Prediction of Gait Speed from Acceleration Based on Long Short-Term Memory

The authors have developed a cane gait-training machine that enables stroke and paraplegic patients to safely rehabilitate on their own. This training machine is pulled by a wire connected to a harness on the patient's waist; thus, the training machine can follow the patient without the use of hands. However, the machine's ability to follow the walker is an issue. Therefore, we motorised the casters and set them to follow the patient's movements. To cope with the transmission, processing, and mechanical delays that occur in this process, and for predictive control, we used long short-term memory, a machine learning method, which predicted the future waist gait speed from the acceleration measured at multiple sites on the body. In this study, we examined the effects on prediction error of varying the combination of acceleration measurement sites used for learning and the prediction horizon, which is the target time to be predicted. Prediction errors under certain conditions enabled the prediction of each subject's average gait speed with an accuracy of 3-5%. Overall, the prediction error increased with longer prediction horizon but temporarily decreased at 0.4 s. Although prediction is possible using only one site on the lower body, we believe that prediction using multiple sites will reduce the error due to noise.