Arrhythmogenic sites identification in post-ischemic ventricular tachycardia electrophysiological studies by explainable deep learning

Background and objective: Abnormal ventricular potentials (AVPs) in intracardiac electrograms (EGMs) are frequently considered as markers of arrhythmogenic sites in post-ischemic ventricular tachycardia (VT) during electroanatomic mapping (EAM) procedures. Their detection is strongly operator-dependent and time-consuming. This work explores the adoption of explainable deep learning to support the discrimination between physiological EGMs and AVPs. Methods: Three convolutional neural networks were trained to discriminate the target signals based on their time-frequency representations by synchrosqueezed wavelet transform. The efficacy of the method was assessed on 2561 real bipolar EGMs collected from nine post-ischemic VT patients. Results: The proposed approach achieved high performance, with accuracy levels reaching up to 89%. It also demonstrated coherent localization of the arrhythmogenic sites with respect to conventional voltage and local activation time maps. Moreover, by using saliency maps, AVPs discriminant signatures were highlighted at high frequencies (i.e., in the 103-125 Hz band, which was generally relevant for every network), in line with prior evidence. Conclusion: For the first time, deep learning has been successfully applied and robustly evaluated in the field. The proposed approach paves the way to the development of effective AI-driven systems. These systems will enable faster, trustworthy and operator-independent identification of AVPs in VT EAM procedures. Furthermore, even without injecting prior knowledge in the adopted models, the analysis of saliency maps revealed that CNNs are prone to autonomously select time-frequency ranges of the EGMs in agreement with the current knowledge.