Separate Assessment of Hemodynamic Information Content and Interchannel Correspondence in the Recognition of False Alarms of Ventricular Tachycardia

The study separately evaluates the additional diagnostic information provided by hemodynamic signals for ventricular tachycardia false-alarm discrimination and the effect of preserving event-specific correspondence between electrical and hemodynamic signals during training. Thirty-second segments of two electrocardiographic leads, photoplethysmography, and invasive arterial blood pressure from the VTaC dataset were analyzed. Morphological, dynamic, and cross-channel features were extracted from the signals and used to train a gradient boosting classifier. Performance was assessed using repeated record-group cross-validation with a separate internal subset for threshold selection. The effect of preserving event-specific correspondence during training was evaluated using a control permutation of hemodynamic signals performed only within the training partitions, followed by complete feature re-extraction. Adding hemodynamic morphology increased ROC-AUC from 0.8612 to 0.8721. Most of the multimodal gain was retained under the control permutation of hemodynamic signals, while the original channel pairings additionally outperformed the permuted configuration in ROC-AUC and PR-AUC. No consistent advantage of the original pairings was observed for balanced accuracy. The results show that hemodynamic signals contain diagnostic information beyond that available from electrocardiography and indicate a small effect of preserving event-specific channel correspondence during training.