Automated gait event detection for exoskeleton-assisted walking using a long short-term memory model with ground reaction force and heel marker data

Traditional gait event detection methods for heel strike and toe-off utilize thresholding with ground reaction force (GRF) or kinematic data, while recent methods tend to use neural networks. However, when subjects' walking behaviors are significantly altered by an assistive walking device, these detection methods tend to fail. Therefore, this paper introduces a new long short-term memory (LSTM)-based model for detecting gait events in subjects walking with a pair of custom ankle exoskeletons. This new model was developed by multiplying the weighted output of two LSTM models, one with GRF data as the input and one with heel marker height as input. The gait events were found using peak detection on the final model output. Compared to other machine learning algorithms, which use roughly 8:1 training-to-testing data ratio, this new model required only a 1:79 training-to-testing data ratio. The algorithm successfully detected over 98% of events within 16ms of manually identified events, which is greater than the 65% to 98% detection rate of previous LSTM algorithms. The high robustness and low training requirements of the model makes it an excellent tool for automated gait event detection for both exoskeleton-assisted and unassisted walking of healthy human subjects.