Scholarlink

<p>Autism Spectrum Disorder (ASD) requires both accurate identification and individualized instructional support, yet most computational solutions treat diagnosis, severity estimation, and personalized learning as separate tasks, limiting their practical utility. This work introduces a unified and interpretable artificial intelligence framework that integrates behavioral screening, severity-aware classification, and adaptive e-learning within a single end-to-end architecture. Noise-resilient latent representations are first extracted from heterogeneous caregiver-reported questionnaires using a Variational Autoencoder (VAE). These embeddings are subsequently classified with XGBoost, while SHAP-based attributions provide transparent and clinically grounded explanations of severity predictions. The resulting diagnostic profile conditions a Generative AI module GPT4All for adaptive textual scaffolding and Stable Diffusion for multimodal visual synthesis, while a LinUCB contextual bandit optimizes content sequencing based on real-time engagement signals. Extensive experiments across toddler, child, adolescent, and adult cohorts demonstrate robust performance in both binary ASD identification (accuracy 0.96–1.00; AUC 0.95–0.99) and fine-grained severity grading. Recommendation evaluations further indicate strong personalization effectiveness, with engagement scores exceeding 0.75 and task-completion rates above 0.85. Beyond achieving high predictive accuracy, the framework operationalizes interpretable severity estimates into dynamically personalized multimodal learning pathways, addressing long-standing gaps in the separation between diagnostic intelligence and intervention design. By uniting representation learning, explainable classification, generative personalization, and contextual recommendation into a coherent pipeline, this study provides a clinically aligned, data-efficient, and scalable foundation for next-generation digital learning environments tailored to the diverse developmental needs of individuals on the autism spectrum.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Unified interpretable AI for autism diagnosis and scalable severity-aware personalized adaptive e-learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月10日

查看全文
  • Vijayalaxmi N. Rathod,
  • R. H. Goudar,
  • Sangeeta Sangani

摘要

Autism Spectrum Disorder (ASD) requires both accurate identification and individualized instructional support, yet most computational solutions treat diagnosis, severity estimation, and personalized learning as separate tasks, limiting their practical utility. This work introduces a unified and interpretable artificial intelligence framework that integrates behavioral screening, severity-aware classification, and adaptive e-learning within a single end-to-end architecture. Noise-resilient latent representations are first extracted from heterogeneous caregiver-reported questionnaires using a Variational Autoencoder (VAE). These embeddings are subsequently classified with XGBoost, while SHAP-based attributions provide transparent and clinically grounded explanations of severity predictions. The resulting diagnostic profile conditions a Generative AI module GPT4All for adaptive textual scaffolding and Stable Diffusion for multimodal visual synthesis, while a LinUCB contextual bandit optimizes content sequencing based on real-time engagement signals. Extensive experiments across toddler, child, adolescent, and adult cohorts demonstrate robust performance in both binary ASD identification (accuracy 0.96–1.00; AUC 0.95–0.99) and fine-grained severity grading. Recommendation evaluations further indicate strong personalization effectiveness, with engagement scores exceeding 0.75 and task-completion rates above 0.85. Beyond achieving high predictive accuracy, the framework operationalizes interpretable severity estimates into dynamically personalized multimodal learning pathways, addressing long-standing gaps in the separation between diagnostic intelligence and intervention design. By uniting representation learning, explainable classification, generative personalization, and contextual recommendation into a coherent pipeline, this study provides a clinically aligned, data-efficient, and scalable foundation for next-generation digital learning environments tailored to the diverse developmental needs of individuals on the autism spectrum.