An Interpretable Machine Learning Model for Predicting Intellectual Disability in Children With Cerebral Palsy

Intellectual disability (ID) affects approximately 45% of children with cerebral palsy (CP), yet early identification is frequently hindered by severe motor and communication impairments. This study aimed to develop and validate an interpretable machine learning (ML) framework for predicting ID risk in children with CP.

In this retrospective, registry-based study, data from 807 children with CP were analysed. To ensure temporal validity, all predictors were restricted to clinical and neuroimaging assessments confirmed by 2 years of age. Eight ML algorithms were trained and compared on an independent test set, and SHapley Additive exPlanations (SHAP) were applied to interpret model output at both the global and the individual levels.

The optimized models achieved robust discriminative performance, with the highest area under the receiver operating characteristic curve (AUC) reaching 0.813 on the independent test set. SHAP analysis revealed a highly skewed distribution of predictive features: The inability to achieve independent sitting by age 2 was the most critical risk factor, followed by early-onset epilepsy, spastic quadriplegia and severe Gross Motor Function Classification System (GMFCS) levels. Baseline perinatal factors demonstrated lower direct predictive utility, and local SHAP analyses successfully mapped individualized risk trajectories.

This transparent ML approach functions as a reliable decision-support tool, translating complex algorithmic output into clinically intuitive insights. It may empower clinicians to move from 'wait-and-see' approaches towards timely, personalized neurodevelopmental interventions for high-risk children.