Development and Validation of a Prognostic Model for Independent Walking in Children With Cerebral Palsy Based on Machine Learning.

Abstract

Objective: To develop and validate machine learning-based models for predicting independent walking ability in children with cerebral palsy (CP).

Design: Retrospective cohort study.

Setting: Data were collected from a national CP registry platform and follow-up assessments were conducted through telephone interviews.

Participants: Children with CP (n=807) registered between January 2016 and December 2020, with follow-up data collected from October 2022 to March 2023.

Interventions: Not applicable.

Main outcome measures: The primary outcome was independently walking before the age of 6 years.

Results: Among the 807 participants, 561 (69.5%) achieved independent walking. Univariate Cox regression identified several predictive factors, including neonatal asphyxia, bilirubin encephalopathy, Gross Motor Function Classification System level before age of 2 years, age of independent sitting, type of CP, magnetic resonance imaging classification, Gross Motor Function Measure-88 scores, epilepsy, intellectual disability, early preterm birth, and very low birth weight (P<.05). Machine learning models demonstrated excellent predictive performance, with logistic regression achieving the highest area under the curve (AUC=0.947), followed by XGBoost (AUC=0.946) and multilayer perceptron (AUC=0.945). Cox proportional hazard models identified key predictors for the timing of independent walking, with a nomogram constructed for clinical application. Internal validation confirmed model reliability, although calibration curves indicated potential overestimation for ages 5-6 years.

Conclusions: Machine learning models accurately predict independent walking ability in children with CP, although calibration analyses indicated potential overestimation for children aged 5-6 years. The proposed nomogram provides clinicians with an interpretable tool for personalized prognosis. Although internal validation demonstrated excellent performance, future external validation in multicenter cohorts will be critical to confirm generalizability.