Sudden cardiac arrest is a significant health issue, with post-ischemic ventricular tachycardia (VT) being one of its main causes. In clinical practice, radiofrequency catheter ablation (RFCA) is the most adopted treatment for scar-related VT, guided by the analysis of intracardiac electrograms (EGMs). However, the identification of arrhythmogenic areas is complex, time-consuming, and operator-dependent, leading to a variable long-term success rate of the procedure ranging from 30% to 70%. This Ph.D. thesis proposes innovative tools to improve the objective and reproducible characterization of intracardiac EGMs, acquired during VT mapping and RFCA procedures. Thus, the main objectives were the creation of a multi-expert validated dataset of post-ischemic VT EGMs, the investigation of novel multi-domain features capable of characterizing EGMs across multiple and complementary signal domains, and the development of automatic models for EGM classification and temporal delineation with the final aim of providing methodological components that may support future computer-aided tools for electroanatomical mapping and RFCA strategies. The research combined signal processing, statistical analysis, and machine learning within a unified framework. First, the work focused on the creation and validation of a multi-expert annotated database of post-ischemic VT EGMs, named ARGO. Being the first structured and validated dataset of its kind, it served as the foundation for all subsequent analyses. On this basis, conventional EGM characterizations used in clinical practice and novel non-conventional features have been investigated and combined. Specifically, EGM properties across time, spectral, time–frequency, and spatial domains were explored. Moreover, entropy measures were investigated to quantify the intrinsic irregularity of EGMs, exploring them as digital biomarkers of pathological activity, along with the study and development of delineation algorithms for both EGMs and their pathological component to quantify local activation timing. Based on these analyses, supervised learning models were developed by integrating multiple-domain features, along with deep learning approaches, to automatically recognize abnormal ventricular potentials (AVPs) associated with arrhythmogenic areas. The ARGO dataset represents the first multi-expert validated dataset for post-ischemic VT, including detailed annotations describing both the physiological or pathological nature of each EGM and the temporal delineation of pathological components. The statistical analysis of the newly introduced features demonstrated potential in discriminating physiological EGMs from AVPs. The AVP delineation framework allowed for accurately identifying activation boundaries, achieving median absolute errors of 11 ms for the AVP onset and 9 ms for its end. Overall, supervised learning models achieved high performance in distinguishing between EGM classes, with feature-based models achieving accuracies up to 94% and high sensitivity and specificity in leave-one-subject-out validation, above those obtained by deep learning models (close to 90%). This work integrates engineering innovation with clinical practice in cardiac electrophysiology. The results enable a more objective and reproducible characterization of the arrhythmogenic substrate, representing a methodological step toward future computer-aided interpretation of intracardiac EGMs in RFCA workflows. The proposed approaches pave the way for future research on other malignant arrhythmias and contribute to the establishment of open and standardized resources.
Signal processing and artificial intelligence tools for automated ventricular tachycardia electrogram analysis
ORRU', MARCO
2026-05-29
Abstract
Sudden cardiac arrest is a significant health issue, with post-ischemic ventricular tachycardia (VT) being one of its main causes. In clinical practice, radiofrequency catheter ablation (RFCA) is the most adopted treatment for scar-related VT, guided by the analysis of intracardiac electrograms (EGMs). However, the identification of arrhythmogenic areas is complex, time-consuming, and operator-dependent, leading to a variable long-term success rate of the procedure ranging from 30% to 70%. This Ph.D. thesis proposes innovative tools to improve the objective and reproducible characterization of intracardiac EGMs, acquired during VT mapping and RFCA procedures. Thus, the main objectives were the creation of a multi-expert validated dataset of post-ischemic VT EGMs, the investigation of novel multi-domain features capable of characterizing EGMs across multiple and complementary signal domains, and the development of automatic models for EGM classification and temporal delineation with the final aim of providing methodological components that may support future computer-aided tools for electroanatomical mapping and RFCA strategies. The research combined signal processing, statistical analysis, and machine learning within a unified framework. First, the work focused on the creation and validation of a multi-expert annotated database of post-ischemic VT EGMs, named ARGO. Being the first structured and validated dataset of its kind, it served as the foundation for all subsequent analyses. On this basis, conventional EGM characterizations used in clinical practice and novel non-conventional features have been investigated and combined. Specifically, EGM properties across time, spectral, time–frequency, and spatial domains were explored. Moreover, entropy measures were investigated to quantify the intrinsic irregularity of EGMs, exploring them as digital biomarkers of pathological activity, along with the study and development of delineation algorithms for both EGMs and their pathological component to quantify local activation timing. Based on these analyses, supervised learning models were developed by integrating multiple-domain features, along with deep learning approaches, to automatically recognize abnormal ventricular potentials (AVPs) associated with arrhythmogenic areas. The ARGO dataset represents the first multi-expert validated dataset for post-ischemic VT, including detailed annotations describing both the physiological or pathological nature of each EGM and the temporal delineation of pathological components. The statistical analysis of the newly introduced features demonstrated potential in discriminating physiological EGMs from AVPs. The AVP delineation framework allowed for accurately identifying activation boundaries, achieving median absolute errors of 11 ms for the AVP onset and 9 ms for its end. Overall, supervised learning models achieved high performance in distinguishing between EGM classes, with feature-based models achieving accuracies up to 94% and high sensitivity and specificity in leave-one-subject-out validation, above those obtained by deep learning models (close to 90%). This work integrates engineering innovation with clinical practice in cardiac electrophysiology. The results enable a more objective and reproducible characterization of the arrhythmogenic substrate, representing a methodological step toward future computer-aided interpretation of intracardiac EGMs in RFCA workflows. The proposed approaches pave the way for future research on other malignant arrhythmias and contribute to the establishment of open and standardized resources.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
