Using Wearable Physiological Sensors to Predict Energy Expenditure

Lower-limb assistive robotic devices are often evaluated by measuring a reduction in the user's energy cost. Using indirect calorimetry to estimate energy cost is poorly suited for real-time estimation and long-term collection. The goal of this study was to use data from wearable sensors to predict energy cost with better temporal resolution and less variability than breath measurements. We collected physiological data (heart rate, electrodermal activity, skin temperature) and mechanical data (EMG, accelerometry) from three healthy subjects walking on a treadmill at various speeds on level ground, inclined, and backwards. Ground truth energy cost was established by averaging steady-state breath measurements. Raw physiological signals correlated well with ground truth energy cost, but raw mechanical signals did not. Correlation of mechanical signals was improved by calculating accelerometer magnitudes and linear envelope EMG signals, and further improved by averaging the signals over several seconds. A multiple linear regression including physiological and mechanical data accurately predicted ground truth energy cost across all subjects and activities tested, with less variability and better temporal resolution than breath measurements. The sensors used in this study were fully portable, and such algorithms could be used to estimate energy cost of users in the real world. This could greatly improve the design, control, and evaluation of lower-limb assistive robotic devices.