Jacqueline Romkes

Three-dimensional instrumented gait analysis (3D-IGA) is widely used to guide surgical decision-making in children with cerebral palsy (CP), but its interpretation is time-consuming and prone to inter-rater variability. We investigated whether a generative large language model (LLM) could consistently generate gait deviation findings and surgical procedure suggestions that align with expert judgement. Kinematic features for lower-limb joints across the gait cycle, stance, and swing were extracted from eight children with unilateral CP using the open-source GaitSharing Toolkit and a structured prompt, then submitted three times per patient to OpenAI's GPT-5.5 model. The model assessed 28 kinematic deviations and 12 surgical procedure groups using majority voting. One gait analyst rated the deviations, and two paediatric orthopaedic surgeons independently rated the procedures on a 0–2 ordinal scale, blinded to all clinical information beyond the kinematic curves and diagnosis. Agreement with the gait expert averaged 84.7% in the sagittal plane and was lowest at the knee (62.6%). Surgeons’ agreement with the LLM reached 83.9% and 73.5%, with the tibialis anterior procedure showing the lowest concordance. Inter-surgeon agreement was 79.2%. The LLM showed high self-consistency (>90% across runs). This work demonstrates the potential of generative LLMs as assistive tools in clinical gait analysis for deviation detection and future treatment planning.