Heart Watch

Abstract

Electrocardiogram (ECG) is an important health monitoring signal that is used in various medical diagnosis, especially identifying potential possibility of heart attacks and strokes. Moreover, many patients are in remote places and in many countries the patients to doctors ration is very poor which calls for a miniature hardware that remotely captures ECG and transmits data to the doctors. However, the exact reproduction of ECG requires high bit rate and thus requires transmitting a compressed set of parameters. Further, sending large volumes of annotated raw data to train diagnostic models also compromises the patients privacy. We design and present a system that generates synthetic ECG signals from clinical data in real-time using a highly minimized set of parameters. The system comprises a nonlinear dynamical model whose parameters are trained in real-time to synthesize a signal which matches clinical data with high accuracy. The parameters of the trained system are then transmitted in each cycle of the ECG wave to reconstruct the original signal using the same model at the medical practitioners’ location. The parameter learning problem is highly complicated as one needs to solve a nonlinear, non-convex dynamic optimization problem, which usually only converges to local optima. To address this issue, we propose a novel two-stage algorithm that automatically chooses an initial set of parameters in the vicinity of the global optimum and then performs stochastic gradient descent iterations. We perform experiments to demonstrate the accuracy and real-time performance of the system. We show that on average our system processes clinical data of one second in 0.68s on a microcontroller, with an RMSE error of 0.0038 the average, and 17 parameters per ECG cycle. Our system is also easy to implement, requires minimal storage i.e. only one ECG cycle at any given time, and does not depend on offline training, unlike existing methods.