Exploring Transfer Learning for Ventricular Tachycardia Electrophysiology Studies

Andrea Pitzus1, Giulia Baldazzi2, Marco Orr├╣1, Alberto Valdes Rey1, Graziana Viola3, Luigi Raffo1, Petar Djuric4, Danilo Pani1
1DIEE - University of Cagliari, 2DIEE, University of Cagliari; DIBRIS, University of Genova;, 3Division of Cardiology, San Francesco Hospital, 4Stony Brook University, Stony Brook


Abstract

Rationale: Post-ischemic ventricular tachycardia (VT) arrhythmogenic sites are usually identified by looking for abnormal ventricular potentials (AVPs) in intracardiac electrograms (EGMs). Unfortunately, the accurate recognition of AVPs is a challenging issue, also because of the intrinsic variability between the AVPs. Given the high performance of deep neural networks in several scenarios, in this work we explored the use of transfer learning (TL) in intracardiac electrophysiology for AVPs detection. Methods: The AlexNet convolutional neural network (CNN) was adopted along with TL to discriminate between physiological EGMs and AVPs. A set of bipolar intracardiac EGMs were collected from nine post-ischemic VT patients during electro-anatomical mapping procedures by the CARTO®3V6 mapping system. From these procedures, a balanced dataset of 752 physiological potentials and 752 AVPs, retrospectively annotated by an expert cardiologist, was considered. For each potential, a 500 ms window around the reference beat was considered and the time-frequency representation was generated by computing the synchro-squeezed wavelet transform (SSWT) . For the re-training of CNN through TL, an early-stop condition on the validation loss was imposed. Results: The efficacy of the method was assessed in a 10-fold cross-validation scheme, by computing accuracy, specificity, sensitivity, precision, and F1-score. The proposed approach allows obtaining high recognition results, reaching values above 90% in all the investigated performance indexes, thus revealing a balanced and highly accurate classification of both AVPs and physiological potentials. Conclusion: This work, exploiting a CNN on the SSWT representation of intracardiac EGMs, demonstrated the effectiveness of deep learning in the recognition of both AVPs and physiological potentials in post-ischemic VT EGMs, paving the way for its use in supporting clinicians in targeting arrhythmogenic sites during substrate-guided mapping procedures. In addition, this study further confirms the efficacy of the TL approach even in case of very limited dataset sizes.