Adaptive ECG Leads Selection for Low-Power ECG Monitoring Systems Using Multi-class Classification

The computer-aided interpretation of ECG signals has become a pivotal tool for physicians in the clinical assessment of cardiovascular diseases during the last decade. Therefore, computerized diagnosis systems depend heavily on machine learning and deep learning models to guarantee high classification accuracy. However, a large amount of power is consumed due to the need for heavy computations to handle the classification tasks which act as a barrier to maintain continuous ECG monitoring. Hence, this work targets energy saving in the constrained embedded environment on a Texas Instruments CC2650 Micro-controller Unit (MCU). We provide a new approach to support energy-efficient ECG monitoring in real-time through the adaptive selection of ECG leads after applying multi-class classification on the raw ECG signals. We deploy two different CNN model scenarios on MIT-BIH and CODE-test datasets, and adjust the number of ECG streamed channels to 1,4, and 8, based on the detected cardiac abnormalities, such as arrhythmias and heart blocks. The adaptive selection of ECG channels achieves 77.7% power saving in the normal cardiac condition and up to 55.5% for the heart blocks, sinus bradycardia, and sinus tachycardia.