Resource-Constrained Device Characterization for Detecting Sleep Apnea Using Machine Learning

Abstract

Automation of measurement and analysis of continuous human body parameters like ECG is a first step towards establishing accessible and distributed medical infrastructure. Currently, the cost of medical devices and use of expertise for analysis puts this out of reach of many patients. Unless their condition becomes life-threatening most patients will avoid going through this process, losing out on the benefits of early detection and treatment of their illness. In this work, we propose the use of cost-effective devices for making a complete self-contained pipeline which includes measurement of ECG signals, cleaning and pre-processing of signals and use of machine learning techniques to analyse them on the device. We have used as a case study, detection of Sleep Apnea using ECG signals. We compare resource-constrained hardware with varying price and capability ranges to study their effectiveness in detecting Sleep Apnea. We propose the use of an artificial neural network model developed using TensorFlow Lite on resource-constrained devices for detection of Sleep Apnea. We report that the results from resource-constrained devices are comparable to more advanced and expensive devices for detection of Sleep Apnea using ECG signals.