Prediction of gait subtypes in children with bilateral spastic cerebral palsy using Edinburgh Visual Gait Score

Early and accurate gait subtype classification in children with spastic bilateral cerebral palsy (CP) is essential for guiding rehabilitation. While three-dimensional gait analysis is limited by cost and accessibility, the Edinburgh Visual Gait Score (EVGS) provides a feasible observational tool. This study evaluates the discriminative value of EVGS-derived features using machine learning to support gait phenotype stratification.

This retrospective study included 40 children with spastic bilateral CP (80 lower limbs). EVGS item-level and subscale scores were extracted from standardized gait videos. Gait patterns were classified into four pathological subtypes by two blinded raters. A Random Forest (RF) classifier was evaluated using 5-fold stratified group cross-validation, with Multinomial Logistic Regression (MLR) serving as a benchmark. To address within-subject dependency, a sensitivity analysis involving random single-limb sampling was performed. Model performance was assessed using balanced accuracy, macro-averaged F1-score, area under the curve (AUC), and class-specific sensitivity, specificity, precision, recall, and F1-score. Feature importance was analyzed by permutation and Gini metrics.

The RF model using item-level features outperformed MLR, achieving a mean balanced accuracy of 0.64 (SD = 0.03) and a macro-AUC of 0.74 (SD = 0.05). Item-level data consistently outperformed aggregated subscale scores. Sensitivity analysis using random single-limb sampling confirmed the robustness of these findings, yielding comparable performance estimates. For class-specific performance, the model achieved moderate discriminative ability for Type 1 (AUC = 0.76) and Type 3 (AUC = 0.69), limited performance for Type 2 (AUC = 0.55) due to feature overlap, and high discriminative ability for Type 4 (AUC = 0.89). Feature importance analysis identified hip peak flexion at swing (Item 13), knee position at terminal swing (Item 10), knee peak extension at stance (Item 9), trunk peak sagittal position (Item 16), and maximum ankle dorsiflexion at swing (Item 7) as key predictors.

EVGS item-level features facilitate moderate yet robust gait subtype classification in spastic bilateral CP. The integration of EVGS with machine learning presents a promising adjunctive approach for preliminary assessment, particularly in resource-limited settings; however, external validation is requisite prior to clinical implementation.