Related Experiment Video
Updated: Jan 8, 2026
Oscillation and Reaction Board Techniques for Estimating Inertial Properties of a Below-knee Prosthesis
Published on: May 8, 2014
Artificial intelligence in lower limb joint moment prediction during typically developed gait: A systematic review
Jianqi Pan1, Zixiang Gao2, Zhanyi Zhou3
1Faculty of Sports Science, Ningbo University, Ningbo, China.
Objective:
Artificial intelligence (AI) methods have been widely applied in gait analysis, yet quantitative comparisons across models and their input-output specifications remain limited. This study aims to systematically review and synthesize the existing literature to evaluate the effectiveness of AI methods in predicting lower limb joint moments during typically developed (TD) gait.
Methods:
Relevant studies published before July 1, 2025, were retrieved from five databases (PubMed, Scopus, IEEE Xplore, ScienceDirect, and Web of Science) using Boolean logic operations and were screened according to predefined criteria. Risk of bias and applicability were assessed with PROBAST. Meta-analyses were performed in R using a multilevel random-effects model to examine differences in predictive performance across AI model group, signal input type, and output joints.
Results:
Eleven studies involving 371 TD participants met the inclusion criteria. Deep neural networks (DNN) showed the best performance for R2 (0.88, 95 %CI 0.52-1.24), while traditional machine learning (ML) models demonstrated relative superiority for nRMSE (0.11, 95 %CI 0.06-0.29). Among input types, surface EMG (sEMG) achieved the highest R2 (0.96, 95 %CI 0.04-1.89), whereas all inputs except "kinematic and speed and anthropometrics" performed well in the nRMSE analysis. For output joints, the ankle was significantly superior to both the knee (p < 0.001) and the hip (p < 0.001) in terms of R2 and nRMSE.
Conclusion:
AI methods can effectively predict lower limb joint moments during TD gait, but differences exist across model group, input type, and output joints. DNN show advantages in fitting complex data, while traditional ML demonstrates greater robustness in small-sample settings. The sEMG, as a process-related input, exhibits high potential, and predictions for the ankle joint are generally superior. Future studies should expand sample size, explore multimodal inputs and advanced modeling strategies, and further validate the applicability of AI methods in pathological gait.
