Deep neural networks and deep deterministic policy gradient for early ASD diagnosis and personalized intervention in children

Early diagnosis and personalized intervention for Autism Spectrum Disorder (ASD) in children can potentially improve developmental outcomes, though current methods often lack scalability and adaptability. This study introduces an integrated system combining a deep neural network (DNN) and a Deep Deterministic Policy Gradient (DDPG) reinforcement learning framework for early ASD detection and adaptive psychosocial intervention. The DNN, trained and validated on diverse datasets spanning toddlers to adolescents (sourced from the University of Arkansas, Vaishnavi Sirigiri, and Afarin Bargrizan), achieved a predictive accuracy of 96.98% with precision (97.65%), recall (96.74%), and ROC AUC (99.75%) on the test sets, demonstrating superior performance compared to traditional models like Random Forest and Logistic Regression. Key features, such as Qchat-10-Score and ethnicity, were identified using multi-strategy selection (LASSO, Random Forest). Building on these predictions, the DDPG-based intervention system simulated personalized strategies over 12 monthly cycles using virtual data to optimize intervention type, frequency, and intensity, resulting in observed improvements of up to 25% in social skills, up to 30% reduction in behavioral issues, and up to 20% improvement in emotional stability, with a reduction in high-risk ASD cases from 65 to 25% in the simulated cohort. This system offers a promising, data-driven approach to ASD management, enhancing early screening and tailoring interventions to individual needs.