Quantifying clinical reasoning in Functional Ambulation Category–based gait assessment using interpretable graph modeling of muscle synergies

Background The Functional Ambulation Category (FAC) is a widely used clinical scale for assessing walking ability in individuals with stroke. However, the neuromechanical coordination patterns that underpin FAC scoring remain poorly understood. This study aimed to quantify the implicit clinical reasoning embedded in FAC-based gait assessments by modeling inter-synergy coordination derived from electromyography (EMG) data. Methods We analyzed EMG data acquired during walking from 102 neurologically intact individuals and 43 individuals with stroke. Bilateral EMG signals from seven lower-limb and trunk muscles were preprocessed and time-normalized over a single gait cycle. Muscle synergies were extracted using non-negative matrix factorization. Time-varying activation profiles of individual synergies were represented as fully connected graphs and used to train a graph neural network (GNN) to classify FAC levels. Model performance was evaluated using stratified five-fold cross-validation with macro area under the curve (AUC), F1-score, and accuracy. Integrated Gradients (IG), an explainable artificial intelligence method, was applied to identify inter-synergy connections most influential for model predictions, enabling quantitative interpretation of clinical reasoning. Group comparisons were performed using non-parametric statistical tests with false discovery rate correction. Results The proposed graph-based model achieved classification performance comparable to conventional approaches, with improved discrimination in intermediate FAC levels. The model yielded a macro AUC of 0.935 (95% CI, 0.891–0.970), an F1-score of 0.770 (95% CI, 0.667–0.885), and an overall accuracy of 0.891 (95% CI, 0.848–0.945). IG analysis revealed that FAC 2 and FAC 3 groups exhibited increased self-connectivity within synergies associated with stance-phase propulsion (FAC 2: \(\:p=0.012,\:r=-0.28\); FAC 3: \(\:p=0.001,\:r=-0.33\)) and decreased self-connectivity within synergies related to swing-phase limb deceleration (FAC 2: \(\:p=0.030,\:r=0.23\); FAC 3: \(\:p=0.013,\:r=0.26\)). In addition, FAC 2 showed reduced coordination between synergies associated with foot clearance and limb deceleration (\(\:p=0.017,\:r=0.26\)), indicating impaired intra-phase integration. These coordination patterns quantitatively captured neuromuscular features underlying clinical gait assessment. Conclusions These findings suggest that implicit clinical reasoning in FAC-based gait assessment reflects structured and quantifiable coordination patterns among neuromuscular synergies. By making these latent patterns explicit, the proposed framework enhances interpretability and may support more precise and physiologically grounded gait evaluation in stroke rehabilitation.