Repositorio Institucional de Documentos

Trabajo Fin de M´aster Arrhythmic Risk Prediction based on the Analysis of Ventricular Repolarization Markers from Surface ECG Autora Julia Ram´ırez Garc´ıa Directores Esther Pueyo Paules Pablo Laguna Lasaosa Escuela de Ingenier´ıa y Arquitectura Instituto de Investigaci´on en Ingenier´ıa de Arag´on Junio 2013 Acknowledgments This master thesis has been supported by grant PIFUZ-2011-B-TEC-001, from biosig division, Communication Technology Group, University of Zaragoza. First of all, I profoundly want to thank my supervisors, Esther and Pablo. Their encouragement has been vital. Every day I learn new things from them, but this period has been marked with patience and wishful work. Secondly, I would like to thank Pr. Jacobs for accepting me so warmly and giving me the chance to widen my field knowledge. I am very grateful to Ana and Violeta because without their previous work and their present help I would not have been able to finish this master thesis. Special mention to my colleagues and friends: David, Jes´us, ´ Alex, Carlos, Alba and Dani, because the results in this master thesis show that laughing improves motivation. Above all, to my best support, Juan. Finally, many thanks go to my family, who always look after me and of which I am so proud to be part of. Arrythmic Risk Prediction based on the Analysis of Ventricular Repolarization Markers from Surface ECG ABSTRACT Heart rate (HR) dependence of action potential duration (APD), also called restitution kinetics, is critical in activation instability of the heart and provides relevant information for ventricular arrhythmic risk stratification. The dynamic APD restitution (APDR) curve quantifies the relationship between the APD and the RR interval (inverse of HR) at steady-state when pacing at different RR values. Heterogeneities in the ventricle lead to non uniform restitution properties, which makes APDR curves present spatial variations. Dispersion is a measure of that spatial variation. An electrocardiogram (ECG) index, Δα, that quantifies dispersion in the dynamic APDR slopes by characterizing the relationship between the T-peak-to-T-end (Tpe) and the RR intervals at different steady-state conditions, was recently proposed. In this master’s thesis a fully automated method to compute Δαin ambulatory recordings has been developed and the value of Δα, as an independent predictor of sudden cardiac death (SCD) in patients with chronic heart failure (CHF), has been evaluated. Consecutive patients with symptomatic CHF were enrolled in the “MUSIC” (MUerte S´ubita en Insuficiencia Cardiaca) study, a prospective, multicenter study designed to assess risk predictors for cardiovascular mortality in ambulatory patients with CHF. The Holter recordings of 609 patients (48 victims of SCD, 64 of other cardiac causes, 25 of non-cardiac death causes and 472 survivors) with sinus rhythm were available for the present study. Preprocessing of the ECG signals performed in this master’s thesis included low pass filtering at 40 Hz to remove electric and muscle noise, cubic splines interpolation for baseline wander removal and ectopic beats detection. A single-lead-and-rules delineation technique was applied to select the samples from the T-wave and compute principal component analysis. Then, the first principal component was delineated using a single-lead technique and, from the delineation marks, the RR and Tpe interval series of the ECG were obtained and subsequently interpolated at a sampling frequency fs= 1Hz. Since each value of the APDR curve represents a stationary state corresponding to a specific HR value, the ECG index Δαproposed to estimate APDR dispersion should in principle be computed using ECG segments of stable HR regimes. Those types of segments are difficult to get in clinical practice and thus the dependence of the Tpe interval on a history of previous RR intervals was modeled and compensated for the Tpe memory lag. The relationship between Tpe and RR was then characterized on the whole ECG of the ambulatory recordings and the index Δαwas calculated. A threshold set in Δα>0.046 showed to discriminate patients from high and low SCD risk (p-value = 0.003). The time to recurrence (SCD) was approximately doubled among patients with Δα≤0.046 in comparison with those with Δα>0.046 (p-value = 0.001). By combining Δαwith other ECG indices, like the index of average T-wave alternans (IAA), stratification of SCD risk was improved (p-value <0.001). This study demonstrates that dispersion in APDR, quantified from Holter ECG recordings, is a strong and independent predictor of SCD in patients with CHF. This findings support the hypothesis that an increased dispersion in APDR reflects abnormal cardiac function predisposing to SCD. IV V Arrythmic Risk Prediction based on the Analysis of Ventricular Repolarization Markers from Surface ECG RESUMEN La dependencia de la duraci´on del potencial de acci´on (APD, del ingl´es “Action Potential Duration”) con el ritmo cardiaco (HR, del ingl´es “Heart Rate”), tambi´en conocida como cin´etica de restituci´on, es cr´ıtica a la hora de generar inestabilidades el´ectricas en el coraz´on y proporciona informaci´on relevante en la estratificaci´on del riesgo a sufrir arritmias ventriculares. La curva din´amica de restituci´on del APD (APDR, del ingl´es “APC restitution”) cuantifica la relaci´on entre el APD y el intervalo RR (inverso de HR) en condiciones estacionarias. Heterogeneidades en el ventr´ıculo dan lugar a propiedades de la restituci´on no uniformes, haciendo que las curvas APDR presenten variaciones espaciales. La dispersi´on es una medida de dicha variaci´on espacial. Recientemente se propuso en la literatura un ´ındice derivado del electrocardiograma (ECG), Δα, que cuantifica la dispersi´on en las pendientes de las curvas din´amicas de APDR mediante la caracterizaci´on de la relaci´on entre los intervalos del pico al final de la onda T (Tpe) y RR bajo condiciones estacionarias diferentes. En este Trabajo Fin de M´aster (TFM) se ha desarrollado un m´etodo autom´atico para obtener y evaluar, a partir de registros ambulatorios, Δα, como predictor independiente de muerte s´ubita cardiaca (SCD, del ingl´es“Sudden Cardiac Death”) en pacientes con fallo cardiaco cr´onico (CHF, del ingl´es “Chronic Heart Failure”). Pacientes con CHF sintom´atico formaron parte del estudio “MUSIC” (MUerte S´ubita en Insuficiencia Cardiaca). La base de datos conten´ıa los registros Holter the 609 pacientes (48 v´ıctimas de SCD, 64 de otras causas cardiacas, 25 de causas no cardiacas y 472 supervivientes) con ritmo sinusal. El preprocesado de las se˜nales ECG realizado en este TFM consisti´o en un filtrado paso bajo a 40 Hz, interpolaci´on de splines c´ubicos y un detector de latidos ect´opicos. Se aplic´o una t´ecnica de delineaci´on “uniderivacional m´as reglas a posteriori” para seleccionar las muestras pertenecientes a la onda T y realizar un an´alisis de componentes principales. A continuaci´on, se deline´o la primera componente principal mediante una t´ecnica uniderivacional y, a partir de las marcas de delineaci´on, se obtuvieron las series de los intervalos RR y Tpe. Posteriormente, se interpolaron a una frecuencia de muestreo fs= 1 Hz. Como cada valor de la curva APDR est´a medido a un valor espec´ıfico de RR, el ´ındice de ECG Δαdeber´ıa calcularse usando segmentos de ECG de ritmos cardiacos estables. Dichos segmentos son dif´ıciles de conseguir en la pr´actica cl´ınica y por lo tanto se model´o la dependencia del intervalo Tpe con una historia de intervalos previos de RR y se compens´o por el retardo de memoria de Tpe. La relaci´on entre Tpe y RR se caracteriz´o en los registros completos de ECG. Un umbral fijado en Δα>0.046 discrimin´o los pacientes en alto y bajo riesgo a sufrir SCD (p-valor = 0.003). El tiempo hasta el evento (SCD) fue aproximadamente el doble en los pacientes con Δα≤0.046 en comparaci´on con los Δα>0.046 (p-valor = 0.001). Al combinar Δαcon el ´ındice de media de alternancias de onda T se mejor´o la estratificaci´on del riesgo a sufrir SCD (p-valor<0.001). Este estudio demuestra que la dispersi´on en APDR, cuantificada a partir de registros ECG Holter, es un predictor de SCD fuerte e independiente en pacientes con CHF. Estos resultados apoyan la hip´otesis de que una dispersi´on de APDR elevada refleja un funcionamiento cardiaco anormal, con predisposici´on a sufrir SCD. Contents 1. Introduction 3 1.1. Background ....................................... 3 1.2. The Electrocardiogram ................................. 3 1.2.1. Electrical Activity of the Heart ........................ 4 1.2.2. Leads ...................................... 5 1.2.3. ECG Waves and Time Intervals ........................ 6 1.3. Electrocardiogram Detection and Delineation .................... 7 1.3.1. Single-lead Delineation ............................. 7 1.3.2. Single-lead-and-rules Delineation ....................... 7 1.4. Action Potential Duration Restitution Dispersion .................. 7 1.5. Objectives ........................................ 8 1.6. Structure of the document ............................... 9 2. Materials and Methods 11 2.1. Materials ........................................ 11 2.1.1. Study Population ................................ 11 2.1.2. Follow-up and End Points ........................... 11 2.2. Methods ......................................... 12 2.2.1. ECG Preprocessing and Delineation ..................... 12 2.2.2. Restitution Dispersion from ECG segments with Stable Heart Rate . . . . 12 2.2.3. Restitution Dispersion from ECG segments with Unstable Heart Rate . . 13 2.2.4. Automated processing of the database .................... 15 2.3. Statistical Analysis ................................... 15 2.3.1. Quantitative statistical differences between populations .......... 16 2.3.2. Classification .................................. 16 2.3.3. Survival analysis ................................ 16 3. Results 19 3.1. Analysis of the relation between Δαand SCD .................... 19 3.2. Selection of a risk threshold .............................. 20 3.3. Analysis of the relation between increments in Δαand SCD ............ 20 3.4. Survival analysis .................................... 21 3.4.1. Influence of Δαand other variables on SCD ................. 21 3.4.2. Cumulative survival curves .......................... 23 3.5. Limitations ....................................... 25 vii 2LIST OF TABLES Chapter 1 Introduction 1.1. Background Cardiovascular disease is the leading cause of death in developed countries. In 2010, 201,781 deaths associated with cardiovascular disease occurred in Spain, 92,224 corresponding to men and 109,557 to women. This number represented 54,4 % of the total number of deaths that occurred in Spain that year [1]. A good number of those cardiovascular deaths were produced by arrhythmias. An arrhythmia is defined as a deviation from or a disturbance of the normal sinus rhythm. The normal sinus rhythm has a rate between 50 and 100 beats/minute at rest. The rhythm is called sinus bradycardia when the rate is below the lower limit and sinus tachycardia when it is above the upper limit. A particular case of arrhythmia is ventricular fibrillation, which is a totally disorganized rhythm during which the ventricles cease to depolarize in an orderly fashion. As a result, a heart undergoing ventricular fibrillation cannot deliver oxygenated blood to the brain. Ventricular fibrillation leads to cardiac arrest, cessation of respiration, loss of consciousness, and, if no immediate treatment is given, it could lead to Sudden Cardiac Death (SCD) and be almost invariably fatal [2]. SCD remains an important cause of mortality in patients with mild-to-moderate heart failure (New York Heart Association [NYHA] classes II and III) [3]. Although previous studies have shown the benefit of implantable cardioverter-defibrillators in this type of population [4], the cost effectiveness of the therapy is low, as only a minority of patients with implantable cardioverter-defibrillators benefitted from this therapy during the follow-up period [5]. A number of indices have been proposed as SCD predictors, including left ventricular ejection fraction (currently the only recommended marker to risk stratify patients [6]) and T-wave alternans [7]. Nevertheless, further research is needed to provide an index or a combination of indices with improved capacity to identify patients at risk of SCD. 1.2. The Electrocardiogram An electrocardiogram (ECG) describes the electrical activity of the heart recorded by electrodes placed on the body surface. The voltage variations measured by the electrodes are caused by the action potentials of the excitable cardiac cells as they make the cells contract. The resulting heartbeat in the ECG is manifested by a series of waves whose morphology and timing 3 4CHAPTER 1. INTRODUCTION convey information that can be used for diagnosing diseases associated with disturbances of the heart’s electrical activity [8]. 1.2.1. Electrical Activity of the Heart The heart is a muscular organ the size of a large fist whose primary function is to pump oxygen-rich blood throughout the body. Its anatomy is divided into two“mirrored”sides, left and right, which support different circulatory systems but which pump in a synchronized, rhythmic manner. Each side of the heart consists of two chambers, the atrium where the blood enters and the ventricle where the blood is forced into further circulation. The wall of the heart is called the myocardium and is primarily composed of muscle cells which produce mechanical force during contraction of the heart. The myocardium also contains specialized muscle cells which are connected into a network (conduction system) that allows an electrical impulse to rapidly spread throughout the heart. A cardiac cycle is created when such an impulse propagates through the conduction system. The electrical impulse is the event that triggers the mechanical force, and thus the electrical event precedes heart contraction. The initialization of this cardiac cycle occurs in a mass of pacemaker cells with the ability to spontaneously fire an electrical impulse. These cells are collectively referred as the sinoatrial (SA) node and are situated in the upper part of the right atrium. Each cardiac cycle is composed of two phases, activation and recovery, which are referred to in electrical terms as depolarization and repolarization and in mechanical terms as contraction and relaxation. Depolarization is manifested by a rapid change in the membrane potential of the cell and constitutes the initial phase of the cardiac action potential (AP). The rapid change in voltage causes neighbouring cells to depolarize, and, as a result, an electrical impulse spreads from cell to cell throughout the myocardium. Depolarization is immediately followed by repolarization during which the membrane potential of the cells gradually returns to its resting state. The ECG describes the different electrical phases of a cardiac cycle and represents a summation in time and space of the action potentials generated by millions of cardiac cells. Of the millions of individual cells in the heart that depolarize during a cardiac cycle, only groups of cells in the myocardium depolarize at any given instant. Each group of cells simultaneously depolarizing may be represented as an equivalent current dipole source to which a vector is associated, describing the dipole’s time-varying position, orientation and magnitude. The related vectors of all these groups can be summed to give a “dominant” vector which describes the main direction of the electrical impulse. Figure 1.1 illustrates how the AP of different cardiac cells generate the ECG signal, in this example viewed by an exploring electrode which is positioned on the chest. During atrial depolarization, the dominant vector is directed downwards towards the AV node. As a result, an atrial wave with positive polarity (“P” wave) is generated in the ECG recorded at the position of the exploring electrode. The amplitude of the resulting wave is low because the muscle mass of the atria that produces the electrical wavefront is relatively small. Ventricular depolarization begins in the wall between the ventricles (septum) in such a way that the associated vector is directed away from the exploring electrode; hence, the related ECG wave (“Q”wave) has negative polarity. Due to the high conduction velocity of the cells in this part of the heart, the Q wave has short duration. During continued ventricular depolarization, the dominant direction of the vector gradually turns toward the exploring electrode (“R” wave). Depolarization terminates with the dominant vector pointing away from the electrode, and thus a wave with negative polarity (“S” 1.2 The Electrocardiogram 5 wave) is produced in the ECG. During ventricular repolarization, a similar sequence of dominant vectors to those during ventricular depolarization appears, and a wave with positive polarity (“T” wave) is produced. Since atrial repolarization is concurrent with ventricular depolarization, the related atrial repolarization wave is masked by the ventricular waves which have much larger amplitudes. Figure 1.1: The morphology and timing of action potentials from different regions of the heart and the related cardiac cycle of the ECG as measured on the body surface. 1.2.2. Leads The electrical activity of the heart is measured on the body surface by attaching a set of electrodes to the skin. The electrodes are positioned so that the spatiotemporal variations of the cardiac electrical field are sufficiently well-reflected. For an ECG recording, the difference in voltage between a pair of electrodes is referred to as a lead. A number of lead systems exist today with standardized electrode positions. The two that have received the most attention, namely, the standard 12-lead ECG and the orthogonal lead system producing a vectorcardiogram (VCG) are described below. Standard 12-lead ECG: The standard 12-lead ECG is the most widely used lead system in clinical routine and is defined by a combination of three different lead configurations: the bipolar limb leads, the augmented unipolar limb leads and the unipolar precordial leads. The 12-lead ECG is recorded by placing 10 electrodes at standardized positions on the body surface. Orthogonal leads: Each standard lead reflects a different point of view of the electrical activity of the heart, depending on the place where the electrode, or pairs of electrodes are located. Sometimes it is interesting to obtain a global ECG which represents the different ECGs obtained from individual leads. Therefore, is is possible to compute an orthogonal 6CHAPTER 1. INTRODUCTION base composed of three leads from the standard 12-lead ECG [9]. This method symbolizes three leads (Frank leads) that would be obtained if three electrodes were placed on the x, yand zaxis of the heart, which is physically unviable [9]. The three orthogonal leads were available in the database analyzed in this master’s thesis. 1.2.3. ECG Waves and Time Intervals To develop signal processing algorithms, the knowledge of ECG wave characteristics, which are described below, along with the wave-naming convention, is central. P wave: Reflects the sequential depolarization of the right and left atria. QRS complex: Reflects depolarization of the right and left ventricles. It is composed of the Q,Rand Swaves. Since the QRS complex has the largest amplitude of the ECG waveforms, it is the waveform of the ECG which is first identified in any type of computerbased analysis. T wave: Reflects ventricular repolarization. Substantial changes in T-wave have shown to provide relevant information about the risk of suffering from ventricular arrhythmias. This master’s thesis mainly analyzes this wave. Figure 1.2 shows ECG waves and important time intervals. T pe (n-1) interval QT(n) interval 19.6 19.8 20 20.2 20.4 20.6 20.8 21 0 0.5 1 1.5 2 ECG (mV) Time (s) P Q Rn-1 S T Rn QTp(n) interval RR(n) interval n-1th beat nth beat Figure 1.2: Wave definitions of the cardiac cycle and important wave durations and intervals. RR interval: Represents the length of a ventricular cardiac cycle, measured between two successive R waves, and serves as an indicator of ventricular rate. QT interval: Represents the time from the onset of ventricular depolarization to the completion of ventricular repolarization. Tpe interval: Represents the time from the peak to the end of T-wave. It is considered as a ventricular repolarization dispersion marker. 1.3 Electrocardiogram Detection and Delineation 7 1.3. Electrocardiogram Detection and Delineation The detection process consists of detecting heartbeats. ECG delineation consists of determining the boundaries of each wave within the PQRST complex so that, with the resulting time instants, wave durations can be computed. 1.3.1. Single-lead Delineation Single-lead delineation identifies the wave boundaries of a lead, independently from the others. In this work an automatic method based on the Wavelet Transform (WT) has been used [10]. WT describes the signal in both time and frequency domains. Therefore, it allows to represent temporal wave characteristics at different levels (scales) depending on its frequency content. This representation is proportional to the signal derivative, so a zero-crossing represents a peak in the original signal. QRS complex needs a different scale of that used to characterize P and T waves, because its frequency content is substantially different [11]. Firstly, QRS complex fiducial point (QRS complex gravity center) is detected and, then, Q, R and S waves are separately delineated. Finally, P and T waves are delineated by sliding the analysis window [10]. Signal to noise ratio constitutes a drawback in single-lead delineation technique because if the noise level is high it is difficult to correctly place the wave boundaries. 1.3.2. Single-lead-and-rules Delineation This method consists of selecting an annotation mark among the marks obtained using single-lead delineation over the available leads in each heartbeat. If the mark is the onset of a wave, all the onset marks from all the leads are sorted and the first one (which marks the position of the first recorded electrical change) is chosen, since it is the most restrictive. To do so, a protection criteria which states that kleads must have their onset mark within a δtime interval needs to be accomplished. kand δvalues are chosen depending on the delineated wave. If the mark is the offset of a wave, the algorithm is the same but choosing the last annotation mark. If the protection criteria is not accomplished, the mark is rejected. If the mark corresponds to the wave peak, the median criteria is used: all the marks are sorted and the one in the middle is selected [12]. 1.4. Action Potential Duration Restitution Dispersion The underlying mechanisms of lethal arrhythmias, which contribute to cardiovascular disease begin the primary cause of death in the industrialized world, are poorly understood [13]. The generation of arrhythmias has been widely studied by dynamically pacing cardiac myocytes, cardiac tissue or whole hearts [14]. These experimental results have revealed that heart rate (HR) dependence of action potential duration (APD), also called restitution kinetics, is critical in activation instability and, therefore, provides relevant information for ventricular arrhythmic risk stratification [15]. The dynamic APD restitution (APDR) curve, measured using the so-called dynamic restitution protocol, quantifies the relationship between the APD and the RR interval (inverse of HR) at steady-state when pacing at different RR values [16] (Figure 1.3). 8CHAPTER 1. INTRODUCTION Figure 1.3: APD dynamic restitution (APDR) curve from one particular cardiac cell. Individual APDR curves have been reported to play an important role in the development of ventricular arrhythmias and APDR curves containing steep slopes are thought to induce alternans of APD and block propagation leading eventually to arrhythmia [17]. On the contrary, for shallow slopes, APD disturbances are smaller and eventually return to a stable activation [16]. Heterogeneities in the ventricle lead to non uniform restitution properties, which makes APDR curves present spatial variations [18]. Dispersion is a measure of that spatial variation. Recent studies have suggested that dispersion in the APDR curves may act as a potent arrhythmogenic substrate [19]. Additionally, increments in that dispersion have been associated with greater propensity to suffer from ventricular tachycardia/fibrillation [20] (Figure 1.4). In [21], a method to indirectly estimate dispersion of restitution slopes by making only use of the surface ECG was developed. An ECG index, Δα, that quantifies dispersion in the dynamic APDR slopes by characterizing the relationship between the T-peak-to-T-end (Tpe) and the RR intervals at different steady-state conditions, was proposed and evaluated. 1.5. Objectives The main objective in this work is to present a fully automated method to analyze APDR dispersion in ambulatory recordings and to show that the ECG index, Δα, proposed in [21], is an independent predictor of SCD in patients with chronic heart failure (CHF). To achieve the proposed objective, a number of steps were followed: To process the “MUSIC” (MUerte S´ubita en Insuficiencia Cardiaca) database, composed of 609 Holter ECG recordings. To compute a representative marker of dispersion of APD restitution curves from changes in the heart rate and time intervals measured on the ECG signal. To compare the obtained responses and to extract conclusions about the relevance of these results to predict ventricular arrhythmias that could lead to SCD in patients with CHF. 1.6 Structure of the document 9 Figure 1.4: Dynamic restitution curves (APDR) for two regions corresponding to APDmin (dashed line) and AP Dlast (solid line). Slopes αmin and αlast are estimated for a change in the RR interval. 1.6. Structure of the document After this brief introduction, the structure of this document is the following: in chapter 2, the analyzed database and the methods used to reliably compute the risk index, Δα, are described. In chapter 3, the results obtained by applying the methodology on the studied database are presented and discussed. In chapter 4, the practical training performed during this work is explained. Finally, in chapter 5, the conclusions and future work are presented. 10 CHAPTER 1. INTRODUCTION Chapter 2 Materials and Methods 2.1. Materials 2.1.1. Study Population Consecutive patients with symptomatic CHF corresponding to NYHA classes II and III were enrolled in the MUSIC (MUerte S´ubita en Insuficiencia Cardiaca) study, a prospective, multicenter study designed to assess risk predictors for cardiovascular mortality in ambulatory patients with CHF [22]. The study protocol was approved by institutional investigation committees and all patients signed informed consent. The Holter recordings of 609 patients (48 victims of SCD, 64 of other cardiac causes, 25 of non-cardiac death causes and 472 survivors) with sinus rhythm were available for the present study. No medications were withdrawn during Holter monitoring. Each recording consisted of three orthogonal ECG leads, sampled at 200 Hz. The clinical characteristics of the studied patients and medications are listed in Table 3.1. Measurements of the index of average alternans (IAA), computed in the MUSIC database after evaluation of other clinical variables, were available for this study [7]. Only the values of IAA corresponding to the patients analyzed in this master’s thesis were considered. 2.1.2. Follow-up and End Points Patients were followed up every 6 months for a median of 48 months, with total mortality as a primary end point and CD and SCD as secondary end points. Information about end points was obtained from medical records, patients’ physicians and family members. Sudden cardiac death was defined as (1) a witnessed death occurring within 60 minutes from the onset of new symptoms unless a cause other than cardiac failure was obvious, (2) an unwitnessed death (<24 hours) in the absence of preexisting progressive circulatory failure or other causes of death, or (3) death during attempted resuscitation. Cardiac death was defined as death from cardiac causes, including SCD, but excluding such vascular causes as pulmonary embolism, aortic aneurysm dissection/aneurysm or stroke. End points were reviewed and classified by the MUSIC Study Endpoint Committee. Table 3.2 summarizes the number of deaths in the study population during the median 48-month period. 11 18 CHAPTER 2. MATERIALS AND METHODS Logrank test In statistics, the logrank test is a hypothesis test to compare the survival distributions of two samples. It is a nonparametric test and appropriate to use when the data are right skewed and censored. It is widely used in clinical trials to establish the efficacy of a new treatment compared to a control treatment when the measurement is the time to event. The logrank test statistic compares estimates of the hazard function of the two groups at each observed event time. It is constructed by computing the observed and expected number of events in one of the groups at each observed event time and then adding these to obtain an overall summary across all time points where there is an event [33]. Chapter 3 Results This chapter presents and discusses the results obtained in this master’s thesis regarding the computation of APDR dispersion in ambulatory recordings from CHF patients and its use for prediction of SCD risk. The relevance of this study is supported by the large amount of clinical studies investigating abnormal ventricular behaviour and its relation to SCD. In this study, Principal Component Analysis transformation was applied over the three orthogonal leads available in the database and the heartbeats from the first principal component were detected and delineated. From the annotation marks provided by the delineation process, RR and Tpe series were computed. The value of Δα, describing dispersion in APDR slopes, was obtained from the RR and Tpe series of each patient in the database. To account for the Tpe memory lag with respect to the RR interval, a model that characterizes the Tpe dependence on RR was used. The results obtained for Δαand their relation to the occurrence of clinical events are described in the following. 3.1. Analysis of the relation between Δαand SCD The risk marker Δαdid not follow a normal statistical distribution. Therefore, the nonparametric Mann-Whitney U test (section 2.3.1) was applied to determine whether SCD victims (group a1) presented Δαvalues that were significantly different from the rest of patients (cardiac death but non-SCD victims, victims of other causes and survivors) (group a2). The 25th, 50th and 75th percentiles of Δαin the study population were 0.005, 0.022 and 0.046, respectively. The median value of Δαwas 0.033 for group a1 and 0.022 for group a2 (p= 0.048) (Figure 3.1(a)). In an equivalent way, the procedure was repeated to determine whether there were statistical differences between CD victims (group b1) and the rest (victims of other causes and survivors) (group b2) Δαvalues. The median value of Δαwas 0.021 for group b1 and 0.022 for group b2 (p= 0.608) (Figure 3.1(b)). Statistically significant differences were found in Δαmedian values between groups a1 and a2 but not in Δαmedian values between b1 and b2 groups. As a result, the idea of the restitution dispersion marker acting as a good SCD predictor may be supported. Increments in Δαwould indicate higher SCD risk. 19 20 CHAPTER 3. RESULTS (a) (b) Figure 3.1: Δαmedian ±95 % confidence interval for SCD and non-SCD victims (a) and for CD and non-CD victims (b). 3.2. Selection of a risk threshold Based on the results in section 3.1 showing that increments in Δαvalues are related with SCD, a risk threshold was set on the continuous variable Δα. A new binary variable was built so that Δα+ (associated with Δαvalues above the threshold) was composed of patients at higher SCD risk and Δα- (associated with Δαvalues below the threshold) contained patients at lower SCD risk. To determine the threshold that better discriminated between group a1 and group a2, a ROC curve was computed, as shown in figure 3.2. As it can be seen, the ROC curve did not show a marked logarithmic tendency but curved enough to separate from the diagonal. The threshold that maximized sensitivity and specificity (closest point to the upper-left corner) was Δα=0.046, which corresponded to the 75th percentile of Δαin the population (marked with a red dot in figure 3.2). 3.3. Analysis of the relation between increments in Δαand SCD Patients were divided into Δαpositive (Δα+) and negative (Δα-) groups by setting a cut-off point of 0.046 for Δα. Of the 609 patients studied, 457 were thus included in the Δαgroup (Δα≤0.046) and 152 in the Δα+ group (Δα > 0.046). Upon comparison of clinical variables between Δα+ and Δαgroups (Table 3.1), significant differences were found for age, gender, treatment with amiodarone and rate adaptation time t90 for the Tpe series. Data are presented as mean ±standard error of the mean (SEM) for continuous variables and as number and percentage for categorical variables. Two-tailed Mann- Whitney and Fisher exact tests (section 2.3) were used for univariable comparison of quantitative and categorical data, respectively. Survival rate was significantly higher in the Δαgroup for primary and secondary end points (p= 0.003) (Table 3.2). 3.4 Survival analysis 21 Figure 3.2: ROC curve for Δαin the classification of SCD victims. Red dot is associated with the selected threshold. 3.4. Survival analysis As explained in section 2.3.3, a survival analysis involves predicting the fraction of the population which will survive to SCD past a certain time and how clinical variables increase or decrease the odds of survival. Survival probability was estimated by using Kaplan-Meier methods with a comparison of cumulative events by using log-rank tests. The prognostic value of Δαin predicting the endpoints was determined with univariable and multivariable Cox proportional hazards analyses. 3.4.1. Influence of Δαand other variables on SCD Use of Δα Univariable Cox analysis computes the hazard function and the hazard ratio explained in section 2.3.3 by only taking into account the variable under study. In this master’s thesis, Δα. Univariable Cox analysis revealed that Δα+ outcome was associated with all-cause mortality, CD and SCD (Table 3.3), however, Δα+ statistically discriminates better the survival rate to SCD than to CD or total mortality. We can conclude that the time to recurrence was significantly prolonged (approximately doubled) among patients with Δα≤0.046 in comparison with those with Δα>0.046. 22 CHAPTER 3. RESULTS Overall population Δα−Δα+p-value (n= 609) (n= 457) (n=152) Age (y) 63 ±0,5 62 ±0,6 64 ±1,00.040 Gender (men) 426 (70.0 %) 333 (72.9 %) 93 (61.2 %) 0.008 NYHA class III 110 (18.1 %) 81 (17.7 %) 29 (19.1 %) 0.716 LVEF ≤35 % 324 (53.2 %) 242 (53.0 %) 82 (53.9 %) 0.852 Diabetes 232 (38.1 %) 180 (39.4 %) 52 (34.2 %) 0.289 Beta-blockers 425 (69.8 %) 327 (71.6 %) 98 (64.5 %) 0.104 Amiodarone 55 (9.0 %) 30 (6.6 %) 25 (16.4 %) 0.001 ARB or ACE inhibitors 494 (81.1 %) 375 (82.1 %) 119 (78.3 %) 0.338 Average heart rate [beats/min] 72 ±0,5 71 ±0,5 73 ±1,1 0.600 Maximum heart rate [beats/min] 115 ±0,8 115 ±0,9 116 ±1,8 0.589 Heart rate range [beats/min] 43 ±0,6 43 ±0,7 43 ±1,3 0.426 t90[s] (Tpe) 94 ±2,3 88 ±2,7 109 ±3,80.001 QRS >120 ms 236 (38.8 %) 178 (38.9 %) 58 (38.2 %) 0.924 Nonsustained ventricular tachycardia and >240 155 (25.5 %) 121 (26.5 %) 34 (22.4 %) 0.335 Ventricular premature beats in 24 h IAA>3.7μV 143 (23.8 %) 108 (23.9 %) 35 (23.5 %) 0.999 Data are presented as absolute frequencies and percentages and as mean ±standard error of the mean. ACE = angiotensin-converting enzyme; ARB = angiotensin receptor blocker; LVEF = left ventricular ejection fraction; NYHA = New York Heart Association; IAA = Index of maximum alternans; Δα+ = dispersion in the dynamic APDR slopes positive group; Δα−= dispersion in the dynamic APDR slopes negative group. Significant differences between Δα+ and Δα−are indicated in bold. Table 3.1: Characteristics of patients Use of Δαin combination with other variables Although a univariable Cox analysis provides useful information about survival rate, it necessarily ignores the impact of any other factors under investigation, which are known to be associated with SCD end-point, such as left ventricular ejection fraction <35 % [6] or index of average alternans >3.7μV [7]. In clinical investigations, it is more common to have a situation where several covariables potentially affect patient prognosis [34]. By considering all the critical covariables in the adjusted model, it is possible to compute the relative SCD risk prediction of each covariable. Multivariable Cox proportional hazard model was constructed by adjusting for: age, gender, NYHA class, left ventricular ejection fraction <35 %, index of average alternans >3.7 μV, diabetes, use of beta-blockers, amiodarone and angiotensin-converting enzyme or angiotensin receptor blocker inhibitors. Δα+ was the variable with the second highest hazard ratio (2.68), after left ventricular ejection fraction <35 % (hazard ratio 2.87; 95 % CI 1.44-5.69; p= 0.003). The index of average alternans had a hazard ratio of 2.33; 95 % CI 1.30-4.20; p= 0.005. Table 3.3 shows all the hazard ratios computed in the multivariable Cox analysis. The unadjusted Δα+ effect (univariable) may be summarized by a time ratio of 2.54 (1.44-4.50), which, having allowed for other covariables (multivariable) increased slightly to 2.68. This fact suggests that joining together more variables instead of using only Δα+ may improve survival prediction. 3.4 Survival analysis 23 Overall population Δα−Δα+p-value (n= 609) (n= 457) (n=152) Total mortality 137 (22.5 %) 91 (19.9 %) 46 (30.3 %) 0.010 CD 112 (18.4 %) 74 (16.2 %) 38 (25.0 %) 0.021 SCD 48 (7.9 %) 27 (5.9 %) 21 (13.8 %) 0.003 Data are presented as absolute frequencies and percentages. CD = cardiac death; SCD = sudden cardiac death; Δα+ = dispersion in the dynamic APDR slopes positive group; Δα−= dispersion in the dynamic APDR slopes negative group. Significant differences between Δα+ and Δα−are indicated in bold. Table 3.2: Events during follow-up. 3.4.2. Cumulative survival curves Figure 3.3 (top panels) shows the event-free curves for SCD (a) and CD (b), respectively, having divided the population into Δα+ (in green) and Δα- (in blue) groups. As it can be seen from the figure, the survival curves for Δα+ and Δαcan be statistically separated for SCD endpoint (p-value = 0.001) and CD end-point (p-value = 0.007). Patients belonging to Δα+ group have a lower survival rate than those from Δαgroup, supporting the idea that higher values of Δαindicate higher SCD risk. Values of cumulative survival corresponding to CD end-point are lower, indicating that Δαbetter separates SCD risk than other types of cardiac deaths. The Kaplan-Meier curves obtained by using IAA>3.7μV as a risk stratifier are shown in Figure 3.3 (middle panels). Both curves can be statistically separated for SCD end-point (pvalue = 0.004) and CD end-point (p-value = 0.026). Comparing these p-values and curves visual separation, we can affirm that Δαon its own statistically discriminates better from SCD and CD risk than IAA. This master’s thesis uses an ECG-derived marker to predict SCD survival. It would be, then, interesting, to measure the SCD survival prediction rate using the combination of ECG- derived variables. The prediction of Δα+ was, then, combined with the one provided by Index of average alternans >3.7 μV. Combining the information provided by Δαand IAA and sorting the population by those who meet Δα>0.046 and IAA>3.7μV and those who do not, it can be seen how both survival curves are more separated with lower p-values (Figure 3.3 (bottom panels)). Therefore, combining Δαand IAA improves the stratification of SCD risk. The results described above can be discussed in terms of sensitivity and specificity. Using only Δαas a risk marker, 21 true positives out of 152 are obtained (positive predictive value = 13.82 %). Using the combination of Δαand IAA, the population of the risk group is reduced to 35 patients, from which 10 are SCD (positive predictive value = 28.57 %). In clinical practice, it is recommended to implant cardioverter defibrillators to 18 patients if only one is known to suffer from SCD (positive predictive value = 5.56 %). Besides, carrying an implantable cardioverter defibrillator is not comfortable, as it is an invasive device and its discharges are often debilitating. Therefore, minimizing non-necessary cardioverter defibrillator implantations is an important matter. The results obtained in this master’s thesis show how Δαon its own helps to improve the positive predictive value and how, in combination with another ECG-derived SCD risk marker, such as the index of average alternans, the survival prediction is increased up to 28.57 %. This study demonstrates that quantification of the dispersion in the dynamic APDR slopes 24 CHAPTER 3. RESULTS Univariable Multivariable Hazard ratio p-value Hazard ratio p-value (95 % CI) (95 % CI) Total mortality Δα+ (Δα>0.046) 1.68 (1.18-2.40) 0.004 1.60 (1.11-2.32) 0.013 Age - - 1.02 (1.00-1.04) 0.013 Gender (men) - - 1.74 (1.15-2.62) 0.009 NYHA class III - - 2.42 (1.67-3.50) <0.001 LVEF - - 2.29 (1.56-3.36) <0.001 IAA>3.7μV1.59 (1.11-2.29) 0.012 1.53 (1.06-2.21) 0.025 Diabetes - - 1.31 (0.92-1.86) 0.129 Beta-blockers - - 0.74 (0.51-1.07) 0.111 Amiodarone - - 1.59 (0.97-2.60) 0.064 ARB or ACE inhibitors - - 1.10 (0.72-1.68) 0.671 CD Δα+ (Δα>0.046) 1.70 (1.15-2.51) 0.008 1.65 (1.10-2.49) 0.016 Age - - 1.02 (1.00-1.04) 0.026 Gender (men) - - 1.92 (1.20-3.07) 0.006 NYHA class III - - 2.34 (1.55-3.53) <0.001 LVEF ≤35 % - - 2.27 (1.48-3.47) <0.001 IAA>3.7μV1.58 (1.05-2.36) 0.028 1.53 (1.01-2.31) 0.043 Diabetes - - 1.39 (0.94-2.05) 0.097 Beta-blockers - - 0.78 (0.52-1.17) 0.229 Amiodarone - - 1.37 (0.78-2.43) 0.281 ARB or ACE inhibitors - - 1.10 (0.68-1.76) 0.703 SCD Δα+ (Δα>0.046) 2.54 (1.44-4.50) 0.001 2.68 (1.50-4.79) 0.001 Age - - 1.02 (1.00-1.04) 0.231 Gender (men) - - 2.57 (1.18-5.60) 0.018 NYHA class III - - 2.01 (1.07-3.76) 0.029 LVEF ≤35 % - - 2.87 (1.44-5.69) 0.003 IAA>3.7μV2.26 (1.27-4.04) 0.006 2.33 (1.30-4.20) 0.005 Diabetes - - 1.29 (0.71-2.34) 0.402 Beta-blockers - - 1.19 (0.62-2.30) 0.605 Amiodarone - - 1.22 (0.50-2.93) 0.665 ARB or ACE inhibitors - - 0.61 (0.26-1.46) 0.269 CD = cardiac death; CI = confidence interval; SCD = sudden cardiac death; Δα= dispersion in the dynamic APDR slopes; IAA = Index of maximum alternans. Adjusted model includes age, gender, New York Heart Association class, left ventricular ejection fraction <35 %, Index of maximum alternans >3.7 μV, diabetes and use of beta-blockers, amiodarone and angiotensin receptor blocker or angiotensinconverting enzyme inhibitors. Statistically significant values are marked in bold. Table 3.3: Association of dispersion in the dynamic APDR slopes index, Δα, with sudden cardiac death, cardiac death and total mortality. 3.5 Limitations 25 is a strong, independent predictor of SCD in patients with CHF. The index Δαquantifying the dispersion in the dynamic APDR slopes in a 24-hour period independently predicted CD and SCD but did not predict noncardiac mortality (results not shown). These findings support the hypothesis that elevated dispersion in the dynamic APDR slopes reflects abnormal cardiac function predisposing to CD and, more specifically, to SCD. 3.5. Limitations Some of the limitations of the present work are listed below: The number of patients in the“MUSIC”database that died from SCD was relatively low in comparison with the group formed by survivors, victims of non-cardiac causes and victims of cardiac-but-non-SCD causes. The quality of the ECG recordings was not optimal. A thorough preprocessing step was necessary to prepare the signals for further analysis. 26 CHAPTER 3. RESULTS (a) (b) Figure 3.3: Event-free curves for sudden cardiac death (a) and cardiac death (b). Above, using Δα+ on its own; middle, using IAA+ on its own; below, a combination of Δα+ and IAA+. Chapter 4 Practical Training 4.1. Introduction In this master’s thesis, I visited the Zentrum f¨ ur Klinische Forschung (ZKF) (Center for Clinical Trials) of Universit¨ atsSpital in Z¨ urich (USZ) (Switzerland). I worked in the group of Dr. Laurence Jacobs, in collaboration with FIFA - Medical Assessment and Research Centre (F-MARC), as the required part of the practical training of the master thesis in Biomedical Engineering. Sports-Related Sudden Cardiac Death (SRSD) is always a devastating event. This is especially true when the victim is an apparently healthy athlete who in the eyes of the public epitomizes health itself. Among athletes a tragedy like SRSD becomes a powerful symbol, one that is vigorously fueled by the media. However, SRSD remains fundamentally mysterious in medical terms, in the sense that no practical diagnosis of its risk exists, and an understanding of the mechanisms involved is lacking. A large amount of research into SRSD has been generated over the past decade, most of it with mixed results as far as identifying plausible causes. In this practical training I learnt new hypotheses about the field and, especially, methodologies and techniques that have never before been brought to bear on this problem. 4.2. Hypothesis The central hypothesis of our study is that SRSD will not occur unless there is an associated impairment in the dynamics of the Autonomic Nervous System (ANS). Based on this idea, an external stimulus, a trigger, will generate an arrhythmia, which, under normal circumstances, is quickly arrested by an ANS response. When the ANS is impaired, however, the arrhythmia is not arrested and SRSD ensues. While the response system associated with inter- and intracardiac rhythms, mediated through various nonlinear feedback loops, is highly complex, it is undoubtedly deterministic. This forms the basis for the proposed analysis, namely, the concept that deterministic complexity [35] is essential for long-term dynamical stability of the cardiovascular system. The degree of deterministic complexity might be the basis of a diagnosis of health and risk associated with the ANS. 27 34 BIBLIOGRAPHY [12] P. Laguna, R. Jan´e, and P. Caminal, “Automatic detection of wave boundaries in multilead ECG signals: Validation with the CSE database,” Computational Biomedical Research, vol. 27, pp. 45–60, 1994. [13] T. Thom, N. Haase, W. Rosamond, V. J. Howard, J. Rumsfeld, T. Manolio, Z. J. Zheng, K. Flegal, C. O´ Donnell, S. Kittner, D. Lloyd-Jones, D. C. Goff, Y. Hong, R. Adams, G. Friday, K. Furie, P. Gorelick, B. Kissela, J. Marler, J. Meigs, V. Roger, S. Sidney, P. Sorlie, J. Steinberger, S. Wasserthiel-Smoller, M. Wilson, and P. Wolf, “Heart disease and stroke statistics-2006 update: a report from the american heart association statistics committee and stroke statistics subcommittee,” Circulation, vol. 113, pp. e85–191, 2006. [14] W. Bian and L. Tung,“Structure-related initiation of reentry by rapid pacing in monolayers of cardiac cells,” Circulation Research, vol. 98, pp. 29–38, 2006. [15] M. L. Koller, M. L. Riccio, and R. F. Gilmour, “Dynamic restitution of action potential duration during electrical alternans and ventricular fibrillation,” American Journal of Physiology, vol. 275, pp. 1635–1642, 1998. [16] M. L. Riccio, M. L. Koller, and R. F. G. Jr, “Electrical restitution and spatiotemporal organization during ventricular fibrillation,” Circulation Research, vol. 84, pp. 955–963, 1999. [17] R. J. Selvaraj, P. Picton, K.˜ Nanthakumar, and V. S. Chauhan, “Steeper restitution slopes across right ventricular endocardium in patients with cardiomyopathy at high risk of ventricular arrhythmias,” American Journal of Physiology, vol. 292, pp. 1262–1268, 2007. [18] K. R. Laurita, S. D. Girouard, and D. S. Rosenbaum, “Modulation of ventricular repolarization by a premature stimulus: Role of epicardial dispersion of repolarization kinetics demonstrated by optical mapping of the intact guinea pig heart,” Circulation Research, vol. 79, pp. 493–503, 1996. [19] R. Coronel, F. J. G. Wilms-Schopman, T. Opthof, and M. J. Janse, “Dispersion of repolarization and arrhythmogenesis,” Heart Rythm, vol. 6, pp. 537–543, 2009. [20] H. Pak, S. Hong, G. Hwang, H. Lee, S. Park, J. Ahn, Y. Moo, and Y. Kim,“Spatial dispersion of action potential duration restitution kinetics is associated with induction of ventricular tachycardia/fibrillation in humans,” Journal of Cardiovascular Electrophysiology, vol. 15, pp. 1357–1363, 2004. [21] A. Minchol´e, E. Pueyo, J. F. Rodr´ıguez, E. Zacur, M. Doblar´e, and P. Laguna, “Quantification of restitution dispersion from the dynamic changes of the T-wave peak to end, measured at the surface ECG,” IEEE Transactions on Biomedical Engineering, vol. 58, pp. 1172–1182, 2011. [22] R. V´azquez, A. Bay´es-Gen´ıs, I. Cygankiewicz, D. Pascual-Figal, L. Grigorian-Shamagian, R. Pavon, J. Gonzalez-Juanatey, J. Cubero, L. Pastor, J. Ordonez-Llanos, J. Cinca, and A. B. de Luna, “The MUSIC Risk score: a simple method for predicting mortality in ambulatory patients with chronic heart failure,” European Heart Journal, vol. 30, pp. 1088–1096, 2009. [23] K. T. Tusscher and A. Panfilov,“Alternans and spiral breakup in a human ventricular tissue model,” American Journal of Physiology, vol. 291, pp. H1088–H1100, 2006. BIBLIOGRAPHY 35 [24] E. Pueyo, P. Smetana, P. Caminal, A. B. de Luna, M. Malik, and P. Laguna, “Characterization of QT interval adaptation to RR interval changes and its use as a risk-stratifier of arrhythmic mortality in amiodarone-treated survivors of acute myocardial infarction,” IEEE Transactions on Biomedical Engineering, vol. 51, pp. 1511–1520, 2004. [25] E. Pueyo, M. Malik, and P. Laguna, “A dynamic model to characterize beat-to-beat adaptation of repolarization to heart rate changes,” Biomedical Signal Processing and Control, vol. 3, pp. 29–43, 2008. [26] P. Hansen, “Rank-deficient and discrete ill-posed problems,” 1998. [27] J. Nocedal and S. Wright, Numerical Optimization. New York: SpringerVerlag. [28] E. Lehmann, Nonparametrics: Statistical Methods Based on Ranks. 1975. [29] R. A. Fisher, “On the interpretation of x2from contingency tables, and the calculation of P,” Journal of the Royal Statistical Society, vol. 85, pp. 87–94, 1922. [30] J. A. Swets, Signal detection theory and ROC analysis in psychology and diagnostics: collected papers. Lawrence Erlbaum Associates, Mahwah, NJ, 1996. [31] E. L. Kaplan and P. Meier,“Nonparametric estimation from incomplete observations,” Journal of American Statistics Association, vol. 53, pp. 457–481, 1958. [32] D. R. Cox and D. Oakes, Analysis of survival data. Chapman & Hall, 1984. [33] N. Mantel, “Evaluation of survival data and two new rank order statistics arising in its consideration,” Cancer Chemotherapy Reports, vol. 50, pp. 163–170, 1966. [34] M. J. Bradburn, T. G. Clark, S. B. Love, and D. G. Altman, “Survival analysis part II: Multivariate data analysis-an introduction to concepts and methods,” British Journal of Cancer, vol. 89, pp. 431–436, 2003. [35] R. Savit and M. Green,“Time series and dependent variables,” Physica D, vol. 1991, pp. 95– 116, 1990. [36] SCHILLER’s medilog Holter system. www.coremed.se/pdf/Medilog.pdf. [37] J. Ram´ırez, A. Minchol´e, P. Laguna, and E. Pueyo, “Characterization of cardiac repolarization response to heart rate changes provoked by a tilt test,” Computers in Cardiology, 2012. 36 BIBLIOGRAPHY APPENDIX Sudden Cardiac Death Survival Prediction from Restitution Dispersion Analysis Julia Ramírez1, 2, Ana Mincholé3, Juan Bolea1, 2, Pablo Laguna1, 2, Esther Pueyo2, 1 1 Grupo de Tecnologías de las Comunicaciones Instituto de Investigación en Ingeniería de Aragón (I3A). Universidad de Zaragoza, Mariano Esquillor s/n, 50018, Zaragoza, Spain. Tel. +34-976762707, Fax +34-976762043, e-mail: [email protected] 2 CIBER – Bioingeniería, Biomateriales y Nanomedicina, Spain 3 University of Oxford, Oxford, United Kingdom Abstract Increase in the dispersion of action potential duration restitution (APDR) has been associated with sudden cardiac death (SCD). A marker, Δα, was proposed to quantify APDR dispersion from the electrocardiogram (ECG). 609 ECG recordings were analysed. The marker Δα stratified patients according to their risk of suffering from SCD. Introduction APDR measures the relationship between the action potential duration and the RR interval at steady-state pacing. Due to heterogeneities in the ventricles, APDR presents spatial variations generally termed APDR dispersion. An increase in APDR dispersion has been associated with higher propensity to suffer from ventricular arrhythmias and SCD. Recently, a marker, Δα, which accounts for the rate normalized differences of the Tpe interval, was proposed to quantify APDR dispersion from the ECG at steady-state conditions [1]. Materials and methods Materials Consecutive patients were enrolled in the MUSIC (MUerte Súbita en Insuficiencia Cardiaca) study, a prospective, multicenter study designed to assess risk predictors for cardiovascular mortality in ambulatory patients. The Holter recordings of 609 patients (48 victims of SCD, 64 of other cardiac causes, 25 of non-cardiac death causes and 472 survivors) with sinus rhythm were available for the present study. Each recording consisted of 3 ortogonal ECG leads, sampled at 200 Hz. In this study, the population in the database was splitted into two groups: SCD victims (group 1) and victims of other cardiac causes, non-cardiac causes and survivors (group 2). Patients were followed up every 6 months for a median of 48 months. SCD was defined as (1) a witnessed death occurring within 60 minutes from the onset of new symptoms unless a cause other than cardiac failure was obvious, (2) an unwitnessed death (< 24 hours) in the absence of preexisting progressive circulatory failure or other causes of death, or (3) death during attempted resuscitation. Methods Preprocessing of the ECG signals included low pass filtering at 40 Hz to remove electric and muscle noise, cubic splines interpolation for baseline wander removal and ectopic beats detection. Principal Component Analysis was applied over the three leads to emphasize T-wave energy and improve delineation. A Single-Lead-and-rules delineation technique was applied to mark the onsets and offsets of the T-wave. From the annotation marks, RR, QT and Tpe series were obtained and subsequently interpolated at a sampling frequency of 1 Hz. The spatial dispersion of APDR slopes was estimated from the ECG as ∆𝛼 =𝜕𝑇𝑝𝑒 𝜕𝑅𝑅 measured at steady-state RR intervals [1]. Results and discussion The mean value of Δα in the study population was 0.028 ± 0.076 and the 25th, 50th and 75th percentiles were 0.005, 0.022 and 0.046, respectively. II Reunión Jóvenes Investigadores del Instituto de Investigación en Ingeniería de Aragón (I3A) Δα discriminated between the group formed by SCD victims (group 1) and the group composed of the other patients (group 2), with mean ± SEM values of: Δα = 0.052 ± 0.013 for the former and Δα = 0.026 ± 0.003 for the latter (p = .048). Patients were divided into Δα positive (Δα+) and negative (Δα-) groups by setting a cut-off point of 0.046 for Δα, corresponding to the 75th percentile of the distribution of Δα in the population. Of the 609 patients studied, 457 (75.0%) were included in the Δαgroup (Δα ≤ 0.046) and 152 (25%) in the Δα+ group (Δα > 0.046). A two-tailed Fisher exact test showed an existing effect of being a SCD victim on having Δα>0.046 (p = .003), with a survival rate higher in the Δαgroup for SCD endpoint. In a survival analysis, Cox regression revealed that a Δα+ outcome was associated with SCD (p = .001), as shown in Figure 1. Conclusions This study demonstrates that quantification of APDR from the ECG is a strong predictor of SCD. This finding supports the hypothesis that elevated APDR dispersion reflects abnormal cardiac function predisposing to SCD. REFERENCIAS [1] Mincholé, A., Pueyo, E., Rodríguez, J.F., Zacur, E., Doblaré, M. and Laguna, P. Quantification of restitution dispersion from the dynamic changes of the T-wave peak to end, measured at the surface ECG. IEEE Transactions on Biomedical Engineering, vol. 58, 2011, pp. 1172-1182. [2] Pueyo, E., Smetana, P., Caminal, P., de Luna, A.B., Malik, M. and Laguna, P. Characterization of QT interval asaptation to RR interval changes and its use as a risk-stratifier of arrhythmic mortality in amiodarone-treated survivors of acute myocardial infarction. IEEE Transactions on Biomedical ENgineering, vol. 51, 2004, pp. 1511-1520. Fig 1: Event-free curves for SCD 2 Analysis of Repolarization Dispersion to Predict Sudden Cardiac Death Survival J. Ramírez1, 2, A. Mincholé2, 1, P. Laguna1, 2 and E. Pueyo2, 1 1Communications Technology Group (GTC), Aragón Institute of Engineering Research (I3A), IIS Aragón, Universidad de Zaragoza, Zaragoza, Spain 2CIBER- Bioingeniería, Biomateriales y Nanomedicina, Spain [email protected], m[email protected], [email protected], epuey[email protected] Abstract - Enhancement of dynamic restitution of the action potential duration (APDR) dispersion has been associated with higher propensity to suffer from ventricular arrhythmias and sudden cardiac death (SCD). Recently, a marker, Δα, was proposed to quantify APDR dispersion from the electrocardiogram (ECG) by computing the ratio between T-peak-to-T-end (Tpe) and RR intervals. Holter ECG recordings of 609 subjects were analysed. RR and Tpe series were computed and the risk marker Δα was derived. The marker Δα stratified patients according to their risk of suffering from ventricular arrhythmias that could lead to SCD, with larger repolarization dispersion indicating lower survival probability. I. INTRODUCTION APDR measures the relationship between the action potential duration and the RR interval at steady-state pacing. Due to heterogeneities in the ventricles, APDR presents spatial variations generally termed APDR dispersion. Enhancement of APDR dispersion has been associated with higher propensity to suffer from ventricular arrhythmias and SCD. A marker, Δα, was recently published to non-invasively estimate the quantification of APDR dispersion from the ECG by computing the ratio between differences in the Tpe and RR intervals at different steady-state conditions [1]. II. METHODS Holter ECG recordings of 609 subjects (48 victims of SCD, 64 of other cardiac causes, 25 of non-cardiac death causes and 472 survivors) from the ``MUSIC'' database were analysed. ECGs were delineated using a single-lead procedure over the first principal component calculated to emphasize the T-wave. RR and Tpe series were computed and Tpe/RR dynamics was estimated using a nonlinear system with memory [2], from which the risk marker Δα was derived, by applying the following equation: ∆𝛼 = 𝜕𝑇𝑝𝑒 𝜕𝑅𝑅 Figure 1.Event-free curves for SCD. III. RESULTS AND DISCUSSION Δα discriminated between the group formed by SCD victims and the group composed of the other subjects, with mean ± SEM values of: Δα = 0.052 ± 0.013 for the former and Δα = 0.026 ± 0.003 for the latter (p = 0.026). Following the hypothesis that an enhancement of APDR dispersion has been associated with higher propensity to suffer from SCD, a threshold on the third quartile of Δα values was set in a survival analysis. Statistically significantly different event probabilities were obtained in both strata of the population (p = 0.001), as shown in Figure 1. IV. CONCLUSIONS The marker Δα stratifies patients according to their risk of suffering from ventricular arrhythmias that could lead to SCD, with larger repolarization dispersion indicating lower survival probability. Future research involves applying this method over a database composed of professional athletes, to predict sports-related sudden cardiac death. REFERENCES [1] A. Mincholé, E. Pueyo, J. F. Rodríguez, E. Zacur, M. Doblaré and P. Laguna, “Quantification of restitution dispersion from the dynamic changes of the T-wave peak to end, measured at the surface ECG”, IEEE Transactions on Biomedical Engineering, vol.58, 2011, pp. 1172-1182. [2] E. Pueyo, P. Smetana, P. Caminal, A. B. de Luna, M. Malik and P. Laguna, “Characterization of QT interval adaptation to RR interval changes and its use as a risk-stratifier of arrhythmic mortality in amiodarone-treated survivors of acute myocardial infarction”, IEEE Transactions on Biomedical Engineering, vol.51, 2004, pp. 1511-1520. Prediction of Sudden Cardiac Death in Chronic Heart Failure Patients by Analysis of Restitution Dispersion Julia Ram´ ırez1,2, Ana Minchol´ e3, Juan Bolea1,2, Pablo Laguna1,2, Esther Pueyo2,1 1Communications Technology Group (GTC), Arag´ on Institute of Engineering Research (I3A), IIS Arag´ on, Universidad de Zaragoza, Zaragoza, Spain 2CIBER - Bioingenier´ ıa, Biomateriales y Nanomedicina, Spain 3Department of Computer Science, University of Oxford, Oxford, United Kingdom Abstract An increase in the dispersion of action potential duration restitution (APDR) has been associated with higher propensity to suffer from ventricular arrhythmias and sudden cardiac death (SCD). Recently, a marker, Δα, was proposed to non-invasively quantify APDR dispersion from the electrocardiogram (ECG) by computing the ratio between differences in the T-peak-to-T-end (Tpe)and RR intervals at different steady-state conditions. Holter ECG recordings of patients with chronic heart failure divided into two groups, one consisting of victims of SCD and the other of victims of other causes and survivors, were analyzed. ECGs were delineated using a single-lead procedure over the first principal component calculated to emphasize the T-wave. Δαdiscriminated between the groups formed by SCD and non-SCD victims, respectively, with mean ±SEM values of: Δα= 0.052 ±0.013 for the former and Δα= 0.026 ±0.003 for the latter (p<0.048). In a survival analysis where a threshold on the third quartile of Δαvalues was set, statistically significantly different event probabilities were obtained in both stratas of the population (p= 0.003). The marker Δαstratifies patients according to their risk of suffering from ventricular arrhythmias that could lead to SCD, with larger restitution dispersion indicating lower survival probability. 1. Introduction Sudden cardiac death (SCD) remains an important cause of mortality in patients with mild-to-moderate heart failure. A number of indices have been proposed as SCD predictors, including left ventricular ejection fraction (currently the only recommended marker to risk stratify patients [1]) and T-wave alternans [2]. Nevertheless, further research is needed to provide an index or a combination of indices with improved capacity to identify patients at risk of SCD. Heart rate dependence of action potential duration (APD), also called restitution kinetics, is thought to be critical in activation instability and, therefore, provides relevant information for ventricular arrhythmic risk stratification [3]. The dynamic APD restitution (APDR) curve, measured using the so-called dynamic restitution protocol, quantifies the relationship between the APD and the RR interval at steady-state when pacing at different RR values. Heterogeneities in the ventricle lead to non-uniform restitution properties, which makes APDR curves present spatial variations [4]. Dispersion is a measure of that spatial variation. Recent studies have suggested that dispersion in the APDR curves may act as a potent arrhtymogenic substrate and increments in that dispersion have been associated with greater propensity to suffer from ventricular tachycardia/fibrillation [5]. The main limitation on the usability of APDR dispersion as a risk index is that its quantification usually requires invasive procedures. In [6], a method to indirectly estimate dispersion of restitution slopes by making only use of the surface electrocardiogram (ECG) was developed. An ECG index, Δα, that quantifies dispersion in the dynamic APDR slopes by characterizing the relationship between the T-peak-to-T-end (Tpe) and the RR intervals at different steady-state conditions, was proposed and evaluated. In this work, we present a fully automated method to analyze APDR dispersion in ambulatory recordings and we show that the ECG index, Δα, proposed in [6], is an independent predictor of SCD in patients with chronic heart failure (CHF). 2. Materials and Methods 2.1. Materials Consecutive patients with symptomatic CHF corresponding to NYHA classes II and III were enrolled in the MUSIC (MUerte S´ ubita en Insuficiencia Cardiaca) study, a prospective, multicenter study designed to assess risk predictors for cardiovascular mortality in ambulatory patients