Efficient Single-Lead ECG Analysis System with Multi-Lead Reconstruction

Although the electrocardiogram (ECG) is essential for diagnosing heart disease and monitoring the condition of cardiac patients in real time, most wearable devices can only detect the ECG signal from the wrist. Therefore, reconstructing the entire 12–lead signal from the single wrist lead is essential for quickly identifying the presence of possible damage to the myocardium or conduction system, monitoring drug–induced effects, or verifying the correct functioning of a pacemaker. In this paper, we introduce a deep learning framework capable of simultaneously reconstructing the complete ECG signal and classifying it into normal or pathological. Specifically, different neural network architectures — 1D convolutional neural networks, WaveNet, and state space models, such as Mamba—were trained and evaluated on the public benchmark PTB–XL, with WaveNet consistently outperforming other alternatives and achieving the lowest mean squared error (MSE) for signal reconstruction, and the highest F1–score for classification. Furthermore, to make the model applicable to resource–constrained wearable devices, we explored the use of 8–bit integer quantization on the best network configuration, achieving 80% reduction in model size with negligible performance loss. Preliminary experimental results achieve state–of–the–art performance on PTB–XL, highlighting the practical utility of the proposed model in performing comprehensive ECG analysis directly from wearable devices, for robust and efficient cardiac monitoring in real–world scenarios.