Pattern recognition approaches for clinical decision in treatment of atrioventricular nodal reentry tachycardia

Atrioventricular Nodal Reentry Tachycardia (AVNRT) is among the most common cardiovascular disorders and is characterized by an abnormal acceleration of the heart rhythm. Catheter ablation represents the standard treatment for several arrhythmias, including AVNRT, and consists of delivering radiofrequency energy to selectively eliminate cardiac tissue responsible for pathological conduction pathways. Despite its effectiveness, the accurate identification of the optimal ablation target site still largely relies on the clinician’s experience. This subjectivity may increase the risk of arrhythmia recurrence, irreversible myocardial damage, or the onset of atrioventricular block. To reduce complications, it is therefore crucial to correctly identify effective, ineffective, and dangerous regions. This study focuses on the extraction of informative and discriminative features based solely on prior knowledge of the electrophysiology of AVNRT signals, with the aim of developing an automatic classification system for these data. The dataset consists of 12 ablation procedures and is labelled based on a broader set of clinical considerations. A preliminary analysis of the data revealed that a subset of traces did not follow the expected morphology described in the literature. Focusing specifically on electrophysiological morphology, a review of the labels was conducted to evaluate how well the signals conform to the expected morphological characteristics of their assigned classes. Three stratified datasets were generated based on the results of the scoring process. The complete dataset and these datasets were used for further analysis in order to evaluate performance on datasets with higher complexity and heterogeneity. The main focus of this work was the exploration of pattern recognition approaches through the development of ad-hoc feature extraction pipelines based on classical cross-correlation method and, subsequently, on a more flexible approach using Dynamic Time Warping (DTW). The proposed pipelines were first evaluated using simulated data, which confirmed their effectiveness in identifying the correct waveform but also highlighted limitations related to amplitude variability and signal morphology. The pipelines were then applied to real EGM signals across all datasets, from which the features were extracted. The extracted features were then used as input for several classifiers, including knowledge-based classifiers, decision tree, random forest, and support vector machine (SVM) models. In general, the dataset containing only traces that reflect the expected class characteristics achieved satisfactory results, particularly with the SVM classifier (reaching an F1-score of 0.99). In contrast, the complete dataset, which includes all traces, showed lower performance due to reduced generalization ability, especially for the dangerous class (reaching an F1-score of 0.74 using knowledge-based classifiers). These results lead to the conclusion that the pipelines used to extract template-matching features produce discriminative features when the data follow electrophysiological expectations, but they are not able to fully generalize the complexity of the traces. This limitation is mainly related to data heterogeneity and the need to include additional feature types to better characterize signal variability. Future research will focus on integrating additional feature extraction approaches, ad-hoc filtering methodologies, and more robust classification models, with the ultimate goal of improving clinical decision support and patient safety.