Abstract P502: Deep Learning-powered Wearable Electrocardiogram Big Data Monitoring for Precision Cardiac Health

Smart health technologies are bringing exciting possibilities to the cardiac healthcare area. Wearable electrocardiogram (ECG) monitoring is expected to establish cardiac big data towards precision cardiac health. However, there are two key obstacles here. Firstly, how to conveniently measure the standard 12-lead ECG in our daily lives is an open question, since the traditional 12-lead ECG is mainly used in clinics or hospitals. The Holter ECG monitor is actually not convenient and comfortable enough for daily and long-term use. The Apple Watch only provides finger-touch-based single lead ECG measurement, neither supporting 12-lead ECG nor continuous tracking. In this study, a long short-term memory neural network-based ECG monitoring system is proposed, which can generate the remaining 9-lead ECG from only 3-lead ECG, offering a very high wearabilty, usability and convenience. Secondly, how to maintain a high ECG quality even when the user has different physical activities is another critical challenge. Usually, the ECG morphology may be contaminated by diverse motions artifacts induced by sensor-to-skin contact variations. This has to be addressed to guarantee the obtained ECG is usable and interpretable. We have introduced bidirectional long short-term memory to deal with these noisy fluctuations, by learning the temporal consistent dynamics among 3-lead ECG. The system has been evaluated on ten human subjects to demonstrate the effectiveness. Compared with the ground truth, the reconstructed 12-lead ECG has a correlation as high as 0.88 and a root mean square error of 0.059 mV, far superior to the traditional linear regression method. The proposed novel monitor is expected to greatly advance precision cardiac health.