Sugiarto Tommy, Lin Yi-Jia, Tsai Hsiao-Liang, Sun Chi-Tien, Hsu Wei-Chun
Graduate Institute of Applied Science and Technology, National Taiwan University of Science and Technology, Taipei, Taiwan.
Graduate Institute of A.I. Cross-disciplinary Technology, National Taiwan University of Science and Technology, Taipei, Taiwan.
Biomed Eng Online. 2025 Jun 24;24(1):78. doi: 10.1186/s12938-025-01409-1.
Wearable sensors combined with deep-learning models are increasingly being used to predict biomechanical variables. Researchers have focused on either simple neural networks or complex pretrained models with multiple layers. In addition, studies have rarely integrated knee alignment information or the side affected by injury as features to improve model predictions. In this study, we compared the performance of selected model architectures, including complex pretrained models, in predicting three-dimensional (3D) ground reaction force (GRF) data during level walking by using data obtained from motion capture systems and wearable accelerometers.
Ten deep-learning models for predicting the 3D GRF were developed using motion capture and accelerometer data with or without subject-specific features. Incorporating subject-specific features improved prediction accuracy for all models except the long short-term memory (LSTM) model. A two-dimensional (2D)-CNN-LSTM hybrid model achieved the best results. Established models, such as ResNet50 and Inception, performed better when trained with pretrained ImageNet weights and subject-specific features, underscoring the value of pretrained knowledge and subject-specific information for improving accuracy. However, these models did not outperform the custom hybrid models in predicting time-series 3D GRF data, indicating that larger models do not necessarily perform better for time-series applications but do always have greater computational demands.
Incorporating subject-specific features, such as alignment information, enhanced the accuracy of GRF predictions during walking. Complex pretrained models were outperformed by custom hybrid models for time-series 3D GRF prediction during walking. Custom models with lower computational demands and using alignment features are a more efficient and effective choice for applications requiring accurate and resource-efficient predictions.
可穿戴传感器与深度学习模型相结合，越来越多地用于预测生物力学变量。研究人员主要关注简单神经网络或具有多层的复杂预训练模型。此外，很少有研究将膝关节对线信息或受伤侧作为特征纳入，以改善模型预测。在本研究中，我们通过使用从运动捕捉系统和可穿戴加速度计获得的数据，比较了包括复杂预训练模型在内的选定模型架构在预测平地上行走时的三维（3D）地面反作用力（GRF）数据方面的性能。
使用有或没有个体特异性特征的运动捕捉和加速度计数据，开发了10个用于预测3D GRF的深度学习模型。纳入个体特异性特征提高了除长短期记忆（LSTM）模型之外所有模型的预测准确性。二维（2D）-CNN-LSTM混合模型取得了最佳结果。诸如ResNet50和Inception等已建立的模型，在使用预训练的ImageNet权重和个体特异性特征进行训练时表现更好，这突出了预训练知识和个体特异性信息对提高准确性的价值。然而，在预测时间序列3D GRF数据方面，这些模型并未优于定制混合模型，这表明更大的模型不一定在时间序列应用中表现更好，但计算需求总是更大。
纳入诸如对线信息等个体特异性特征，提高了行走过程中GRF预测的准确性。在行走过程中时间序列3D GRF预测方面，定制混合模型优于复杂预训练模型。对于需要准确且资源高效预测的应用，计算需求较低且使用对线特征的定制模型是更高效有效的选择。
