Estimating Steady-State Metabolic Energy Cost From Wearable Sensors Using LSTM Neural Networks During Level and Inclined Walking

To reduce the physical effort of walking, lower limb exoskeletons have been developed to assist with human movement. Recent studies on human-in-the-loop optimization have greatly improved exoskeleton performance, allowing for personalized assistance strategies for wearers. Currently, metabolic cost is widely employed as the primary physiological objective in optimization processes. However, estimating metabolic cost typically involves noisy, sparsely sampled data and requires extended experimental durations, which can result in participant fatigue and cardiopulmonary drift. This study aimed to predict energy expenditure based on physiological data collected from wearable sensors, offering improved temporal resolution and reduced variability compared to traditional respiratory measurements. Physiological data, including heart rate, electromyography (EMG), and limb accelerations, were collected from subjects walking at various treadmill speeds and slopes. Energy expenditure was estimated using various machine learning models, including multiple linear regression (MLR), BP neural networks (BPNN), and long short-term memory (LSTM) networks. The results indicated that heart rate muscle activities have a strong correlation with metabolic cost. Among the three models, the LSTM model demonstrated the highest accuracy in predicting energy expenditure. Our results offer valuable insights for developing personalized exoskeleton assistance strategies.