Online Adaptive and LSTM-Based Trajectory Generation of Lower Limb Exoskeletons for Stroke Rehabilitation

Abstract

Lower Limb Exoskeletons (LLEs) are promising in stroke rehabilitation, but the challenge is how to design an adaptive and appropriate trajectory for each stroke survivor to encourage active engagement. To achieve this, online adaptive trajectory generation based on synergies is proposed. In neurology, a gait involves not only the movement of lower limbs but also the rhythmic interjoint coordination (i.e., synergies) among different limbs. Studies also showed the promising applications of synergies in stroke rehabilitation. In this paper, Long Short-Term Memory (LSTM) network is adopted for the first time to interpret and exploit inter-limb synergy for trajectory generation of rehabilitative LLEs. The reference trajectory is generated online for the leg of the paretic side of stroke patients based on the motion data of their upper and lower limbs by LSTM-based synergy extracted from healthy people. Gait experiments on healthy subjects are conducted using a wearable motion capture system to get motion data. One side's hip and knee angle data of a randomly selected subject are estimated, based on the other side's motion data by an LSTM model trained by motion data of other healthy subjects. The estimation results are compared with estimation based on other methods. Results indicate that LSTM has better estimation performance and stability over statistical regression methods such as PCA, which has been widely adopted to analyze human motion synergy. In addition, LSTM shows better inter-individual adaption. The feasibility of the proposed trajectory generation based on LSTM has been validated, although the therapeutic effects or possible benefits of applying synergies into rehabilitation need further exploration.