ECG-Based Arrhythmia Analysis and Short-Term Ventricular Event Prediction Using Deep Learning

Early prediction and characterization of ventricular arrhythmias are important for reducing the risk of sudden cardiac death, a leading cause of mortality worldwide. Conventional ECG analysis commonly relies on full waveform processing, which can be computationally intensive and may provide limited interpretability at the individual-beat level. In this study, we propose a lightweight Hybrid CNN framework for beat-level ECG analysis and short-term prediction of ventricular arrhythmic events. ECG signals are segmented into individual heartbeats, and eleven statistical, temporal, and morphological features, namely the mean, standard deviation, maximum, minimum, RR interval, RR ratio, QRS amplitude, QRS width, energy, skewness, and kurtosis, are extracted from each beat. At each prediction time point, a sequence containing up to 300 consecutive heartbeats is organized as the input to the proposed Hybrid CNN, with a minimum sequence length of 20 beats. The primary task is binary prediction of the occurrence of a ventricular arrhythmic event within the subsequent five-minute window, where a ventricular event is defined by a “V” annotation in the MIT-BIH Arrhythmia Database. To assess the reliability and reproducibility of the proposed framework, the experimental analysis includes annotation-parser validation, explicit definition of ventricular event labels, automatic R-peak detection, feature ablation analysis, quantitative explainability, and confusion-matrix-based performance evaluation. The current implementation uses offline preprocessing, and its suitability for strictly causal real-time prediction requires further investigation. Experiments conducted on the MIT-BIH Arrhythmia Database demonstrate that the proposed Hybrid CNN achieves a sensitivity of 90.22% and a specificity of 89.55% for short-term ventricular event prediction. The ablation further evaluates the contribution and temporal robustness of the proposed beat-level representation. Overall, the proposed framework provides an integrated and lightweight solution for beat-level ECG representation, arrhythmia characterization, and short-term ventricular event prediction, with potential applicability to real-time wearable and edge-AI ECG monitoring systems.