Long Short-Term Memory and Hidden Markov Model for Predicting Gait Through Piezoresistive Sensors in Real-Life Terrains

In wearable robotics, the accurate prediction and classification of gait patterns play a fundamental role in optimizing exoskeleton control systems. By leveraging long short-term memory (LSTM) models, it is possible to predict key parameters of human gait. This study explores the application of LSTM and hidden Markov models (HMMs) using data from piezoresistive sensors to predict gait parameters for exoskeleton control systems, such as the percentage of gait progression, the center of pressure (CoP), heel strike (HS), and toe off (TO) events. Eleven healthy subjects participated in the experimental protocol and performed walking tasks in four different contexts. An inertial sensor and an instrumented insole with piezoresistive sensors were used to measure gait characteristics, including CoP and rate of gait progression, along with gait cycle events. The findings demonstrate promising accuracy in predicting gait parameters and phases under diverse locomotive conditions, with computational costs falling within acceptable limits. The results for CoP and gait progression show a similar trend in root-mean-square error (RMSE) for prediction intervals of 0.05 and 0.10 s, regardless of walking conditions. For prediction intervals greater than or equal to 0.20 s, a sudden increase in error values is observed (15%). Regarding gait event identification, acceptable errors (~0.15 s) are observed even for prediction steps longer than 200 ms for both HS and TO events. The results suggest that the proposed method could be promising in implementing an adaptable exoskeleton control system for various real locomotion conditions and computational demands.