Signals classification to assist clinician decision-making in cardiac ablation for atrioventricular nodal reentry tachycardia

Ablation procedures are widely recognized as the primary intervention for atrioventricular reentrant tachycardia, leveraging advancements in multipolar catheters, high-definition voltage mapping, and robotic assistance. However, these procedures remain associated with recurrence risks and complications such as steam-pop, stroke, and procedural fatigue. This study aims to develop a robust classifier as a foundational step toward an assistive predictive algorithm capable of distinguishing atrial signals recorded during ablation as effective, indifferent, or dangerous. Electrogram (EGM) signals from 12 ablation procedures were analysed, categorized into indifferent, effective, and dangerous. Signal alignment was performed using ECG R-wave detection to ensure consistent ventricular referencing. Filtering approaches, including band-pass Butterworth and wavelet-based filtering, were evaluated, but due to the non-stationary nature of EGM signals, filtering was ultimately excluded. Feature extraction encompassed envelope-based metrics, template-matching analysis, and Short-Time Fourier Transform (STFT) features, alongside literature-inspired descriptors. A multi-step feature selection strategy was employed, including univariate analysis, correlation assessments, and feature importance evaluation across models. Classification performance was assessed using knowledge-based, decision tree, multinomial logistic regression (MLR), and Support Vector Machine (SVM) models. The best knowledge-based classifier ( \textit{KB-V2}) achieved an F1-score of 0.72, demonstrating systematic misclassification issues. The decision tree model improved overall performance (F1-score: 0.79), yet struggled with effective and dangerous signal differentiation. MLR exhibited superior recall for effective signals but does not ensure sufficient performances (F1-score: 0.69). The SVM classifier outperformed all models, yielding an overall F1-score of 0.84 and significantly reducing misclassification errors. Feature importance analysis revealed that envelope-derived metrics were the most influential, followed by morphological features and STFT-based descriptors. However, limitations in data structure, filtering, and feature extraction constrained classification performance. The presence of misclassification errors, particularly for dangerous signals, highlights the need for further refinement. This study establishes a preliminary framework for an assistive classification system for cardiac ablation procedures. Future research will focus on optimizing feature extraction, integrating advanced filtering methodologies, and refining classification models to enhance reliability, with the ultimate goal of improving clinical decision support and patient outcomes.