A Lightweight Channel-Attention CNN for Robust Beat-Level Arrhythmia Detection

Reliable beat-level arrhythmia detection from electrocardiogram (ECG) signals is essential for continuous patient monitoring. However, many existing deep learning models remain computationally expensive and are sensitive to severe class imbalance in arrhythmia datasets. In this study, we propose a lightweight one-dimensional convolutional neural network enhanced with a squeeze-and-excitation (SE) channel-attention mechanism for efficient heartbeat classification. The model operates on per-beat ECG feature vectors and incorporates Synthetic Minority Oversampling Technique (SMOTE) applied exclusively to the training data to mitigate class imbalance while preventing evaluation leakage. The architecture consists of two Conv1D layers with batch normalization and SE attention, followed by global average pooling and a compact fully connected classifier, enabling efficient inference with a minimal parameter footprint. The proposed model is evaluated on four publicly available ECG databases: MIT-BIH Arrhythmia, INCART-2, Sudden Cardiac Death Holter, and MIT-BIH Supraventricular Arrhythmia. Experimental results show accuracies of 97.76%, 98.38%, 97.17%, and 97.40%, respectively, while maintaining strong macro-averaged precision and recall across beat classes. The model contains approximately 48K parameters (0.19 MB) and requires fewer than 0.9 million FLOPs per inference, enabling efficient execution on standard CPUs. Post-training INT8 quantization further reduces the model size by over 60% with negligible accuracy loss. These results demonstrate that the proposed lightweight attention-based CNN provides accurate and computationally efficient arrhythmia detection, making it suitable for real-time ECG monitoring in resource-constrained and wearable healthcare environments.