Abstract: Alterations in joint contact forces (JCFs) are thought to be important mechanisms for the onset and progression of many musculoskeletal and orthopaedic pain disorders. Computational approaches to JCFs assessment represent the only non-invasive means of estimating in-vivo forces; but this cannot be undertaken in free-living environments. Here, we used deep neural networks to train models to predict JCFs, using only joint angles as predictors. Our neural network models were generally able to predict JCFs with errors within published minimal detectable change values. The errors ranged from the lowest value of 0.03 bodyweight (BW) (ankle medial-lateral JCF in walking) to a maximum of 0.65BW (knee VT JCF in running). Interestingly, we also found that over parametrised neural networks by training on longer epochs (>100) resulted in better and smoother waveform predictions. Our methods for predicting JCFs using only joint kinematics hold a lot of promise in allowing clinicians and coaches to continuously monitor tissue loading in free-living environments.
The Challenge of Measuring Joint Loading
Understanding joint contact forces (JCFs) is essential for managing musculoskeletal disorders like osteoarthritis and preventing sports injuries. However, measuring these forces is incredibly difficult. In-vivo measurements require invasive surgical implants, while the standard non-invasive alternative—musculoskeletal modeling—requires expensive laboratory equipment like force plates and optical motion capture.
Because force plates cannot be easily used outside a lab, we lack a way to monitor joint loading in "free-living" environments (e.g., during a marathon or a daily walk). This research investigates whether deep learning can bridge this gap by predicting complex 3D joint forces using only kinematics (joint angles), which are much easier to collect via wearables or video.
Methodology: The XCM Network
The researchers employed a specific deep learning architecture called XCM (eXplainable Convolutional neural network for Multivariate time series).
Unlike standard networks, XCM uses 2D and 1D convolutional filters in parallel. This allows the model to simultaneously extract spatial features across different joints and temporal features across the duration of the gait cycle. The models were fed 3D joint angles from the hip and knee, plus sagittal ankle angles, to predict 3D JCFs for the ankle, knee, and hip.
A key focus of the study was the effect of training duration. While machine learning practitioners often use "early stopping" to prevent overfitting, the authors tested training for up to 1,000 epochs to investigate the impact of "over-parameterization" on the quality of the predicted force waveforms.
Key Insights: "Double Descent" and Smoothness
Traditionally, training a model for too many cycles leads to "wiggly" predictions as the model starts memorizing noise. However, the authors found that training for more than 100 epochs actually made the waveforms smoother and more physiologically realistic. This aligns with the "double descent" phenomenon in modern AI, where very large models or long training times eventually lead to better generalization after an initial period of overfitting.
While numerical error metrics (RMSE) are important, the visual quality of the waveform is critical for biomechanical interpretation. Over-training might only slightly lower the mathematical error, but it significantly improves the physiological plausibility of the predicted force, making it more useful for clinical assessment.
Results and Performance
The study proves that joint angles alone contain enough information to estimate internal forces within a clinically acceptable margin of error. Prediction errors were generally within the "Minimal Detectable Change" (MDC) values of traditional modeling.
- Accuracy: The lowest error was 0.03 BW (ankle medial-lateral force in walking), while the highest was 0.65 BW (knee vertical force in running).
- Correlation: The average correlation between predicted and actual knee JCFs was 0.87, which is comparable to much more complex modeling approaches.
- Training Impact: Increasing training from 25 to 50 epochs saw the biggest jump in accuracy. For walking, 1,000 epochs yielded the best results, whereas running models peaked around 200 epochs.
- Gait Specificity: A model trained on walking is poor at predicting running forces (and vice versa), suggesting that the neural network is learning task-specific biomechanical strategies.
Limitations and Future Work
While promising, the models were trained to mimic musculoskeletal simulations (the "gold standard"), not actual in-vivo sensor data. If the simulation has inherent errors, the AI will learn them as well. To truly move into "free-living" environments, these models need to be tested using data from wearable Inertial Measurement Units (IMUs) rather than high-precision optical cameras. Future work should also explore larger datasets to account for the full diversity of human movement patterns and pathological gaits.
Summary updated