Large language models for gait classification and treatment recommendation in cerebral palsy
Related Experiment Video
Updated: Oct 30, 2026
Deep-Learning Based Multi-Joint Synchronous Tracking for Objective Quantification of Hindlimb Locomotor Kinematics in Rats
Published on: April 3, 2026
Meltem Çelik1, Sema Ertan-Birsel2, Elif Demirci1
1Ortopediatri Istanbul, Academy of Pediatric Orthopaedics, Istanbul, Turkey.
Background:
Large language model (LLM)-based artificial intelligence systems are increasingly being explored for the interpretation of clinical and biomechanical data. However, their ability to classify cerebral palsy gait patterns and support treatment-related decision-making remains unclear.
Research Question:
How accurately can commonly used LLM-based AI systems classify sagittal gait patterns in children with spastic diplegic cerebral palsy using kinematic graphs and three-dimensional skeletal gait videos, and how well do their treatment recommendations agree with expert decisions?
Methods:
Fifteen children with spastic diplegic cerebral palsy, including three cases from each Rodda sagittal gait pattern, were retrospectively evaluated. ChatGPT, Gemini, and Claude independently classified each case using bilateral sagittal-plane kinematic graphs and three-dimensional skeletal gait model videos and provided treatment recommendations. Expert consensus served as the reference standard. Agreement was assessed using Cohen's and Fleiss' kappa, together with classification accuracy.
Results:
Classification accuracy was higher in the graph-based condition than in the video-based condition. Claude showed the highest graph-based performance (73.3% accuracy; κ=0.67), while ChatGPT and Gemini each achieved 46.7% accuracy (κ=0.33). Video-based accuracy decreased to 26.7% for ChatGPT and Gemini (κ=0.08) and 20.0% for Claude (κ=0.00). Inter-model agreement was fair for graph-based classification (κ=0.31) but poor for video-based classification (κ=-0.07). Treatment recommendation agreement was generally low across models and modalities.
Significance:
Under the tested conditions, LLMs achieved higher accuracy with kinematic graphs, although the difference cannot be attributed to input modality alone. Graph-based classification showed promising performance; however, limited treatment concordance indicates that current LLMs should be considered supportive rather than independent clinical decision-making tools.
