ECG-Based Cardiac Arrhythmia Detection Using a Lightweight Attention-Augmented CNN-BiLSTM Network
TL;DR
A lightweight Convolutional Neural Network and Bidirectional Long-Term Memory (CNN-BiLSTM) architecture is presented to help identify multi-class ECG arrhythmias and demonstrates clinically measurable QRS activation, paving the way for wearable monitoring platforms with interpretable and deployable cardiac AI.
Abstract
The use of automated electrocardiogram (ECG) analysis in clinical settings with limited resources has made cardiac arrhythmias a significant risk factor for sudden cardiac death and stroke. Despite their high accuracy, existing deep learning models are not as practical due to computational overhead or insufficient temporal interpretation. A lightweight Convolutional Neural Network and Bidirectional Long-Term Memory (CNN-BiLSTM) architecture is presented in this paper to help identify multi-class ECG arrhythmias. The proposed model employs a coarse-to-fine morphological feature hierarchy, with three progressive convolutional blocks with filter counts of 32, 64, and 128 and a BiLSTM encoder with hidden dimension 128 capturing full bidirectional temporal context across the heartbeat sequence. The soft temporal attention layer focuses representational mass on waveform segments of utmost diagnostic importance, such as the QRS complex and P-wave. The MIT-BIH Arrhythmia Database was used under the AAMI-EC57 five-class protocol (N, S, V, F, Q); the results show 98.21% accuracy, 97.68% macro-averaged sensitivity, and 98.74% macro-averaged specificity, outperforming CNN-only (93.42%), LSTM-only (91.5%), and unidirectional CNN-LSTM (96.31%) baselines. An ablation study quantifies the independent contributions of bidirectionality (+1.14%) and attention (+0.76%). Attention weight visualisation demonstrates clinically measurable QRS activation, paving the way for wearable monitoring platforms with interpretable and deployable cardiac AI.