MACHINE LEARNING-BASED INTEGRATION OF BEHAVIORAL AND CLINICAL FEATURES FOR AUTISM SCREENING IN YOUNG CHILDREN

Abstract

Background: Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder marked by ongoing deficits in social communication and behavior that is restricted or repetitive. Early identification improves outcomes, but traditional screening relies on time-consuming clinician-administered tools and parent questionnaires that are open to variable interpretation and specialist availability. Multimodal artificial intelligence (AI) and machine learning (ML) approaches have been proposed to enable faster and more consistent preliminary screening. Objective: This study presents a structured multimodal data collection instrument for ASD screening and evaluates a planned machine learning pipeline (Random Forest, Support Vector Machine and XGBoost) for classification of screening risk, illustrated using a synthetic, diagnostically labelled dataset of 40 patient records. Methods: The screening proforma was field-tested for usability  on forty real cases (18–22 months); because these forms were not diagnostically labelled, they could not be used for model development. To demonstrate and evaluate the planned classification pipeline, a synthetic dataset of 40 records (P001–P040; ages 18–60 months) with M-CHAT, CARS and ADOS scores and corresponding diagnostic categories (ASD, Non-ASD Developmental Delay, Typical Development) was analysed. Three classifiers (Random Forest, SVM, XGBoost) were evaluated using repeated stratified 5-fold cross-validation, with simulated measurement noise added to the screening scores to approximate realistic inter-rater/test–retest variability rather than the perfect separability of the raw synthetic thresholds.. Results: Of the 40 synthetic cases, 19 (47.5%) were labelled ASD, 12 (30.0%) Non-ASD Developmental Delay, and 9 (22.5%) Typical Development. Cross-validated classifiers achieved accuracy of 92.4–96.5% (mean 94.9%), sensitivity 90.2–96.8%, specificity 94.4–96.4%, and AUC 0.97–0.99. CARS, ADOS and M-CHAT scores were the dominant predictors (combined feature importance ≈ 0.97), with demographic variables contributing minimally. Conclusion: The structured multimodal ASD screening proforma was found feasible to administer, and the planned Random Forest/SVM/XGBoost pipeline achieved strong (≈95%), but not perfect, classification performance on a synthetic, diagnostically-labelled dataset. Because the underlying dataset was synthetically generated with distinct score thresholds per diagnostic category, even this ~95% figure likely overstates real world performance; it should be read as a feasibility demonstration of the analysis pipeline, not as evidence of clinical validity. Larger multicentre data collection with confirmed, real diagnostic labelling will be required to assess genuine predictive performance or clinical utility.