LAGRANGIAN OPTIMIZATION-BASED DEEP Q-NETWORK FOR ROBUST AUTISM PREDICTION USING BEHAVIOURAL SCREENING DATA ACROSS AGE GROUPS

Autism Spectrum Disorder (ASD) prediction remains a major challenge due to behavioural heterogeneity, data imbalance, and the absence of stable cues across age groups. Existing studies have relied on conventional classifiers or deep learning models that either fail to capture sequential dependencies in screening data or suffer from instability caused by noisy, incomplete, and imbalanced records. This gap highlights the lack of constraint-aware models that can both generalize and provide reliable decision confidence for early ASD detection. To address this, the present work introduces a Lagrangian optimization-based Deep Q-Network (LO-DQN), a novel reinforcement learning framework that embeds constraint-driven mechanisms entropy regularization, replay prioritization, Q-value bias suppression, and target synchronization. This design ensures stability in learning from diverse behavioural datasets and reduces the risks of overestimation and premature convergence common in existing methods. The study contributes new knowledge by demonstrating how constraint-guided reinforcement learning can be adapted to ASD screening contexts. Experimental results on structured, non-imaging datasets show improved classification performance and stability across varied diagnostic profiles. The findings establish LO-DQN as a scalable, interpretable, and constraint-resilient framework, opening a new pathway for early, low-cost, and clinically relevant ASD prediction.