Abstract
<title>Abstract</title> <p>Automated cardiac arrhythmia classification from single-lead ambulatory electrocardiogram (ECG) recordings is a fundamental challenge in clinical engineering, with direct implications for wearable monitoring and early diagnosis. This paper presents a two-stage hybrid framework that couples sub band statistical feature extraction using the Daubechies-4 (db4) discrete wavelet transform (DWT) with a compact one-dimensional convolutional neural network (1D-CNN). Raw ECG signals from the MIT-BIH Arrhythmia Database are conditioned through a Butterworth high-pass filter (0.5Hz), a 60Hz IIR notch filter, and z-score normalisation, then segmented by the Pan–Tompkins algorithm into 217-sample beat windows. A five-level db4 DWT produces six sub-bands; four statistical descriptors per sub-band mean, standard deviation, sub-band energy, and Shannon entropy form a compact 24-dimensional feature vector, achieving 9:1 compression over the raw waveform. A lightweight 1D-CNN with fewer than 200K trainable parameters classifies these vectors into five rhythm classes: Normal (N), Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Premature Ventricular Contraction (PVC), and Atrial Premature Contraction (APC). On 3244 held-out test beats under a stratified 80/10/10 intra-patient split, the system achieves 99.08% overall accuracy, 99.41% best validation accuracy, and a mean AUC-ROC of 0.9884, with RBBB attaining perfect sensitivity of 1.0000. The distinct contribution of this work is the use of wavelet sub band statistical summaries rather than raw coefficients or raw waveforms as a structured, interpretable feature space for a lightweight CNN, yielding strong accuracy at a parameter count suitable for embedded deployment.</p>