AI-driven oncogenic risk stratification in pediatric autism spectrum disorder: A multi-omic machine learning framework for early cancer prevention.

12

Background: Children with Autism Spectrum Disorder (ASD), particularly those with co-occurring intellectual disabilities or PTEN/TSC1/2 mutations, exhibit an elevated cancer risk (OR up to 4.8). Despite this, standardized oncological screening for neurodivergent populations remains suboptimal, especially in Low-Middle Income Countries (LMICs). This study evaluated the efficacy of a multi-task deep learning framework in identifying high-risk oncogenic profiles in pediatric ASD cohorts. Methods: A retrospective analysis was conducted using a multi-omic dataset of 5,120 pediatric profiles (2018–2025) from LMIC healthcare settings. The study assessed the integration of genomic features (rare coding variants), maternal metabolic history, and environmental chemical exposures. A machine learning architecture utilizing Random Forest (RF), XGBoost, and SHAP-based Explainable AI (XAI) was implemented to determine risk-stratification accuracy and identify shared ASD-cancer biomarkers. Results: Of the 1,120 profiles analyzed, the AI framework identified high-risk oncogenic signatures in 12.4% of the cohort. The model achieved a high diagnostic accuracy with an AUC-ROC of 0.92 (p < 0.001). Genomic feature extraction highlighted 138 shared genes; patients with PTEN mutations were significantly more likely to be flagged for high risk compared to those with non-syndromic ASD (88.4% vs. 15.2%, p < 0.001). High-risk stratification was more frequent in nulliparous maternal lineages (42.1%) and those with documented early-life chemical exposure (64.5% vs. 28.3%, p < 0.01). Explainable AI (XAI) modules provided interpretable clinical pathways for 97.8% of flagged cases, demonstrating high robustness even in data-sparse environments typical of underprivileged regions. Conclusions: The current provision of cancer surveillance in ASD populations is disproportionately reactive rather than preventive. This AI-driven framework offers a scalable, objective tool for early risk stratification, bridging the gap between neurodevelopmental monitoring and oncological screening. Integrating such digital biomarkers into primary pediatric care offers an underutilized opportunity for early intervention, particularly for underprivileged populations where access to advanced genetic testing is limited.