Deep learning to estimate power output from breathing

Abstract

Activity tracking devices are widely used for monitoring and measuring different aspects of physical activity, such as heart rate, speed, energy expenditure and power output. Breathing variables like ventilation and oxygen uptake have also been used for estimating physical effort during exercise. The link between exercise intensity and the muscles’ increased need for oxygen makes breathing a universally applicable metric across many activity forms. Breathing can be measured by standard equipment such as exercise spirometers, but they are impractical to use in normal exercise situations, because they cover the face or mouth and often require stationary recording equipment. Respiratory inductive plethysmography (RIP) is a method for measuring the movement of the rib cage and abdomen caused by breathing, and it enables us to have a portable, non- invasive way of recording breathing during exercise. Power output is a direct measurement of the work performed by a person doing physical exercise, and is a way of expressing the person’s energy expenditure. This thesis studies whether we can use RIP signals to create predictive models for estimating power output during exercise. An N-of-1 study was performed on a highly active adult male of age 25, who performed 21 workouts on a stationary bike. RIP signals from the rib cage and abdomen, in addition to heart rate and power output were recorded during the workouts. The data acquired was used to build predictive models, by using three different deep learning methods: Dense neural networks (DNN), convolutional neural networks (CNN) and long short-term memory (LSTM) networks. A person’s breathing pattern may have characteristics that differ from that of another person, which is a reason for building personalized models. A CNN trained on a combination of features derived from RIP signals and heart rate obtained a mean absolute percentage error (MAPE) of 0.20, which means a 20% error on average with respect to the ground truth. The model’s ability to give precise predictions during workouts with large fluctuations in power output significantly outperformed the DNN and LSTM models tested in this study. We conclude that deep learning techniques can be used for creating personalized models that estimate power output from RIP signals.