| Sumario: | Accurate classification of gait patterns in children with cerebral palsy (CP) is critical for guiding treatment but requires expert interpretation of data. Automated methods have potential to improve consistency and scalability; however, clinically meaningful automated tools are limited. This work aimed to determine if automated algorithms based on established gait classifications can reproduce expert clinical classification in children with CP. An automated MATLAB algorithm was developed to assign gait classifications using two established systems (Rodda & Graham and Rozumalski & Schwartz). A sample of children with CP who met criteria for crouch gait underwent automated classification. Three expert gait analysts independently classified all trials using standardized definitions corresponding to each system. Fleiss's κ quantified inter-rater reliability, while Cohen's κ, weighted κ, percent agreement, and macro F1 scores quantified agreement between each reviewer and the automated classifications. Inter-rater reliability among reviewers was substantial for Rodda & Graham (κ = 0.753), with high agreement between reviewers. Inter-rater reliability among reviewers was moderate for Rozumalski & Schwartz (κ = 0.456) and agreement between raters was lower and more variable, with some classifications demonstrating full disagreement among reviewers. Automated gait classification based on quantitative gait criteria can achieve agreement with gait analysis experts for systems with clear biomechanical boundaries. More complex cluster–based systems yield lower agreement, reflecting inherent ambiguity in cluster overlap. These findings support use of automated tools as reliable, objective ground-truth for large-scale analyses and for training markerless or video-based assessment algorithms aimed at expanding gait evaluation beyond specialized motion laboratories. • There is a need for clinically meaningful automated gait classification tools. • Rodda & Graham reported almost perfect agreement with expert gait interpretation. • Cluster–based systems yield lower agreement, reflecting more overlap in clusters. • These methods can serve as objective ground truth for training of large-scale analyses.
|
|---|