Pathogenic Genetic Variants, Comorbid Autism and Adaptive Developmental Quotient as Independent Predictors of Intellectual Disability in Children With Global Developmental Delay: An Interpretable Machine Learning Model With Calibrated Risk Estimation

BACKGROUND: Global developmental delay (GDD) frequently precedes intellectual disability (ID), but no validated multivariable prognostic tool exists to support individualised counselling during the initial diagnostic work-up. Existing risk indicators are typically considered in isolation, and their joint contribution within an interpretable predictive framework remains uncertain. METHODS: We retrospectively analysed 2453 children diagnosed with GDD between January 2014 and December 2023 at a provincial tertiary children's rehabilitation centre, all followed to a minimum age of 60 months. Twenty-eight candidate predictors covering perinatal, developmental, neuroimaging, electrophysiological, genetic and comorbidity domains were retained after multiple imputation, multicollinearity screening with random-forest importance protection and standardisation. Five algorithms-L2- and L1-regularised logistic regression, random forest, XGBoost and LightGBM-were trained on a stratified 70% training partition with class-weight rebalancing; no synthetic minority over-sampling was applied. Probabilities from the L2 model were post hoc recalibrated by Platt scaling. We evaluated discrimination, calibration (slope and intercept after Platt scaling), Brier score and net benefit on the held-out 30% test set, with 1000 bootstrap confidence intervals. Sensitivity analyses excluded post-baseline candidate predictors, and we benchmarked the full model against parsimonious one-, three- and five-feature regressions. RESULTS: Of the cohort, 2037 children (83.0%) progressed to ID, reflecting the referral profile of a tertiary centre. The Platt-calibrated L2 logistic regression achieved an AUC of 0.783 (95% CI 0.735-0.828) with calibration slope 1.01 and intercept -0.005, and a Brier score of 0.113 (95% CI 0.097-0.130). All five algorithms performed within a 0.015 AUC band. The strongest independent risk factors were pathogenic genetic variant pathogenicity (OR 2.13, 95% CI 1.86-2.44), comorbid autism spectrum disorder (OR 1.69, 95% CI 1.41-2.03) and EEG epileptiform discharges (OR 1.43, 95% CI 1.18-1.74); higher Gesell adaptive developmental quotient was the strongest protective factor (OR 0.45 per standardised unit, 95% CI 0.34-0.60). At the Youden-optimal threshold of 0.85, sensitivity, specificity, positive and negative predictive values were 66.1%, 76.8%, 93.3% and 31.7%, respectively. Removal of early intensive intervention from the model lowered AUC by only 0.012 (95% CI - 0.002 to 0.026), indicating that retrospective treatment information was not the primary driver of model performance. A parsimonious five-feature model achieved AUC 0.762, recovering most of the discriminative signal. Discrimination was robust to strict exclusion of post-baseline predictors (AUC 0.769), to complete-case analysis (0.839) and to restriction to genetically tested children (0.864). CONCLUSIONS: Among children referred to a tertiary centre with GDD, an interpretable, well-calibrated logistic regression model integrating routinely available clinical, neurophysiological and genetic data stratifies the risk of progression to ID with moderate discrimination and high positive predictive value. The model is suitable for high-risk triage; its application to community populations will require refitting and external validation. The accompanying nomogram and SHAP explanations support clinician-facing risk communication.