Ventricular arrhythmias (VAs) can lead to sudden cardiac death if not promptly managed. Implantable cardioverter defibrillators (ICDs) typically deliver therapy after VA onset, leaving a limited window for intervention and potentially causing adverse effects such as inappropriate shocks. This study investigates predicting VA onset using intracardiac electrograms (EGMs) recorded by subcutaneous ICDs immediately preceding the VA event.
The training set included 10,913 EGM recordings, with 236 upstream EGMs recorded before VA onset from Emory University Hospital, while the test set included 3,712 recordings with 51 upstream EGMs from Stanford Hospital. A deep learning model—a residual neural network with an attention mechanism—was first pretrained on a large ECG dataset due to the small size of the VA dataset, and then fine-tuned end to end for the VA prediction task.
During method development, we trained five models corresponding to five cross-validation folds using Emory dataset. On the independent Stanford test set, the model achieved strong performance, with a mean AUROC of 0.95 ± 0.01 and an AUPRC of 0.62 ± 0.06. Temporal analysis showed that VA can be predicted up to 37 seconds before onset, achieving high sensitivity (0.89 ± 0.02) and specificity (0.93 ± 0.02), extending the prediction window by up to 32 seconds compared to previous studies. Notably, the proximity of upstream EGMs to the VA onset did not significantly affect AUROC, suggesting that earlier detection may allow sufficient time for therapeutic intervention. Although precision remains relatively low (0.15 ± 0.03), it is important to consider that precision is prevalence-dependent, and in our case, the prevalence of upstream EGMs is only 1.4%. Nonetheless, further improvements in precision are necessary for clinical application.
These findings highlight the potential of predicting VA onset from intracardiac EGMs, enabling timely therapeutic intervention before the arrhythmia fully develops.