Motion-Resilient Deep Learning for ECG Signal Classification: A Multi-Electrode Approach Using Adaptive Feature Extraction

Wearable ECG monitoring systems are finding applications in continuous health surveillance. However, motion artifacts typically destroy the signal’s integrity and degrade diagnostic accuracy. In this research, we present a newly developed deep-learning framework that mitigates motion-induced signal degradation by exploiting multi-electrode spatial redundancy and adaptive feature extraction techniques. In this work, a convolutional recurrent neural network is proposed to classify four electrode configurations over different physical activities, as recorded in the MACECGDB database. A new type of multistage feature extraction strategy has been developed that integrates time-domain, frequency-domain, and motion-correlated features within a dynamic reconstruction of ECG signals using the proposed model with enhanced noise robustness. Experimental results have confirmed accuracy in the classification of standing, walking, and jumping conditions well above the traditionally used methods of artifact reduction. This work further enhances the reliability of ambulatory ECG monitoring by integrating machine learning into sophisticated signal processing techniques to provide a robust solution for real-world physiological signal analysis in mobile health applications.