MedNet-lite: an interpretable ultra-lightweight 1D-CNN with CBAM for real-time ECG anomaly screening in resource-constrained edge computing environments

Deploying artificial intelligence (AI) in Internet of Things (IoT) healthcare ecosystems requires models that balance predictive reliability, computational efficiency, and intrinsic interpretability under strict edge computing constraints. This study presents MedNet-Lite, an interpretable, ultra-lightweight 1D convolutional neural network (1D-CNN) incorporating a 1D-adapted Convolutional Block Attention Module (CBAM) for real-time electrocardiogram (ECG) anomaly screening. Evaluated on the MIT-BIH Arrhythmia Database following AAMI-compliant beat categorization, MedNet-Lite achieved 98.40 % accuracy, 99.33 % precision, and 98.88 % recall for normal beats, alongside 90.91 % precision and 94.34 % recall for abnormal beats on an independent 20,282-beat test set. It yielded an ROC-AUC of 0.985 and PR-AUC of 0.942. Quantitative evaluation of the embedded attention mechanism demonstrated stable spatial localization, achieving a weighted attention centroid RMSE of 5.84 samples relative to R-peaks and a mean QRS energy concentration of 14.21 %. For edge deployment, computational profiling on CPU and ONNX Runtime environments showed the model contains only 4,222 parameters, occupies 0.0249 MB, and requires 2.43 MFLOPs. It achieves 1.01 ms native CPU inference, 0.34 ms ONNX inference, and 2.28 ms end-to-end latency, with 4.82 MB peak memory consumption. These results demonstrate that MedNet-Lite effectively balances predictive performance, deployment efficiency, and interpretability for resource-constrained monitoring. Crucially, the framework is intended for non-clinical ECG screening rather than clinical diagnosis.