Gait2Hip-60: Deep Learning Predicts Hip Forces from Walking Data

A new study published on arXiv introduces Gait2Hip-60, a unified deep learning benchmark for predicting hip muscle forces and joint moments from multi-cadence gait kinematics. The researchers collected gait data from 60 healthy adults walking at three metronome-guided cadences, using ten bilateral lower-limb joint angles as inputs. Reference outputs were derived from OpenSim musculoskeletal simulations. They evaluated three sequence models—LSTM, Transformer, and Mamba—under a consistent protocol with subject-level splits, same preprocessing, and identical metrics. The dataset and code are publicly available to encourage further research.

Transformer emerged as the top performer, achieving RMSE of 1.33 N/kg and R² of 0.819 for hip muscle force prediction, and RMSE of 0.11 Nm/kg with R² of 0.862 for hip joint moments. It maintained an advantage across all cadence conditions. In a zero-shot external validation on 9 patients with osteonecrosis of the femoral head (ONFH), Transformer retained moderate predictive ability (force RMSE=1.51 N/kg, R²=0.537; moment RMSE=0.17 Nm/kg, R²=0.569). The results demonstrate the feasibility of estimating hip dynamics directly from gait kinematics without time-consuming musculoskeletal simulation, though broader pathological validation is needed before clinical deployment.