Efficacy of Spectral Signatures for The Automatic Classification of Abnormal Ventricular Potentials in Substrate-Guided Mapping Procedures

Giulia Baldazzi1, Marco OrrĂ¹2, Mirko Matraxia3, Graziana Viola4, Danilo Pani5
1DIEE, University of Cagliari; DIBRIS, University of Genova;, 2DIEE, University of Cagliari, 3Medical Concept Lab, 4EP STAFF, Santissima Annunziata Hospital, Sassari, Italy, 5DIEE - University of Cagliari


Aims: Recent studies highlighted some peculiar spectral signatures of post-ischemic ventricular tachycardias (VT) electrograms (EGMs) that could be used as features in machine learning (ML) applications for the automatic recognition of arrhythmogenic targets for the VT treatment. This study aimed to investigate the impact of the information retrieved from the frequency-domain analysis in the modeling of supervised ML tools for the classification of physiological and abnormal ventricular potentials (AVPs) in EGMs.

Methods: The power spectrum and the power spectral density (PSD) were computed for each EGM, and four features describing the PSD morphology, i.e., the peak frequency, the mean frequency, the mean spectral power, and the power spectrum ratio, were estimated. The relative power contents of the ventricular potentials in distinct 20-Hz sub-bands below 320 Hz were also computed and considered as features. Two different classifiers were considered, i.e. support vector machine (SVM) and a K-nearest neighbor (KNN). In order to assess the efficacy of exploiting these spectral features for AVPs recognition, the two models were assessed using a 10-time 10-fold cross-validation scheme and computing accuracy, specificity, sensitivity, and F1-score metrics. To this aim, 1504 bipolar intracardiac EGMs from nine post-ischemic VT patients, acquired with the CARTO 3V6 mapping system, were retrospectively annotated by an expert cardiologist.

Results: Recognition accuracies of over 80% were obtained. In particular, in both classifiers, accuracy, and F1-score values were found to be above 81%. Furthermore, it is possible to appreciate balanced sensitivity and specificity scores.

Conclusion: The achieved results suggest that the use of spectral features can be successfully adopted to train ML models for the automatic identification of AVPs. These findings deserve further investigations, also exploiting additional features extracted from multiple domains, to assess their robustness and the possibility to improve the performance in targeting arrhythmogenic sites.