Interlimb and Intralimb Synergy Modeling for Lower Limb Assistive Devices: Modeling Methods and Feature Selection

Abstract

The concept of gait synergy provides novel human–machine interfaces and has been applied to the control of lower limb assistive devices, such as powered prostheses and exoskeletons. Specifically, on the basis of gait synergy, the assistive device can generate/predict the appropriate reference trajectories precisely for the affected or missing parts from the motions of sound parts of the patients. Optimal modeling for gait synergy methods that involves optimal combinations of features (inputs) is required to achieve synergic trajectories that improve human–machine interaction. However, previous studies lack thorough discussions on the optimal methods for synergy modeling. In addition, feature selection (FS) that is crucial for reducing data dimensionality and improving modeling quality has often been neglected in previous studies. Here, we comprehensively investigated modeling methods and FS using 4 up-to-date neural networks: sequence-to-sequence (Seq2Seq), long short-term memory (LSTM), recurrent neural network (RNN), and gated recurrent unit (GRU). We also conducted complete FS using 3 commonly used methods: random forest, information gain, and Pearson correlation. Our findings reveal that Seq2Seq (mean absolute error: 0.404° and 0.596°, respectively) outperforms LSTM, RNN, and GRU for both interlimb and intralimb synergy modeling. Furthermore, FS is proven to significantly improve Seq2Seq’s modeling performance ( P < 0.05). FS-Seq2Seq even outperforms methods used in existing studies. Therefore, we propose FS-Seq2Seq as a 2-stage strategy for gait synergy modeling in lower limb assistive devices with the aim of achieving synergic and user-adaptive trajectories that improve human–machine interactions.