Effective therapeutic intervention for individuals with Autism Spectrum Disorder (ASD) demands continuous adaptation to rapidly shifting affective states across neurological, facial, and physiological modalities. Classical rule-based recommendation systems are inherently static and cannot learn from interaction outcomes over time. We propose a Double Dueling Deep Q-Network (D3QN) framework driven by a novel Emotion-Shaped Reward (ESR) function that decomposes the immediate therapeutic signal into four clinically grounded components, intervention efficacy, arousal regulation, emotional valence, and social engagement, weighted by evidence from the ASD intervention literature, augmented by a mismatch penalty for contraindicated stimulation under high cognitive load. Experiments on a synthetic multimodal cohort (N = 250; ASD= 163, TD= 87) with 24 features spanning brain connectivity, facial affect, and physiological arousal demonstrate that D3QN-ESR achieves a mean normalised reward of 0.200±0.102, outperforming random (0.143), rule-based (0.137), and best-fixed-action (0.163) baselines by statistically significant margins ((p < 10−4; Cohen's d ≥ 0.88). The learned policy allocates 41.7% of selections to Sensory Break for high-severity profiles, exhibiting strong clinical coherence. These results establish emotion-shaped reward shaping as a principled methodology for personalised ASD intervention, with direct implications for real-time adaptive support systems embedded in multimodal sensing platforms.
Ayat et al. (Mon,) studied this question.