Introduction
Clinical Gait Analysis (CGA) is a cornerstone in the treatment of pediatric cerebral palsy, providing clinicians with critical insights into a patient's mobility. Traditionally, this analysis relies on two disparate data types: static data, consisting of physical measurements like bone alignment and muscle strength taken at rest, and dynamic data, which are time-series curves representing body movement during walking trials. Currently, the correlation between these datasets is established through qualitative expert intuition, a process prone to subjectivity. This paper introduces the Correlated Static-Dynamic Model (CSDM), a quantitative framework designed to automate and explain the link between physical impairments and gait abnormalities.
Methodology
The CSDM is a generative probabilistic graphical model that treats gait patterns as latent variables. It assumes that a patient's static profile ($y$) determines a hidden gait pattern ($h$), which in turn influences the sequence of phase states ($z$) and the observed dynamic gait curves ($x$).
To bridge the gap between messy, mixed-type static data and latent patterns, the model employs Probability Estimation Trees (PETs). Unlike black-box models, PETs offer high interpretability, allowing clinicians to trace specific anatomical thresholds to predicted movement patterns. The model parameters are learned using an Expectation-Maximization (EM) algorithm, where the E-step infers hidden states via a modified forward-backward algorithm, and the M-step updates the decision tree structure and Gaussian distributions of the movement curves.
Results and Discussion
The model was validated using both synthetic data and a real-world dataset of 99 patients from the Royal Children’s Hospital in Melbourne. The CSDM significantly outperformed standard Hidden Markov Models (HMMs) by leveraging static data as a prior, resulting in higher log-likelihood scores.
As shown in the figure above, the model successfully identified four distinct gait patterns. Pattern 4 was categorized as "near-normal," while Pattern 1 represented highly abnormal movement. Crucially, the generated decision trees aligned with clinical logic, identifying specific anatomical markers—such as internal hip rotation thresholds—that correlate with surgical-level gait deviations. This joint modeling approach ensures that the timing of abnormalities (e.g., swing vs. stance phase) is preserved and contextualized by the patient's physical exam.
Conclusion
The CSDM represents a significant step toward automated, interpretable clinical decision support in gait analysis. While this pilot focused on hip rotation, future work aims to expand the model to multi-dimensional analysis across all lower-limb joints. Additionally, incorporating semi-supervised learning could further refine the model's accuracy by integrating expert-labeled "gold standard" gait patterns.
Summary updated