Cost-Effective Prediction of Knee Joint Angle fromSurface Electromyography Signals During FreeMotion using Transformer Regression

This paper outlines a proof-of-concept for the precise forecasting, using a Time-Series Transformer (TST) regression framework, of the angle of the knee joint based on Surface Electromyography (sEMG) data from just three sensors on the upper thigh. To facilitate real-time prediction, the Transformer was given a hybrid input of autoregressive and exogenous data from two Inertial Measurement Units (IMUs) and the three sEMG sensors. Furthermore, the flexibility of a Transformer Model meant that the data could be continuously recorded during spontaneous, unplanned motion of the lower leg. A sliding-window approach was used to allow the model to continuously predict from a stream of live data. The model achieved high cross-validation accuracy (70-15-15 split) predicting the knee joint angle (in degrees) on a relatively small dataset of 35,613 samples--RMSE=3.97, nRMSE=0.0208, MAE=2.61, MBE=-0.45, R2=0.9928 on unseen data--and outperformed benchmarks set by state-of-the-art methods, such as CNN and LSTM, while making use of fewer sensors and demonstrating sub-millisecond prediction time per sample. Additionally, the limited number of sensors and equipment used in this method allows for greater accessibility to patients in need.本文阐述了一项概念验证研究：利用时序Transformer（Time-Series Transformer, TST）回归框架，仅基于大腿上部3个表面肌电（Surface Electromyography, sEMG）传感器采集的数据，实现膝关节角度的精准预测。为满足实时预测需求，该模型采用自回归数据与两类惯性测量单元（Inertial Measurement Units, IMUs）及3个sEMG传感器采集的外源性数据的混合输入方式。此外，Transformer模型的灵活性使得研究人员可在小腿自发、无计划运动过程中持续采集数据。本研究采用滑动窗口方法，使模型能够基于实时数据流开展连续预测。在包含35613个样本的小型数据集上，该模型针对膝关节角度（单位：度）的预测采用70-15-15的数据集划分方式进行交叉验证，在未知测试数据上取得了优异性能：均方根误差（RMSE）为3.97，归一化均方根误差（nRMSE）为0.0208，平均绝对误差（MAE）为2.61，平均偏差（MBE）为-0.45，决定系数（R²）为0.9928。该模型不仅优于卷积神经网络（Convolutional Neural Network, CNN）、长短期记忆网络（Long Short-Term Memory, LSTM）等当前前沿方法所设定的基准性能，还仅使用更少的传感器，且单样本预测耗时低至亚毫秒级。此外，本方法所用传感器与设备数量有限，能够为有需求的患者提供更高的可及性。