Determining Human Activity Through ECG Motion Artifacts

Remote wearable single-lead electrocardiograms (ECGs) can contain motion artifacts (MA), caused by day-to-day human activity, making it difficult to identify QRS complexes and compute R-R intervals.This issue requires the development of complex signal processing techniques or the use of multi-sensor fusion modalities to eliminate MAs which increase the computational power needs of small battery-powered devices.This work proposes an exploration of embedded friendly machine learning (ML) models in the classification of capacitive electrode-based ECGs that contains MAs in the form of different types of human movements.Adult capacitiveelectrode based ECGs (n=5) are obtained under different upper-limb movements.48 statistical features are extracted from the original ECG signals and its wavelet coefficients.The Ensemble-based KNN approach is shown to provide the highest accuracy of 96% and sensitivity of 95% in binary classification between ECGs with movements and ECGs with no activity.While SVM showed the highest accuracy of 68% in multi label classification between unilateral, bilateral, and no activity.This work is a preliminary step in the investigation of explainable AI (XAI) techniques to identify different human activities solely from MA noise information presented in dry electrode-based ECGs.