AI-Assisted Early Prediction of Autism Spectrum Disorder in Children Using Clinical and Developmental Features

Abstract

Background: Machine-learning models are increasingly applied to autism spectrum disorder (ASD) screening and frequently report exceptionally high classification effectiveness. However, such results may be inflated when recorded screening labels are derived from questionnaire responses that also serve as model predictors. This study considers the influence of questionnaire-derived features on predictive performance and examines whether demographic and contextual characteristics provide independent discriminatory information in toddler and child autism screening datasets.

Methods: Two publicly available datasets, comprising 1,054 toddler and 292 child screening records, were analyzed using five machine-learning algorithms: logistic regression, random forest, support vector machines, gradient boosting, and XGBoost. Models were evaluated using stratified 75/25 training–testing splits, five-fold training cross-validation, and 400-resample bootstrap confidence intervals. Three predictor configurations were compared: complete features, exclusion of aggregate screening scores, and exclusion of both aggregate scores and individual questionnaire responses. Predictive discrimination was assessed primarily using the area under the receiver operating characteristic curve (ROC-AUC).

Results: Aggregate screening-score thresholds reproduced 100% of recorded classification labels in both datasets, showing a deterministic relationship between screening scores and outcomes. For toddlers, random-forest held-out ROC-AUC was 1.000 with all features, 0.995 after aggregate-score removal, and 0.670 (95% CI: 0.598–0.747) using contextual features alone. Corresponding values in children were 1.000, 0.985, and 0.567 (95% CI: 0.428–0.685), respectively. Across all five algorithms, context-only models showed weak discrimination in children (ROC-AUC: 0.526–0.580). The substantial reduction in performance following questionnaire exclusion suggests that near-perfect classification primarily reflects reconstruction of questionnaire-derived labels rather than independent identification of autism-related clinical characteristics.

Conclusions: Exceptionally high predictive validity in questionnaire-based autism screening datasets may largely reflect target leakage due to overlap between predictors and recorded outcomes. Removing aggregate screening scores alone does not resolve this issue if the underlying questionnaire items remain available. Independent clinician-adjudicated diagnostic results, external validation, and rigorous feature-selection procedures are essential before such models can be considered reliable for clinical applications