A Multicenter Standardized Gait Database from the ORITEL Network for Cerebral Palsy Gait Analysis

Gait analysis is a cornerstone of clinical decision-making in cerebral palsy (CP), yet multicenter variability limits comparability and translation. This work reports the creation of a standardized, multicenter gait database and a machine learning (ML) classifier to distinguish gait patterns of CP hemiplegia, CP diplegia, and typically developing group to validate the database. Data is contributed by eight ORITEL motion analysis laboratories (156 sessions, 582 trials) using two systems (Vicon/BTS), with unified nomenclature, unit normalization, and resampling. For classification, 95 sessions (413 trials) meeting inclusion criteria were retained. Feature sets combined kinematic, spatiotemporal, and anthropometric variables. Class imbalance was addressed using sample weighting, dimensionality was reduced with PCA (95% variance), and models were tuned with Optuna and evaluated using stratified group cross-validation and a held-out set. A multilayer perceptron (MLP) with sample weighting achieved the best performance (test accuracy 82%; class F1-scores: diplegia 0.81, hemiplegia 0.63, control 0.98). Feature importance and nonparametric statistics highlighted clinically interpretable discriminants-double support, stride length, swing/stance phases, and minimum knee flexion. An exploratory linkage between classified patterns and reported treatments suggested distinct tendencies across hemiplegia and diplegia, aligning with evidence-based recommendations. The proposed resource and pipeline advance multicenter data standardization and provide a clinically grounded baseline for data-driven support of rehabilitation planning in CP.