Dual-Model Artificial Intelligence Framework Integrating AI-Based Markerless Motion Capture for Dynamic Gait Prediction and Health-Status Classification.

Background/

Objectives

Human gait analysis is essential for identifying biomechanical alterations associated with pathological conditions. However, conventional laboratory systems that combine optical motion capture and force plates remain costly, space-demanding, and difficult to implement in routine or accessible settings. This study proposes a dual-model artificial intelligence framework designed to bridge kinematics and dynamics and subsequently support gait health-status classification from motion data.

Methods

The first model was developed to estimate ground reaction forces (GRF) and center of pressure (CoP) signals from kinematic inputs. Public datasets containing synchronized kinematics and dynamics were used to train and evaluate long short-term memory (LSTM) and one-dimensional convolutional neural network (CNN1D) architectures. A robustness stage further adapted the dynamic prediction model to markerless-like kinematic inputs through domain-adaptation training. The second model was implemented as a multichannel convolutional classifier using normalized GRF/CoP signals and metadata. Three variants were compared: signals only, signals with minimal metadata, and signals with rich metadata. Finally, a bridge block connected both models, and a proof-of-concept deployment was performed using markerless kinematics obtained with Move AI and Blender.

Results

The final dynamic prediction model achieved an overall RMSE of 0.0676, with reconstructed-signal RMSE of 0.0500 and reconstructed contact accuracy of 0.9736. The best classification variant achieved a balanced accuracy of 0.9608, while the minimal-metadata variant was selected for pipeline integration due to its compatibility with accessible data acquisition. In the full pipeline evaluation, the predicted dynamics correctly classified the healthy-control samples. In the Move AI/Blender proof-of-concept, all five healthy participants were classified as healthy controls, with a mean non-pathological classification probability consistent with this outcome.

Conclusions

The proposed dual-model framework demonstrates the operational feasibility of linking kinematic acquisition, dynamic prediction, and gait classification within a single artificial intelligence pipeline. The Move AI/Blender stage represents a preliminary proof of concept rather than clinical validation, and further evaluation with pathological participants and synchronized force-plate measurements is required.