Temporal Forecasting of Sit-to-Stand Motion for Assistive Device Control Based on Muscle Synergy Prediction

Sit-to-stand (STS) is a common and important motion for maintaining the mobility and independence of a person. Assistive devices such as automatic chairs or exoskeletons are expected to assist people experiencing difficulties during STS. To control such assistive devices, it is important to predict STS motion intention with sufficient anticipation time before an actual STS motion begins, because most devices require a certain time before applying sufficient assistive force, and the delay of device activation significantly degrades the quality of assistance and user self-agency. Another challenge is that STS is a non-cyclic motion, which lacks of preceding information, making motion forecasting is more difficult. In this paper, we designed a deep neural network (DNN) structure with multilayer (i.e., multiresolution) long short-term memory (LSTM) that forecasts the temporal muscle synergy of STS motions observed by surface electromyography (EMG). Prediction accuracy and anticipation time were evaluated using STS data from 10 participants, and an average accuracy of 87.97% was achieved for a forecast step (time) of 300 ms. The anticipation time increased with a longer forecast step. In addition, we found that the DNN with multiresolution LSTM outperformed the DNN with a single-resolution LSTM. The prediction performance was evaluated using a real assistive chair that was controlled based on the predicted data. The leg muscles were less activated when the chair was triggered before significant muscle activation began. These results suggest that using our prediction scheme before actual posture changes effectively enables assistive devices to reduce users’ muscle strength required for STS.