Deep Learning Models Predict Dynamic Oxygen Uptake Responses from Wearable Sensor Data during Moderate‐ and Heavy‐Intensity Exercise

Abstract

Wearable technologies and artificial intelligence have enabled continuous and non-intrusive cardiorespiratory monitoring. Machine learning techniques, such as random forest (RF) and long short-term memory (LSTM) models, have been used to predict the oxygen uptake (V̇O2) response to exercise, but their abilities to track V̇O2 during changes in work rate lack precision. Here, we propose using a sequential deep learning model based on temporal convolutional networks (TCN) to estimate V̇O2 from wearable sensor data. Twenty-two healthy adults (9 females, age: 26±5 yr, peak V̇O2: 42±6 ml·min-1·kg-1) completed a 25 W·min-1 ramp cycling test to exhaustion, and a combination of 3 different pseudorandom binary sequence (PRBS) cycling tests to simulate non-constant work rate exercise ranging from low to moderate, low to heavy, and ventilatory threshold to heavy-intensity exercise, respectively. Breath-by-breath V̇O2 was measured using a portable metabolic device, and wearable sensor data were simultaneously collected using a Hexoskin® sensor shirt. Work rate, heart rate, percent heart rate reserve, estimated minute ventilation, and breathing rate were used as model inputs to predict instantaneous V̇O2. Participant data were split into 3 groups to train (n=10), validate (n=7), and test (n=5) the newly proposed TCN model and previously reported RF and LSTM models. Repeated measures Bland-Altman analysis (bias; 95% limits of agreement) revealed that the TCN model (-1 ml·min-1; -237 to 235 ml·min-1) was more accurate at estimating the dynamic V̇O2 responses to ramp and PRBS exercise than the LSTM (-10 ml·min-1; -321 to 301 ml·min-1) and RF models (62 ml·min-1; -259 to 383 ml·min-1), despite containing fewer trainable parameters (63% and 97% reduction vs. LSTM and RF, respectively). These results suggest that the TCN model is more accurate and efficient at estimating V̇O2 than other previously used machine learning models across a wide range of exercise intensities. Our findings are an important step in the development of a framework to non-intrusively assess aerobic fitness levels and energy expenditure during real-life situations outside of the laboratory without cumbersome equipment.