Foundation model for gait kinetics estimation across neurological and developmental cohorts from flexible kinematic inputs
PathoGaitFM is a foundation model that accurately estimates gait kinetics across diverse neurological and developmental cohorts using flexible kinematic inputs, demonstrating high generalizability and robustness without requiring retraining for specific populations or measurement protocols.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Walking is a complex mechanical act that most of us perform without a second thought, but for doctors treating conditions like cerebral palsy, stroke, or Parkinson's disease, understanding the precise forces involved in every step is crucial. To plan surgeries or tailor rehabilitation, clinicians need to know not just how a person moves, but how much force their joints are absorbing and generating. Traditionally, measuring these internal forces has been a difficult task reserved for specialized laboratories. It requires expensive equipment, such as force-sensitive floors and cameras that track reflective markers placed on the body, along with a team of experts to process the data. This setup is so cumbersome that it is rarely available in regular clinics, community centers, or at home, leaving a significant gap in how doctors monitor and treat patients outside of a research setting.
A team of researchers has developed a new approach that aims to bridge this gap using a type of artificial intelligence called a foundation model. Instead of relying on force plates, this system, named PathoGaitFM, can estimate the forces acting on the body using only the angles of the joints as they move. The researchers trained this model on a vast amount of data from healthy adults, teaching it the fundamental physics of walking. They then tested whether this single, shared model could accurately predict forces for people with different neurological conditions, including children with cerebral palsy, adults who have had a stroke, and those with Parkinson's disease. The goal was to see if one universal model could understand the mechanics of walking across very different types of bodies and movement patterns without needing to be retrained for each specific group.
The results show that the model is remarkably effective at this task. When tested on data from four different groups of patients, the system successfully reconstructed the force patterns of walking with a high degree of accuracy, matching the reference measurements with a correlation score of 0.811. This performance was slightly better than other existing methods that had been specifically trained for each individual group. More importantly, the model demonstrated a strong ability to generalize. When the researchers applied the same model to five completely new datasets from different laboratories and populations—without making any adjustments or retraining it—the accuracy remained high, with a correlation score of 0.898. This suggests that the model has learned a deep, shared understanding of how walking mechanics work, rather than just memorizing the specific details of the people it was trained on.
Another key finding is the model's flexibility regarding the data it receives. In real-world scenarios, sensors might fail, or a patient might not be able to wear markers on every joint. The researchers tested whether the model could still work if certain joint angles were missing. They found that even when they removed information about the hips, knees, or ankles, the model's accuracy dropped only slightly, by a small margin of at most 0.041 in its correlation score. This indicates that the system can adapt to incomplete information, a vital feature for clinical use where perfect data is not always available. Furthermore, the model performed well when fed data from video cameras and wearable motion sensors, achieving results comparable to those obtained from the traditional, marker-based systems.
The study also compared this new approach against other established methods. The researchers found that their foundation model outperformed three other published models that had been retrained specifically for the same groups of patients. This suggests that the strategy of pre-training on a large dataset of healthy walking and then using a technique to fill in missing data is more powerful than building separate models for each specific condition. The model was able to estimate forces for the hip, knee, and ankle, as well as the vertical force pushing up from the ground, across a wide range of ages and conditions. While the errors were slightly larger for patients with more severe movement impairments, the estimates were still within a range that is comparable to the natural variation seen when a person walks multiple times in a lab.
This work points toward a future where detailed biomechanical analysis is no longer confined to the laboratory. By demonstrating that a single model can accurately estimate forces from flexible inputs across diverse populations, the researchers have provided a strong basis for using portable technology in routine care. The findings suggest that doctors could soon use video recordings or simple wearable sensors to gain insights into joint loading and movement mechanics, making advanced gait analysis accessible for outpatient clinics and home-based rehabilitation. While the model does not replace the precision of force plates for research purposes, it offers a promising alternative for settings where such equipment is logistically impossible to use, potentially transforming how movement disorders are monitored and treated in everyday practice.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.