Deep Learning for Automated Electrocardiogram Analysis: Real-Time Arrhythmia Classification Using 1D-CNN

Abstract

Cardiac arrhythmias serve as important biomarkers of cardiovascular disease and therefore require continuous, accurate and timely monitoring to enable clinical intervention. Ambulatory electrocardiogram (ECG) recordings produce large volumes of data, making manual review for diagnosis labor-intensive and subject to fatigue-induced error. Currently, automated deep learning solutions are often computationally expensive or fail due to severe class imbalance common in clinical datasets missing detection of rare but clinically important pathological beats. A lightweight one-dimensional Convolutional Neural Network (1D-CNN) for five-class arrhythmia classification was introduced in this study. For extreme skew in the epidemiological data, we create a hybrid resampling pipeline with additional synthetic Gaussian noise augmentation (σ=0.5) to apply at training time that regularizes decision boundaries against clinical noise. The three-stage convolutional model was tested on the independent unseen testing partition of the publicly available benchmark PhysioNet MIT-BIH Arrhythmia Database, which consists of 21,892 unaugmented cardiac beats. The framework demonstrated a total diagnostic test accuracy of 95.5%, with strong macro-averaged sensitivity across all AAMI categories, achieving 88.87, 92.05, and 88.27% for supraventricular ectopic, ventricular ectopic, and fusion beats respectively the proposed model achieves sub-millisecond inference latency while being deployed in real-time on wearable edge telemetry platforms by reducing parameters to 118,341. This method is clinically feasible and computationally efficient for continuous cardiac health monitoring.