Effectiveness Evaluation of Adaptive Difficulty Adjustment Algorithms with Multimodal Feedback for Social Skills Training in Children with Autism Spectrum Disorder

Abstract

Children diagnosed with autism spectrum disorder display substantial heterogeneity in social communication skills, necessitating intervention strategies that dynamically adapt to individual developmental trajectories. We propose a hierarchical reinforcement learning framework that integrates multimodal behavioral streams to guide difficulty progression in therapeutic social scenarios. Skill development is modeled as a constrained Markov decision process, with difficulty vectors d in D evolving according to composite performance signals p(t) and engagement indicators e(t), where the optimization objective J = E [sum over t of gamma ^ t * (r_skill (s_t, a_t) + lambda * r_engage (s_t, a_t))] balances immediate skill gains against sustained participation. Three synchronized channels-facial landmarks tracked via 68-point models sampled at 120 Hz, acoustic features represented by 13-dimensional MFCCs, and skeletal configurations captured across 25 anatomical joints-are processed through temporal convolutional networks. Attention-weighted aggregation f_fusion = sum over i of alpha_i * phi_i (x_i) allows each modality-specific encoder phi_i to contribute proportionally to its instantaneous reliability. Clinical trials involving 124 participants (ages 6-14, ADOS-2 scores 12.4 ± 3.2) demonstrate a 42.3% acceleration in competency acquisition compared with therapist-directed baselines (hierarchical model coefficient beta_time×condition = 1.42, SE = 0.18, t (1984) = 7.89, p < 0.001). Transition prediction between difficulty states achieves 87.4% accuracy. Power-law retention analysis indicates reduced forgetting in the adaptive framework (b_adaptive = 0.084 versus b_control = 0.162), with 78.4% of acquired competencies maintained at a 12-week follow-up.