Abstract
Accurate estimation of joint reaction moments (JRMs) and ground reaction moments (GRMs) during gait is critical for diagnosing musculoskeletal disorders but remains constrained by lab-dependent methods requiring force plates or full-body kinematics. This study introduces a novel deep learning framework to predict sagittal plane JRMs and GRMs using only single-marker positional data (hip, knee, ankle), eliminating the need for complex instrumentation. Leveraging sequence-to-sequence models (GRU, LSTM, 1D-CNN) optimized via Bayesian hyperparameter tuning and Leave-One-Subject-Out (LOSO) cross-validation, we analyzed gait data from 80 asymptomatic adults. Models achieved RMSE values of $31.1-50.1 \text{Nmm} / \text{kg}$ across joints, with normalized RMSE (nRMSE) between $\mathbf{1 6} \boldsymbol{-} \mathbf{2 1 {\%}}$, comparable to prior studies using more extensive inputs. GRU excelled at ankle moment prediction $(31.1 \pm 25.8$ vs. $38.6 \pm 30.0$ Nmm/kg, left/right), while LSTM showed robustness for knee moments $(35.0 \pm 25.6$ vs. $37.2 \pm 30.1 \text{Nmm} / \text{kg})$. Asymmetries in hip RMSE $(48.7 \pm 25.7 \mathrm{L}$ vs. $42.4 \pm 23.4 \mathrm{R})$ highlighted joint-specific variability. The framework's minimal kinematic requirements reduce computational burden, offering a scalable alternative to traditional gait analysis. Despite strong performance in asymptomatic individuals, validation in pathological populations and integration with Markerless systems are needed for clinical translation. This work advances accessible, lab-free biomechanical assessment, demonstrating the potential of deep learning to bridge gaps between research and clinical practice.

This publication has 14 references indexed in Scilit: