An Open Access Dataset for Supervised Machine Learning to Estimate Gait Biomechanical Characteristics

There are limited gait datasets in the literature due to the complexity and accessibility of the motion capture system. Moreover, all of them only comprises of young adult datasets. This paper presents a comprehensive dataset of human gait that contains the kinematics and muscle activities of lower limb captured using Electromyography (EMG), and Inertial Measurement Unit (IMU). The data were collected from 65 male and female participants with ages ranging from 19 years old to 73 years old. A case study that utilizes the dataset and the supervised machine learning models i.e. Feedforward Neural Network (FNN) and Long Short-term Memory (LSTM) network is proposed to demonstrate the feasibility of the dataset to estimate the dynamics of human gait, particularly the lower extremity muscle activity. The models were tested on an unseen dataset and other online dataset. The results showed that LSTM outperformed FNN. The LSTM models achieved root mean square error (RMSE) below 11%, correlation coefficient (r) above 90%, and peak timing differences below 10% when predicting EMG signals in test dataset. The dataset is expected to accelerate the adoption of supervised machine learning in clinical and rehabilitation settings, particularly gait analysis.

受运动捕捉系统（motion capture system）的复杂性与可获取性限制，现有文献中的步态数据集数量有限，且全部仅涵盖青年成人样本。本文提出一款全面的人体步态数据集，该数据集包含通过肌电图（Electromyography, EMG）与惯性测量单元（Inertial Measurement Unit, IMU）采集的下肢运动学参数与肌肉活动信号。本数据集的采集对象为65名男女受试者，年龄跨度为19岁至73岁。本文还提出一项案例研究，借助本数据集与两类监督机器学习模型——前馈神经网络（Feedforward Neural Network, FNN）及长短期记忆网络（Long Short-term Memory, LSTM），验证该数据集用于估算人体步态动力学（尤其是下肢肌肉活动）的可行性。所提模型在预留测试集与另一公开在线数据集上完成了测试验证。实验结果显示，LSTM模型的表现优于FNN。在测试集上预测肌电信号时，LSTM模型的均方根误差（Root Mean Square Error, RMSE）低于11%，相关系数（correlation coefficient, r）高于90%，峰值时序偏差低于10%。本数据集有望推动监督机器学习技术在临床与康复场景（尤其是步态分析领域）的普及应用。