Early prediction of ventricular fibrillation through continuous ECG analysis and machine learning techniques

Ventricular fibrillation (VF) is a life-threatening arrhythmia in which timely intervention is critical to prevent sudden cardiac death (SCD). Early prediction can improve survival rates by enabling pre-emptive clinical actions or early patient response. However, most existing VF prediction methods rely on isolated electrocardiogram (ECG) segments at fixed time points, limiting their practical application. This study introduces a continuous ECG analysis framework for early VF prediction using heart rate variability (HRV) features. ECG data from the PhysioNet Sudden Cardiac Death (SCD) and Normal Sinus Rhythm (NSR) databases were analysed. HRV features, including time-domain and nonlinear metrics, were extracted from 91-minute windows preceding VF onset. A Gaussian kernel Support Vector Machine (SVM) was trained to classify pre-VF and normal rhythms, with performance assessed via leave-one-out cross-validation. Alarm generation was based on a 10-minute moving average of SVM outputs exceeding a defined threshold. The proposed algorithm achieved 100% sensitivity, 77.8% specificity, and 89.5% accuracy. The average earliest prediction time was 77.0 ( $$\pm 15.9$$ ) minutes, providing substantial lead time for clinical intervention. This study demonstrates the feasibility of continuous ECG-based VF prediction, supporting its potential integration into clinical decision support systems and long-term monitoring platforms. Future work will focus on validation in wearable technologies and clinical workflows.