW, SA, and SD) was decoded through a CNN. A Parameter Optimal Iterative Learning Control method was implemented in the mid-level. This method was designed for a soft lower limb exoskeleton to provide hip and knee assistance according to the tension of Bowden cables. The tension of these cables was modeled according to the hip and knee joint moments when performing the LW, SA, and SD tasks. The low-level was responsible for driving the exoskeleton according to the tension of the Bowden cables. The high-level developed in the study [43] was different from the one developed in the study [117]. Firstly, the LM was distinguished between static and dynamic tasks at the high-level. Then, in the case of a dynamic task, the LM was classified (among LW, SA, and SD tasks), followed by a gait phase estimation algorithm to distinguish the stance from the swing phase. During the stance phase, three other subphases (heel-strike, heel-off, and toe-off) were identified at the high-level. In addition, the mid-level developed in the study [43] was hybrid. It used a zero-torque mode during the St task and the swing phase of LW, SA, and SD tasks in an attempt to reduce the constraints associated with the actuator’s frictions. Moreover, during the first and second subphases (from the heel strike to heel-off events and from heel-off to toe-off events, respectively) of the SA task, the knee should perform the extension and flexion movements, respectively. Based on these assumptions, the knee exoskeleton provided a closed-loop torque control to extend and flex the knee joint, assisting the user during the first and the second subphases of the SA task. For the stance phase of the LW and SD tasks, the muscles around the knee joint provide negative torque to help the knee to move and avoid excessive knee flexion, supporting the human body. For this reason, an open-loop damping control was adopted to provide resistance at the knee joint to support the body and absorb the shock. The low-level was responsible for generating assistive commands as a response to the comparison between the desired torque (generated by the mid-level) and the measured torque (measured by an embedded load cell). In the study [116], another three-level hierarchical control was adopted to assist the St, LGW, SA, and SD tasks, and transitions between them. The high-level was responsible for decoding LMs based on a threshold algorithm (to distinguish between static and dynamic tasks) and a Fuzzy-logic-based algorithm (to distinguish between LGW, SA, and SD tasks). The mid-level controller employs adaptive oscillators to Chapter 6 100 track the desired torque curve, while the low-level controller drives the motors to maintain the force feedback loop. Additionally, providing efficient assistance according to the decoded user’s LM implies an accurate and timely identification of the user’s intentions. This is of utmost importance since the higher the anticipation time in identifying the LM, the more time remains to switch the control to assist the users according to their needs timely. This means that, ideally, the upcoming LM should be predicted before its occurrence, in order to adapt the assistance of the wearable assistive device according to the LM decoded. Considering the five studies that adapted the assistance according to the decoded LM, only the study [43] demonstrated the ability to predict and assist specific LMs. The study [43] successfully predicted the transition from (i) SD-LGW at 0.70 ± 58.7 ms, (ii) LGW-SA at 5.40 ± 74.6 ms, (iii) SA-LGW at 46.7 ± 46.6 ms, and (iv) LGW-SD at 78.5 ± 25.0 ms. Nevertheless, the study [43] required offline training to create a user-specific decoding model for each participant prior to real-time experiments, and the assistance was only provided during the stance phase. Despite recent advances, it is still unclear how to control the device to properly and timely assist users according to their locomotion intentions. The available studies are limited by (i) only actively assisting during specific phases of the gait cycle (stance/swing or between gait cycle events) [40] [43] or LMs [42], [117]; (ii) the adaptation the control according to the decoded LM occurs after the critical instant (i.e., after the user enters on the new LM) [40], [42], [43], [116], [117]; (iii) testing the control performance in a single scenario, which is commonly used to train the LM decoding tool; and (iv) only address the walking speeds typically adopted by able-bodied users. This research advances the current state-of-the-art by proposing a three-level assistive control strategy to timely adapt the assistance of the SmartOs system according to the decoded LM (among St, LGW, SA, and SD), while actively assisting throughout the entire gait cycle. For that, the proposed control strategy integrates the DL tool developed in Chapter 5 into the high-level of the proposed control strategy. The robustness of the proposed LM-driven trajectory control was assessed (i) in different scenarios (indoor and outdoor); and (ii) in a range of speeds, all within the preferred values of stroke patients. 6.2.2 METHODOLOGY A. Hierarchical Control Architecture Once the optimal fusion between sensor data and DL algorithms has been found for decoding LMs (Chapter 5), a hierarchical three-level control strategy was developed to adapt the assistance of Chapter 6 101 the SmartOs system according to the output of the LM decoding tool (i.e., a classification between St, LGW, SA, or SD). The proposed control strategy is presented in Figure 6.1. Figure 6.1 – Proposed hierarchical control architecture. LM is the predicted LM. ϴ ref.h and ϴ ref.o are the human and orthosis ankle joint reference angles, respectively. e ϴ is the orthosis position error. u is the Proportional-Integral-Derivative (PID) command. SS is a speed scaling block achieved by (2). PID stands for the PID controller. The high-level runs the LM decoding tool. As depicted in Figure 6.1, new input data from the InertialLab system arrives at the LM decoding tool ( LM Decoding Tool block) at 100 Hz. Therefore, the DL model provides a new classification every 10 ms. This classification is used to update a 3sample buffer that contains the two previous classifications. The output of the 3-sample buffer was used together with the gait cycle phase to determine the human reference trajectory of the ankle joint angle (ϴ ref.h ) through the Trajectory Setting block. This block is responsible for updating the ϴ ref.h according to the decoded LM. The set ϴ ref.h will assist the user in performing the decoded LM, i.e., a trajectory that supports the user during the St, LGW, SA, or SD tasks. The trajectories (depicted in Figure 6.1) were obtained from the dataset collected for developing the LM decoding tool (Chapter 5). Next, the Speed Scaling block that operates at 100 Hz, receives the ϴ ref.h generated by the Trajectory Setting block and the gait speed. This occurs at the mid-level of the architecture. The ϴ ref.h LM Decoding Tool Sensor Data Type Trajectory Setting SS Orthosis uman PID u meas o e ref o Mid level ( )Low level ( k ) igh level ( ) LM ref h ait speed LM LM is t LM is L LM is LM is ref h ref h ref h ref h rajectory etting Gait Cycle ( ) Ankle Angle (deg) Gait Cycle ( ) Ankle Angle (deg) Gait Cycle ( ) Ankle Angle (deg) Gait Cycle ( ) Ankle Angle (deg) Chapter 6 108 (ii) depend on Hill-type models to estimate the user’s joint torque, which may result in complex, userdependent calibration, and time-consuming methods [28], [35], [36], [38], [145]; (iii) are not adaptable since they rely on fixed assistance ratios to define the desired joint torque trajectories [28], [38], [145]; and (iv) were not developed to assist the walking motion, but rather focus on isolated flexion and extension movements of the lower limb joints [28], [35]–[38]. This Ph.D. thesis tackles the above-mentioned limitations by proposing an AAN EMG-based control strategy to automatically assist the walking motion throughout the entire gait cycle. The system implements an outer loop torque control and an inner loop position control, which adapts the level of assistance to be provided to the user in real-time and automatically. To do this, the system presents two DL regressors: (i) one regressor to define user-oriented torque reference trajectories based on the user's body height, body mass, and gait speed [146], without the need for using fixed assistance ratios; and (ii) another regressor to determine the user's engagement (estimating the user’s joint torque, in real-time, through a DL regressor fed by a fusion of kinematic, EMG, anthropometric, and demographic data [65]). The control strategy was designed, implemented, and validated into the SmartOs system to assist the ankle joint while considering the user's engagement. 6.3.2 METHODOLOGY A. Hierarchical Control Architecture The AAN EMG-based control proposed in this research follows a hierarchical architecture organized into high-, mid-, and low-levels, as depicted in Figure 6.4. It combines a torque outer loop and a position inner loop control. In this approach, the outer loop introduces a "softening" effect in the human-robot interaction, while the inner loop ensures joint stiffness. This helps suppress undesirable disturbances, eliminating the necessity for compensation based on anticipatory control models [27]. The high-level controller, implemented at 100 Hz, is responsible for (i) defining the human ankle joint reference angle (ϴ ref.h ) and torque trajectories (Ԏ ref.h ); (ii) estimating the human ankle joint torque trajectory (Ԏ est.h ), in real-time, through the Human-Torque Estimation (HTE) block; and (iii) determine the variation of the Ԏ ref.h and Ԏ est.h ( ∆ Ԏ ref.h and ∆ Ԏ est.h , respectively). The ϴ ref.h is generated by a regression model proposed in [147], considering the user’s body height and gait speed. The methodology for defining Ԏ ref.h is defined in sub-section B. Estimation of Ankle Joint Reference Torque Trajectories . In the proposed control strategy, the Ԏ ref.h is defined according to the ϴ ref.h and its Chapter 6 109 derivatives (angular velocity and angular acceleration), gait speed, anthropometric (body height and mass, shank, and foot lengths), and demographic (gender and age) data. The HTE block is explained in detail in sub-section C. Estimation of Real-time Ankle Joint Torque Trajectories . The HTE block estimates the Ԏ est.h in less than 2 ms, considering the EMG signals from the tibialis anterior and gastrocnemius lateralis muscles (normalized by the corresponding MVCs), hip joint kinematics, walking speed, anthropometric (body height and mass, shank, and foot lengths), and demographic (gender, age) data. Figure 6.4 – AAN EMG-based torque control, combining a torque outer loop with a position inner loop. Ԏ ref.h and Ԏ est.h are the human ankle joint reference and estimated torques, respectively. ∆ Ԏ ref.h and ∆ Ԏ est.h are the variations of the Ԏ ref.h and Ԏ est.h , respectively. ∆ Ԏ ref.o and ∆ Ԏ est.o are the ∆ Ԏ ref.h and ∆ Ԏ est.h interpolated to the low-level frequency, respectively. e Ԏ .o is the orthosis torque error between ∆ Ԏ ref.h and ∆ Ԏ est.h . ϴ ref.h and ϴ ref.o are the human and orthosis ankle joint reference angles, respectively. e ϴԎ .o is the orthosis torquederived position error. ϴ meas.o is the orthosis ankle joint measured angle. e ϴ .o is the orthosis position error. u is the PID command. EMGTA and EMGGAL are the EMG signals measured from the tibialis anterior and gastrocnemius lateralis muscles, respectively. ϴ meas.h is the human hip joint measured angle. GS is the gait speed. BH and BM are the human body height and mass, respectively. SL and FL are the shank and foot lengths, respectively. K is a fixed factor equal to 1.33 N.m/radians. SS is a speed scaling block achieved by Equation 6.1. PID is a Proportional-Integral-Derivative (PID) controller. HTE is a Human-Torque Estimation (HTE) block. In the mid-level controller, which operates at 100 Hz, the human ankle joint reference angles (ϴ ref.h ) , ∆ Ԏ ref.h , and ∆ Ԏ est.h are interpolated (achieving the orthosis ankle joint reference angles (ϴ ref.o ) , ∆ Ԏ ref.o , and ∆ Ԏ est.o , respectively) to the low-level control frequency through the Speed Scaling (SS) block (see Equation 6.1). The ∆ Ԏ ref.o and ∆ Ԏ est.o are compared, generating the orthosis torque error ( e Ԏ .o ). The e Ԏ .o is converted into an orthosis torque-derived position error ( e ϴԎ .o ) through a fixed scaling est h e o K Exo uman ref h e o PID M M L meas h u e o TE meas o Mid level ( ) igh level ( ) ref o M L L ender ge est h ref h ref h SS SS est o ref o Chapter 6 110 value K set at 1.33 N.m/radians. This value was empirically found for the used AO and it represents the rotational stiffness. At the low-level controller (working at 1 kHz), it was empirically verified that was required to apply a feedforward position control to prevent the ankle orthosis from moving to positions that could cause injury to the user and could damage itself. This feedforward position control was also proposed in a study [27] to improve trajectory tracking. To this end, a feedforward position control was applied to the inner loop, adding the ϴ ref.o that the orthosis should move to while taking into account the Ԏ ref.h of the outer loop. The e ϴԎ .o , ϴ ref.o , and the orthosis ankle joint measured angle (ϴ meas.o ) are compared, generating the orthosis position error ( e ϴ.o), fed to the PID controller. The PID controller was implemented using proportional, integral, and derivative gains of 95, 1.5, and 1.5, respectively. Considering the e ϴ, the PID controller computes a PID command ( u ) that is limited to maximum and minimum values of 2500 and -2500, through a saturator. This command is interpreted by the ankle orthosis, generating the corresponding motor torque. With this control architecture, it is intended that if the user performs (i) an ankle joint angle and torque trajectories closer to the references, the assistance of the ankle orthosis should be minimal; (ii) an ankle joint angle and/or torque trajectories different from the references, the ankle orthosis should provide assistance in order to compensate for these differences. B. Estimation of Ankle Joint Reference Torque Trajectories Since the proposed AAN EMG-based control strategy must follow a user-centered design, it was required that the AO provides assistance oriented to the user's needs. This may be achieved by defining a control reference variable, i.e., ankle joint torque, adapted to each user. In this connection, regressor-based DL models were trained to estimate ankle joint reference torque trajectories (Ԏ ref.h ), according to the user’s body height, body mass, and gait speed. In this Ph.D. thesis, the focus is given to CNN. Further details of estimating user-oriented torque reference trajectories with the LSTM model can be found in the study [146]. To train the DL models, a data collection was performed. The study involved thirteen ablebodied adult participants (6 males and 7 females) with an average age of 24.2 ± 1.85 years, an average body mass of 65.2 ± 10.3 kg, and an average body height of 168 ± 12.0 cm. Balanced gender distribution was tackled considering possible biomechanical gender differences [148]. Before Chapter 6 111 undergoing the experiments, each participant gave written and informed consent according to the ethical conduct of the University of Minho Committee (CEICVS 006/2020). As depicted in Figure 6.5, participants were instrumented with a total of 12 pairs of retroreflective markers, which combined with a twelve-camera motion-capture system (Oqus; Qualisys— Motion Capture System, Göteborg, Sweden) and five force platforms embedded on the floor (FP4060; Bertec, Ohio, OH, USA) were utilized to collect the lower limb joint angles and torques (at the ankle, knee, and hip joints). Moreover, it was studied how the use of EMG signals affects the ankle joint torque estimation. For that, an 8-channel surface EMG system [Trigno Avanti (Delsys Incorporated, Natick, USA)] was employed to collect EMG data from the tibialis anterior , gastrocnemius lateralis , biceps femoris , and vastus lateralis muscles. Each subject was instructed to perform ten forward walking trials, walking sequentially at seven controlled walking speeds (1.0, 1.5, 2.0, 2.5, 3.0, 3.5, and 4.0 km/h). These speeds were controlled by a metronome and were chosen to cover the typical walking speeds of post-stroke patients and able-bodied users. The participants were asked to look ahead and walk naturally according to the beats of the metronome. Further details of the protocol are presented in [65], [121]. Figure 6.5 – EMG configuration and marker-set adopted: (A) vastus lateralis ; (B) tibialis anterior ; (C) biceps femoris ; (D) gastrocnemius lateralis ; (1) Anterior superior iliac spine; (2) trochanter; (3) thigh; (4) lateral knee; (5) medial knee; (6) shank; (7) lateral ankle; (8) medial ankle; (9) foot metatarsal 5; (10) foot metatarsal 1; (11) foot metatarsal 2; (12) posterior superior iliac spine. Out of the thirteen participants, one was randomly selected to test the trained model. A LOSOCV method among the remaining twelve participants was implemented to evaluate the model Chapter 6 112 generalization and optimize the model hyperparameters. For the CNN, the parameters explored were (i) the kernel size (2 × 2, 5 × 5, and 10 × 10); (b) the number of filters per convolutional layer (8, 16, 32, 64); and (c) the number of convolutional layers (1, 2, 3). Further, an empirical analysis was applied to select (i) the normalization method (considering the max-min, z-score, and robust normalization methods); (ii) the optimal batch size; and (iii) the dropout percentage. In addition, both neural networks’ weights and biases were updated according to an adaptive moment estimation optimization algorithm (ADAM) considering the mean square error (MSE) ([36,51]). The regression model performance was evaluated during the LOSOCV and model testing procedures. Three metrics were determined, namely, r , R2 , and the computational load (in milliseconds). The predictions performed by the trained model were also evaluated using the Bland– Altman Plot [151]. All predictions were performed in a Hewlett-Packard computer with an Intel® Core™ i7-4710MQ CPU @ 2.50 GHz processor and a Random Access Memory with 16.0 GB. C. Estimation of Real-time Ankle Joint Torque Trajectories Considering that the SmartOs system is not able to measure the human ankle joint torque, it was required to explore a methodology to estimate user-oriented ankle joint torque trajectories (Ԏ meas.h ) in real-time, in a similar way as a torque sensor. In this connection, this Ph.D. thesis proposes the HTE block (Figure 6.4), which represents an EMG-based torque estimation DL regressor. For that, a CNN model was explored considering the following data fusion: gait speed, kinematic (hip kinematics), EMG ( tibialis anterior and gastrocnemius lateralis muscles), anthropometric (body height and mass, ranging from 1.50 to 1.90 m and 50.0 to 90.0 kg, respectively, and shank and foot lengths) and demographic data (age, gender). Further details can be found in the study [65]. The CNN model was trained with a walking dataset collected from seventeen healthy participants (9 females and 8 males) with a mean body height of 168.0 ± 10.31 cm, a mean body mass of 70.11 ± 14.26 kg, and a mean age of 28.05 ± 3.66 years. Participants were instrumented with (i) an 8-channel surface EMG system [Trigno Avanti (Delsys Incorporated, Natick, USA)] to collect EMG data from the tibialis anterior , gastrocnemius lateralis , biceps femoris , and vastus lateralis muscles; (ii) a total of 12 pairs of retro-reflective markers, in which combined with a twelve-camera motion-capture system (Oqus; Qualisys—Motion Capture System, Göteborg, Sweden) and Force Plate-Instrumented Treadmill (Side-by-Side Treadmill – AMTI, MA, USA) were used to acquire the lower limb joint angles and torques (at the ankle, knee, Chapter 6 113 and hip joints). After, each participant was instructed to walk for 4 minutes on the instrumented treadmill, performing 2 minutes at 1.5 km/h and then 2 minutes at 2 km/h. An empirical analysis was conducted to select (i) the kernel size (2, 10, 20, 40, 60); (ii) the number of convolutional layers (1, 2, 3, 4); (iii) the number of filters per convolutional layer (8, 16, 32, 64, 128); (iv) the sequence length (40, 50, 60, 80, 100, 120, 150); (v) the batch size; and (vi) the dropout rate. The normalization method was also studied among max-min, z-score, and robust normalization methods. Furthermore, the ADAM algorithm based on the MSE was used to update the weights and biases of the CNN. This empirical analysis was conducted through the LOSOCV procedure. Of the seventeen participants, one (female 27 years old and a body mass and height of 73.4 kg and 1.63 m, respectively) was randomly selected to test the model. Thus, the model was trained with data from 16 subjects. Three evaluation metrics were employed to evaluate the offline CNN’s performance, namely, the RMSE, the Normalized Mean Square Error (NMSE), and the r . These metrics were computed between the predicted and the real (ground truth) ankle joint torque trajectories during the LOSOCV and the model testing procedures (both performed in a Hewlett-Packard computer with an Intel® Core™ i7-4710MQ CPU @ 2.50 GHz processor and a Random Access Memory with 16.0 GB). Further, an experimental protocol was performed to evaluate the real-time performance, i.e., the prediction time of the CNN into SmartOs CCU. This protocol involved one able-bodied male participant (27 years old) with a body height of 1.70 m and a body mass of 81.2 kg. In the beginning, the participant's gender, age, body mass, height, leg length, and foot length were measured and introduced in the mobile graphical application of the SmartOs system. Then, the participant performed two MVCs for each muscle to normalize the EMG data. After that, the participant was instructed to perform a 5-seconds standing calibration trial for calibrating the InertialLab system. At last, the participant walked on the instrumented treadmill for 5 minutes at 1.5 km/h, while the ankle joint torque trajectories were estimated in real-time. C. Experimental Validation of AAN EMG-based Control To validate the complete AAN EMG-based control proposed in Figure 6.4, five healthy participants (3 males and 2 females with an average age, body height, and body mass of 27.8 ± 3.1, 170.4 ± 8.2, and 67.8 ± 11.4, respectively) without evidence of motor disorders were involved. All participants provided informed consent according to the University of Minho Ethics Committee Chapter 6 114 (CEICVS 006/2020). It is important to note that these participants were not involved in training the regressors described in sub-sections B and C . The protocol started by collecting the participant’s gender, age, body height and mass, and foot and shank lengths. Then, participants were equipped with (i) two EMG sensors placed on the right tibialis anterior and the right gastrocnemius lateralis muscles, following the SENIAM recommendations [105]; and (ii) 2 IMUs from the InertialLab system positioned in the pelvis and in the right thigh, to collect hip joint angles. All these data are used in the HTE block, as depicted in Figure 6.4. Once instrumented with the sensor systems, the participant performed two MVCs per muscle. After that, the participant was instrumented with the SmartOs system (Figure 6.6). Figure 6.6 – Male participant equipped with the InertialLab, Trigno Avanti, and SmartOs systems. Once instrumented, participants underwent a period of familiarization with the device following a trajectory-tracking position control. After familiarization, the data collection started with participants walking in two different sessions. One session consisted of participants being assisted by the AO following a trajectory-tracking position control. The other session consisted of participants being assisted by the AO following an AAN EMG-based control. Both sessions were performed at three speeds (1.2, 1.4, and 1.6 km/h). However, the participants did not know which control strategy was controlling the AO in order to evaluate the users’ perception of which strategy was more or less adapted to their real-time motion. In this connection, a researcher randomly selected the assistive control strategy. Subsequently, participants performed three tasks. In the first, they were asked to remain in a standing position with their right leg raised (the leg instrumented with the SmartOs system), with no movement, for 20 seconds. Secondly, participants were asked to walk for 1 minute on the treadmill, which was initially set at a speed of 1.2 km/h. This task is referred to as the unconditioned task , in which participants were asked to actively perform the walking motion. In the IMU sensors (InertialLab) EMG sensors (Delsys) Ankle Orthosis Backpack Chapter 6 115 third test, the participants were asked to lean on the sidebars of the treadmill, so that they were almost suspended (only their feet touched the treadmill mat, but without supporting their body) for 1 minute. At this point, participants were asked to simulate a pathological condition by trying not to activate the tibialis anterior and gastrocnemius lateralis muscles to perform the gait. In other words, participants had to let the AO do most of the walking motion. Thus, this task is referred to as the conditioned task , in which participants were asked to passively perform the walking motion. Given this condition, it would be expected that there would be a decrease in the activation of the aforementioned muscles and, consequently, the AO would increase its assistance when guided by an AAN EMG-based control. When guided by a trajectory-tracking position control, it would be expected that the assistance provided by the AO during both conditioned and unconditioned tasks would be identical, since this control strategy does not consider the user’s participation. After completing these three tasks, participants performed the same continuous tasks at the same gait speed (1.2 km/h) for the second assistive control strategy. After performing the two sessions (i.e., walking under the trajectory-tracking position control and AAN EMG-based control in random order), participants were given a four-question open questionnaire with the following questions: 1. Among the tasks performed, which condition provided more variable assistance? 2. Among the tasks performed, which condition provided more adequate assistance? 3. Considering the task of walking suspended from the sidebars, in which condition did you feel that the orthosis provided more assistance? 4. Considering the possibility of fatigue, which condition would you choose for prolonged use of the AO? After answering the questions, the protocol was repeated for the remaining speeds, i.e., 1.4 and 1.6 km/h. 6.3.3 RESULTS AND DISCUSSION A. Estimation of Ankle Joint Reference Torque Trajectories In this Ph.D. thesis, it is presented a proof-of-concept of DL regressors’ applicability to achieve an accurate and generalized method for estimating user-oriented ankle torque trajectories for the entire gait cycle, considering gait speeds ranging from 1.0 to 4.0 km/h and for individuals with body Chapter 6 116 height and mass varying from 1.51 to 1.83, and from 52.0 to 83.7 kg, respectively. This enables the estimation of ankle joint reference torque trajectories for a widespread population, including SmartOs’ users. Table 6.3 presents the best results achieved for the DL-based regressor algorithm, for both validation and test conditions, when considering/not considering the EMG signals as input. Pursuing the hypothesis that the ankle joint torque trajectories can be accurately estimated without using EMG signals, a statistical analysis was conducted to evaluate possible significant differences. The Shapiro–Wilk normality test showed that all data were parametric, and the assumptions of homoscedasticity and the existence of outliers were accomplished. Thus, a two-tailed and paired t - test was conducted with a level of confidence of 95%. The results indicate that there were no significant differences between both approaches, with a p -value > 0.05 having been achieved for all metrics ( r : p -value = 0.17, and R2 : p -value = 0.28). Since the p -values were greater than 0.05 for all metrics, it is concluded that there were no significant differences between the two CNN models. This suggests that the CNN model can estimate ankle joint reference torque trajectories accurately even without using EMG signals. The best parameters found for the CNN were the following: (i) kernel size: 2x2; (ii) number of convolutional layers: 2; (iii) number of filters: 8 and 16 in the first and second convolutional layers, respectively; (iv) normalization method: robust; (v) batch size: 20; and (vi) dropout: 25%. Table 6.3 – DL models performance when EMG signals were/were not included as inputs Inputs Deep Learning Model r R2 Prediction Time (ms/sample) LOSOCV Test LOSOCV Test Body height, mass, gait speed, and ankle joint angles CNN 0.89 ± 0.03 0.92 0.91 ± 0.03 0.94 0.51 Body height, mass, gait speed, ankle joint angles, and EMG from tibialis anterior and gastrocnemius lateralis 0.90 ± 0.04 0.89 0.92 ± 0.04 0.92 0.78 Given its prominent capacity to estimate the ankle joint torque trajectories, the predictions performed by CNN were compared to the real ankle joint torque trajectories of the test dataset using the Bland–Altman Plot. The results depicted in Figure 6.7 – a) show that the predictions made by CNN are closer to the real ankle joint torque trajectories since the majority of the measures are Chapter 6 117 within the limits of agreement, the bias is close to 0 N.m, and the limits of agreement are small. Furthermore, Figure 6.7 – b) illustrates the average and standard deviation values of the real and CNN-based predicted ankle joint torque for the test dataset. These results show that the CNN produced reference ankle joint torque trajectories closely similar to the real ones. Figure 6.7 – (a) Bland–Altman plot results; (b) Average (dashed lines) and standard deviation (SD) (filled area) of the real and CNN-based predicted ankle joint torque for the test dataset. The achieved results were encouraging when considering the findings presented in previous studies [152]–[156]. The study [152] reached an R2 of 0.48 for predicting the peak ankle dorsiflexion torque using linear and quadratic equations. The study [153] estimated ankle, knee, and hip angles and torques for the stance phase using feedforward and LSTM neural networks, revealing a r of 0.98. In the study [154], peak ankle, knee, and hip joint angles and torques were estimated using linear, quadratic second-order, and quadratic third-order equations. The highest value of R2 for peak ankle joint torque estimation was 0.93. Other studies [155], [156] have reported that a Multilayer Perceptron can estimate knee joint torque trajectories with an r of 0.97 by combining EMG signals with kinematic sensors. Despite presenting different datasets and study conditions, the CNN tool proposed by this research presents an estimation performance in line with the studies [152]–[156], by predicting ankle joint torque trajectories for the entire gait cycle with an R2 of 0.91 ± 0.03 ( r of 0.95 ± 0.18). The tool developed in this thesis advances by estimating torque for the complete gait cycle, which is advantageous over the estimation for specific gait events [152]–[156] because the torque curve is characterized not only in magnitude but also temporally [157]. Further, the proposed tool predicts ankle joint torque trajectories without the need for complex data acquisitions related to EMG [155], [156]. Chapter 6 124 Table 6.7 – Variation of the EMG signals from tibialis anterior and gastrocnemius lateralis , range of motion (ROM) of the hip joint, and human ankle joint torque between unconditioned and conditioned tasks during AAN EMG-based control strategy (a negative and a positive sign means a reduction and an increase, respectively, when comparing the variable measured at the conditioned task to the unconditioned task) ID EMG from tibialis anterior EMG from gastrocnemius lateralis Hip ROM Human torque 1.6 km/h 1.4 km/h 1.2 km/h 1.6 km/h 1.4 km/h 1.2 km/h 1.6 km/h 1.4 km/h 1.2 km/h 1.6 km/h 1.4 km/h 1.2 km/h 1 -29.7% -32.0% -33.5% -70.5% -64.7% -57.3% -22.7% -24.7% -16,1% -14.3% -15.0% -14.4% 2 -63.1% -64.5% -63.0% -67.1% -62.6% -59.8% -4.9% -18.6% -1,0% -14.0% -11.0% -5.8% 3 -55.7% -63.4% -32.0% -22.2% -9.4% -11.1% -22.0% -17.8% -22,8% -16.3% -22.7% -20.1% 4 -1.8% -2.0% -16.0% -48.9% -18.9% -26.8% -48.8% -44.3% -38,0% -7.0% -9.3% -1.2% 5 -1.7% -9.3% -14.2% -37.3% -64.5% -59.8% -22.8% -21.3% -17,6% -3.4% -6.7% -8.9% Average ± STD/Speed -30.4% ± 25.9% -34.2% ± 26.2% -31.7% ± 17.5% -49.2% ± 18.2% -44.0% ± 24.6% -42.9% ± 20.3% -24.2% ± 14.1% -25.3% ± 9.8% -19,1% ± 11,9% -11.0% ± 4.9% -12.9% ± 5.6% -10.1% ± 6.6% Total average ± STD -32.1% ± 23.2% -45.4% ± 21.0% -22.9% ± 11.9% -11.3% ± 5.7% Chapter 6 125 Table 6.8 – Variation of the dorsiflexion and plantar flexion motor torques between unconditioned and conditioned tasks (a negative and a positive sign means a reduction and an increase, respectively, when comparing the variable measured at the conditioned task to the unconditioned task). Maximum dorsiflexion, and plantar flexion angles measured at the conditioned task ID Dorsiflexion Motor Torque Plantar Flexion Motor Torque Maximum Dorsiflexion Angle (º) Maximum Plantar Flexion Angle (º) 1.6 km/h 1.4 km/h 1.2 km/h 1.6 km/h 1.4 km/h 1.2 km/h 1.6 km/h 1.4 km/h 1.2 km/h 1.6 km/h 1.4 km/h 1.2 km/h 1 18.4% 10.4% 8.1% 11.6% 10.9% 10.5% 11.2 11.8 12.6 -6.3 -8.1 -7.6 2 6.1% 6.9% 6.9% 11.4% 12.3% 17.6% 10.4 11.9 12.9 -7.4 -7.1 -7.4 3 16.9% 1.8% 1.4% 8.7% 1.2% 3.8% 12.7 13.8 14.6 -6.9 -7.9 -7.9 4 12.9% 26.8% 14.5% 11.0% 21.6% 5.9% 11.4 13.8 13.4 -8.8 -7.9 -8.7 5 2.4% 2.8% 12.9% 6.4% 1.5% 10.6% 11.9 13.2 14.7 -7.2 -7.2 -8.6 Average ± STD 11.3% ± 6.2% 9.7% ± 9.1% 8.8% ± 4.7% 9.8% ± 2.0% 9.5% ± 7.6% 9.7% ± 4.8% 11.5 ± 0.7 12.9 ± 0.9 13.6 ± 0.9 -7.3 ± 0.8 -7.7 ± 0.4 -8.0 ± 0.5 Total average ± STD 9.9% ± 6.6% 9.7% ± 4.8% 12.7 ± 0.83 -7.7 ± 0.59 Chapter 6 126 In addition, the contributions of the proposed AAN EMG-based control strategy were compared to the contribution of the trajectory-tracking position control. For that, the EMG signals of the tibialis anterior and gastrocnemius lateralis , the motor torque, and the human ankle joint angle trajectories were analyzed considering the unconditioned and conditioned tasks. Table 6.9 presents the achieved results. Table 6.9 – Variation of the EMG signals from tibialis anterior and gastrocnemius lateralis , and dorsiflexion and plantar flexion motor torques between unconditioned and conditioned tasks (a negative and a positive sign means a reduction and an increase, respectively, when comparing the variable measured at the conditioned task to the unconditioned task). Maximum dorsiflexion, and plantar flexion angles measured at the conditioned task Average ± STD AAN Position EMG from tibialis anterior -32.1% ± 23.2% -29.2% ± 20.9% EMG from gastrocnemius lateralis -45.4% ± 21.0% -42.8% ± 19.4% Dorsiflexion Motor Torque 9.7% ± 6.4% 0.8% ± 2.6% Plantar Flexion Motor Torque 9.7% ± 4.8% -0.9% ± 4.1% Maximum Dorsiflexion Angle (º) 12.7 ± 0.8 10.5 ± 0.1 Maximum Plantar Flexion Angle (º) -7.7 ± 0.6 -5.8 ± 1.0 Table 6.9 shows that, during conditioned tasks of both control strategies (AAN EMGbased and trajectory-tracking position controls), there was a reduction in the muscle activation of tibialis anterior and gastrocnemius lateralis muscles compared to the unconditioned tasks. The decrease in activation of the tibialis anterior muscle during the AAN EMGbased control strategy (32.1% ± 23.2%) was similar to the decrease in activation of the same muscle during the application of the trajectory-tracking position control (29.2% ± 21.0%). An identical result was obtained for the decrease in activation of the gastrocnemius lateralis muscle, with decreases of 45.4% ± 21.0% and 42.8% ± 19.4% during the application of the AAN EMG-based and the trajectorytracking position control strategy, respectively. This means that the quasi-passive walking Chapter 6 127 motion performed by participants at different gait speeds was similar between the two control strategies, as there was a similar reduction in muscle activation in both. Nonetheless, despite a similar reduction in muscle activation, the ability of the AO to assist participants differed between the two control strategies. While during the AAN EMG-based control strategy the AO increased its dorsiflexion motor torque contribution (9.7% ± 6.4%) to compensate for the lower activation of the dorsiflexor muscle ( tibialis anterior ), during the trajectory-tracking position control the increase in the orthosis contribution was very small (0.8% ± 2.6%). The same behavior was observed for the plantar flexion motor torque contribution, in which (i) the AAN EMG-based control increased the motor torque of the AO (9.7% ± 4.8%) to compensate for the lower activation of the plantar flexors ( gastrocnemius lateralis ); (ii) the trajectory control even decreased the motor torque of the AO (-0.9% ± 4.1%) in the presence of lower plantar flexor muscle activity. This result is in line with what would be expected since trajectory-tracking position controls do not take into account the real-time muscular needs of the participant [27], [28]. Thus, these strategies typically do not adapt to the user's participation. Conversely, the proposed AAN EMGbased control strategy precepts the user’s muscular participation and accordingly adapts the AO assistance. Thus, an increase in motor torque is expected as a result of a decrease in the EMG signals of the tibialis anterior and gastrocnemius lateralis . As a result, the ability to reach the maximum dorsiflexion angle (12.0º) was closer with the AAN EMG-based control (12.7 ± 0.8º) than with the trajectory-tracking position control (10.5 ± 0.1º). This difference in joint angles was also seen in the plantar flexion angle, where the AAN EMG-based control achieved a value (-7.7 ± 0.6º) closer to the maximum plantar flexion angle (-11.0º) than the trajectory-tracking position control (- 5.8 ± 1.0º). These results are also in line with what would be expected since the AAN EMG-based control not only invokes the user’s participation but also attempts to correct the user's walking motion and provide assistance when and as much as needed. In addition, when comparing the reference and measured ankle joint angles, the use of the AAN EMG-based control achieved an average RMSE and delay of 4.94º ± 0.24º and 182.0 ± 15.4 ms, respectively. This is reflected in a reduction in RMSE of 8.0% ± 4.2% and a reduction in the delay of 20.7% ± 5.9% compared to the trajectory-tracking position control. At last, the replies of each participant to the applied questionnaire were analyzed. According to the provided replies (depicted in Figure 6.11), (i) 100% of the participants recognized that the AAN EMG-based control provided more variable assistance; (ii) 91.7% of the participants Chapter 6 128 recognized that the AAN EMG-based Control provided more adequate assistance; (iii) 91.7% of the participants felt that the AAN EMG-based Control provided more assistance; and (iv) 100% of the participants preferred the AAN EMG-based Control for the prolonged use of the AO considering the existence of fatigue. These results are promising in the sense that most of the participants who tested the proposed AAN EMG-based control strategy recognized its main advantages, i.e., the ability to provide assistance tailored to the user's needs at the time and in the amount required (when and as much as needed). Figure 6.11 – Representation of the answers to the open questionnaire. 6.4 HUMAN-IN-THE-LOOP CONTROL 6.4.1 CRITICAL ANALYSIS OF RELATED WORK In addition to changes in the physiological, kinematic, and spatiotemporal parameters post-stroke patients typically start to face fatigue three months after stroke [72], [160], [161]. Between 25% and 85% of post-stroke patients report fatigue, i.e., they have an increased need to rest or a lack of energy with high frequency (every day or nearly every day). These fatigue episodes may influence the expected rehabilitation outcomes since they affect the patients’ active participation during rehabilitation sessions. Consequently, the ability to accomplish daily activities is compromised, negatively impacting patients’ quality of life [160]. Chapter 6 129 According to studies [37], [134], patients' active participation should be encouraged to achieve neuromuscular recovery during rehabilitation training. Thus, AAN HITL control strategies have been proposed for wearable assistive devices. With this control scheme, control parameters are adapted based on physiological signals obtained from the user [31], [162], [163]. With the AAN HITL control strategy, it is expected that fatigue can be controlled while still allowing the patient to actively participate in the therapy. In AAN HITL control strategies, the physiological parameter commonly optimized is energy expenditure. However, the standard approach for estimating this signal relies on indirect calorimetry through a respirometer, which requires expensive and non-portable equipment, is time-consuming, produces noisy estimates, and is impractical for real-world applications [164]. At this level, different AI algorithms have been employed to estimate energy expenditure using data collected from wearable sensors (EMG, IMUs, and heart rate sensors) [52], [165]–[168]. owever, to the author’s best knowledge, only the study [168] has integrated the energy expenditure AI algorithm into a knee exoskeleton to perform a HITL control scheme. In HITL control schemes, the reduction of energy expenditure commonly occurs by adjusting the torque profiles of the wearable assistive device [31], [50], [67], [168]–[170]. However, most of the available studies (i) depend on indirect calorimetry to estimate energy expenditure, which limits the wearability of the technology [31], [50], [170]; (ii) take too long to find the optimal control parameters, typically ranging from 24 to 72 minutes [31], [50], [67], [169]; and (iii) use wearable assistive devices with an electric cable-driven transmission [31], [50], [67], [169], [170]. The fact that wearable assistive devices use an electric cable-driven transmission means that an ankle torque pattern similar to the natural human torque profile can be used. However, according to Chapter 2, electric motor-based controllers with a gear-based transmission correspond to the most used actuators employed in wearable assistive devices for post-stroke rehabilitation. As already exposed for the knee joint in the study [168], for electric motor-based controllers with a gear-based transmission, the ankle torque pattern to be interpreted by the wearable assistive device must be different from the natural human torque pattern. Otherwise, the gait pattern will not be met because the instants of dorsiflexion and plantar flexion that should be performed will not be correctly executed by the wearable assistive device. Therefore, when designing a trajectory-tracking torque controller for an electric motor-based controller with a gear-based transmission, the torque profile to be used as the reference trajectory must be manually adapted to the intended application. Chapter 6 130 The AAN HITL control strategy proposed here is designed and integrated into the SmartOs system to address the limitations of existing strategies in the literature. The proposed control employs a regressor proposed in the study [168] to estimate energy expenditure based on data from four IMUs (eliminating the need for a respirometer). The innovation stems from (i) the development of a HITL control that allows the ankle torque trajectory to be optimized according to the user’s energy expenditure; and (ii) the proposal and implementation of a torque control for an electric motor-based controller with a gear-based transmission. 6.4.2 METHODOLOGY A. Hierarchical Control Architecture The AAN HITL control proposed in this research follows a hierarchical architecture organized into high-, mid-, and low-levels, as depicted in Figure 6.12. Figure 6.12 – AAN HITL control. BMI : body mass index defined through the mobile graphical application of the SmartOs system. Inertial Data : 3D acceleration data from the chest, right wrist, right waist, and left ankle recorded with the InertialLab system. EE : energy expenditure estimated every 10 seconds. IT and TI : cumulative interaction torque, and the integral of the reference ankle torque, respectively. CP : control parameters (magnitude of the plantar flexion and dorsiflexion torques at the push-off instant and midswing event, respectively). Ԏ ref.h and Ԏ meas.o : human ankle joint reference torque and the measure SmartOs’ motor torque, respectively. e Ԏ: torque error. u : PID command. SS : speed scaling block (Equation 6.1). PID : Proportional-Integral-Derivative controller. The high-level controller is responsible for (i) receiving and processing the 3D acceleration data from 4 IMUs (InertialLab system) positioned at the chest, right wrist, right waist, and left ankle, and the body mass index ( BMI ); and (ii) estimating the user’s energy expenditure ( EE ) every 10 seconds. Details about data processing steps and energy expenditure estimation can be found in B. Energy Expenditure Estimation . Energy Expenditure Estimation Cubic Spline Interpolator SS Orthosis uman PID u meas o e ref o Mid level ( )Low level ( k ) igh level ( ) ref h CMA ES C nertial ata M Chapter 6 131 Once estimated, the energy expenditure value is sent to the mid-level, where it is used together with the cumulative interaction torque ( IT ) and the integral of the reference ankle torque ( TI ), as inputs to the HITL optimizer ( CMA-ES ). These two variables ( IT and TI ) are computed for each gait cycle and are reset at the beginning of the gait cycle. The CMA-ES block presents a cost function (detailed in D. CMA-ES optimizer ) that tends to be minimized through the variation of two control parameters ( CP, representing the CMA-ES outputs), namely, the magnitude of the plantar flexion and dorsiflexion torques at the push-off instant and mid-swing event, respectively. These control parameters act as input in a cubic spline interpolator (detailed in C. Ankle Torque Profile ), generating the human ankle joint reference torque (Ԏ ref.h ). The Ԏ ref.h is interpolated to the Ԏ ref.o , through the Speed Scaling ( SS ) block, according to Equation 6.1. The Ԏ ref.o is then sent to the low-level and it is compared with the SmartOs’ motor torque (Ԏ meas.o ), generating a torque error ( e Ԏ). The e Ԏ feeds the PID controller, using proportional, integral, and derivative gains of 300, 7.5, and 2.5, respectively. Considering the e Ԏ, the PID controller computes a PID command ( u ) that is limited to maximum and minimum values of 2500 and -2500, through a saturator. This command is interpreted by the ankle orthosis, generating the corresponding motor torque. With this control architecture, it is expected the adaption of the control parameters to reduce the user's energy expenditure while still providing ankle support. B. Energy Expenditure Estimation In this Ph.D. thesis, the energy expenditure was estimated in real-time using a machine learning regression model, namely an Exponential Gaussian Process Regression (EGPR), presented in the study [171]. This regression model was chosen since it outperformed other AI models, including boosted decision trees, bagged decision trees, support vector machines, and CNNs [171]. The EGPR model was trained and validated offline using data from a publicly available 10-participant dataset [172]. The training involved LOSOCV with a dataset including acceleration measurements from the chest, right wrist, left waist, and right ankle, as well as the participants' body mass index. The trained regression model was implemented into the SmartOs high-level to estimate the energy expenditure in real-time, every 10 seconds. This time window was chosen to ensure that at least two respiratory cycles were captured, which is considered sufficient for accurate instantaneous metabolic cost estimation [171]. Moreover, despite training the EGPR with participants' body mass index and 3D acceleration data from the chest, right wrist, left waist, and right ankle, real-time tests Chapter 6 132 were conducted with the following data: participants' body mass index and acceleration data from the chest, right wrist, right waist, and left ankle. This alteration was done since (i) the SmartOs system does allow an IMU sensor to be positioned at the right ankle joint; and (ii) no differences were found in the energy expenditure estimation when using the IMU sensors at the right/left waist and ankle. To estimate energy expenditure, the following data processing was applied. After receiving the 3D acceleration signals from the four specified body locations, these signals were filtered with a 4thorder Butterworth low-pass filter at 20 Hz. This frequency was chosen according to studies [165]– [167], since it provides a balance between signal attenuation and preservation of relevant information. Then, the filtered data were reorganized into 10-second windows, as suggested in the study [171]. At last, the mean absolute deviation was computed for each signal since it corresponds to a good feature to distinguish different physical activity intensities ([173]), being then normalized by the z-score method. Once processed, the data were interpreted by the EGPR, and the energy expenditure was estimated. C. Ankle Torque Profile The ankle torque profile represents the variation of the torque magnitude and sign in time, during a gait cycle. Figure 6.13 depicts the natural human ankle joint angle and torque, during levelground walking at several gait speeds ranging from 1.0 to 4.0 km/h. Figure 6.13 shows that if the natural human ankle torque profile is used as a reference in a trajectory-tracking torque control applied to electric motor-based actuators with a gear-based transmission, the device will perform (i) small dorsiflexion immediately after the heel-strike event; (ii) a continuous plantar flexion until the end of the stance phase; and (iii) a torque around 0 N.m throughout the swing phase. In this way, the kinematic movement of the ankle joint would not be similar to the trajectory typically followed by this joint during the gait cycle. In addition, the area of the curve corresponding to the dorsiflexion motion must be identical to the area of the curve corresponding to the plantar flexion motion. Otherwise, the position of the ankle at the heel-strike event will be different for each gait cycle. It is therefore necessary to create a torque trajectory specific for electric motor-based actuators with a gear-based transmission. Chapter 6 133 Figure 6.13 – Human ankle torque (top view) and ankle angle (bottom view). Adapted from the study [121]. In an attempt to find the best torque profile to use, an empirical test was performed with a healthy participant walking for one minute on a treadmill, assisted by the SmartOs system operating with trajectory-tracking position control, at a gait speed of 1.5 km/h. The average motor torque curve generated by the SmartOs system was then recorded for each gait cycle. Figure 6.14 depicts the registered motor torque profile. According to Figure 6.14, the torque pattern was similar to the ankle angle pattern shown in Figure 6.13 (bottom view), as the dorsiflexion and plantar flexion phases coincided in time. On the other hand, the sum of the dorsiflexion and plantarflexion areas at the end of each gait cycle was approximately 0 N.m, indicating that the ankle position was approximately the same at the beginning of each gait cycle. Thus, the torque profile shown in Figure 6.14 seems to be a good candidate to be used as a reference in trajectory-tracking torque controls. . . . nkle angle ait Cycle Average trial ( . km h) Average trial ( . km h) Average trial ( . km h) Average trial ( . km h) Average trial ( . km h) Average trial ( . km h) Average trial ( . km h) Collected trials Plantar Flexion Dorsiflexion Plantar Flexion Dorsiflexion nkle torque m kg