Feasibility of single-beat detection of ventricular late potentials on standard ECG leads via interpretable gradient...

Ventricular late potentials (VLPs) are markers of arrhythmogenic substrate, but conventional assessment using signal-averaged ECG (SAECG) requires prolonged acquisition and operator-dependent artifact handling, limiting scalability and ambulatory use. Single-beat detection of VLP-like activity from standard surface ECG remains insufficiently validated.

To evaluate the technical feasibility of interpretable single-beat detection of VLP-like perturbations from standard ECG leads without signal averaging.

Using the MIMIC-IV-ECG database, we analyzed 120,000 beats from leads II, V2, and V6. Because large public datasets with beat-level clinically adjudicated VLP labels are not currently available, physiologically constrained synthetic VLP-like signals were injected into a subset of beats to create a controlled feasibility benchmark. For each beat, more than 200 features were extracted, including time-domain statistics, frequency-domain measures, wavelet coefficients, autocorrelation features, and localized windowed summaries. Ten classifiers were optimized using nested patient-wise cross-validation and evaluated in five settings: single-lead detection, cross-lead generalization, mixed-lead training, reduced training size, and class-imbalance robustness.

Gradient-boosted ensembles, particularly XGBoost and CatBoost, achieved strong discrimination on held-out single-beat data (AUC > 0.99; F1 > 0.93), while remaining stable with 10% of the training data and 5% positive-class prevalence. Performance was also robust in lead-transfer experiments. SHAP analysis identified localized entropy, dispersion, and related high-frequency descriptors in late post-R windows as the dominant predictors.

These findings support the methodological feasibility of interpretable single-beat detection of VLP-like signatures from routine surface ECG under controlled synthetic conditions. Validation on clinically adjudicated cohorts and external datasets is required before clinical translation.