An Automated Healthcare Framework for Single-Lead ECG Arrhythmia Classification using Lightweight CNN-BiLSTM and Temporal Attention

Abstract

Continuous monitoring of the electrocardiogram on wearable and ambulatory devices has made the automatic recognition of abnormal heartbeats a practical necessity, yet the beats that matter most clinically are precisely the ones that occur least often. This study presents a compact recurrent-convolutional model that classifies single-lead heartbeats into the four Association for the Advancement of Medical Instrumentation categories: normal, supraventricular, ventricular, and fusion. Two ideas drive the design. First, each beat is expanded into a three-channel derivative-domain representation that stacks the raw waveform with its first and second temporal derivatives, making the velocity and curvature of the signal explicit to the network. Second, a small one-dimensional convolutional encoder is followed by a bidirectional long short-term memory layer and an additive temporal-attention stage that weights the diagnostically informative parts of the beat. Class imbalance is countered by combining Borderline-SMOTE oversampling of the training partition with a class-weighted loss. On the MIT-BIH Arrhythmia Database the model attains an accuracy of 99.12 percent and a macro-averaged F1 of 0.927 using only 60,788 trainable parameters, a footprint compatible with microcontroller-class deployment. An ablation study confirms that the derivative channels, the recurrent layer, the attention stage, and the oversampling each contribute measurably, and the result is competitive with recent and substantially larger models.