Do deep learning models learn beyond average gait? A comparison of architectures, kinematic representations, and marker sparsity for vertical ground reaction force prediction

Abstract

Background: Vertical ground reaction force (vGRF) is a critical biomarker of joint loading and functional recovery, yet its measurement typically requires force plates in gait laboratories, limiting use outside specialized facilities. Deep Learning (DL) models can estimate vGRF from more accessible kinematic data, but the diversity of approaches and evaluation protocols makes it difficult to identify what truly drives performance. Methods: In this study, we evaluated three popular DL models (LSTM, BiLSTM, and Transformer) for stance-phase vGRF prediction using two kinematic input representations (3D marker trajectories and joint angles), and we additionally tested marker-set reduction down to three markers and a single ankle marker. To assess whether models learn beyond the typical healthy stance profile, we benchmarked all methods against a training-set mean-waveform baseline. Results: Across input modalities and architectures, prediction accuracy was high (RMSE <10% BW and R2 >0.9). However, only the Transformer significantly outperformed the baseline, whereas LSTM and BiLSTM did not improve beyond this reference, despite similarly good absolute metrics. Marker trajectories and joint angles yielded comparable accuracy, indicating that performance is robust to the chosen kinematic representation. Finally, marker-set reduction for the Transformer showed no significant degradation when using only three markers, whereas an ankle-only input no longer outperformed the baseline. Conclusions: Transformer-based vGRF estimation can remain feasible with sparse kinematic inputs, supporting simplified out-of-lab gait analysis. Reporting an explicit mean-waveform baseline is recommended to avoid over-interpreting high absolute metrics in healthy gait, where waveform variability is limited.