Near-term prediction of sustained ventricular arrhythmias applying artificial intelligence to single-lead ambulatory electrocardiogram.

Summary

Using 14-day single-lead ambulatory ECGs from six countries, a deep learning model predicted sustained VT within the next 13 days with AUROC ~0.95 and high specificity. Saliency analyses implicated premature ventricular complex burden and early depolarization timing as key predictive features.

Key Findings

• Deep learning on single-lead ambulatory ECGs achieved AUROC 0.957 (internal) and 0.948 (external) for near-term sustained VT prediction.
• At fixed specificity of 97%, sensitivities were 70.6% (internal) and 66.1% (external).
• Predicted 80–81% of rapid sustained VT (≥180 bpm) and 90% of VT degenerating to VF.
• Saliency maps highlighted premature ventricular complex burden and early depolarization time as important predictors.

Clinical Implications

Could inform intensified monitoring, expedited electrophysiology referral, remote alerting in wearables, and personalized ICD programming to mitigate near-term VT/VF risk.

Limitations

• Retrospective design; lack of prospective, real-time clinical deployment and outcomes testing.
• Event rate was low (0.5%), raising potential class imbalance and calibration challenges.

Future Directions

Prospective, randomized implementation studies to test clinical impact (alerts, workflow integration) and evaluation across diverse devices and populations.