Scholarlink

This study presents a scalable and interpretable deep learning framework for the simultaneous prediction of attention deficit hyperactivity disorder (ADHD) and Diabetes, leveraging heterogeneous health data. The architecture integrates clinical metrics from electronic health records (EHRs), physiological signals from wearable devices, and behavioural assessments. A hybrid feature selection strategy combining LASSO, Recursive Feature Elimination (RFE), Mutual Information, and XGBoost was used to reduce over 400 input variables to 38 high-impact predictors. These were input into a multi-branch neural network: a CNN for static features, an LSTM for temporal data, and an MLP fusion layer producing dual-diagnosis outputs. The model achieved high predictive accuracy (93.2% for Diabetes, 92.5% for ADHD) with AUCs exceeding 0.94. Integrated explainable AI tools (SHAP, LIME, Grad-CAM) revealed clinically relevant predictors, including HRV entropy, glucose variability, BMI, and impulsivity. The framework demonstrated strong generalizability across diverse demographic groups and maintained robustness in comorbid cases. Unlike existing models, the proposed pipeline embeds explainability into the training loop, enabling both interpretability and refinement. This work advances the state-of-the-art in comorbidity-aware AI systems and offers a practical solution for early and reliable multi-disease diagnosis in clinical settings.

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

A Scalable Deep Learning and Explainable AI Framework for Dual Diagnosis of ADHD and Diabetes Using Multi-source Health Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tanguturu S. P. Madhuri,
  • G. S. Raghavendra

摘要

This study presents a scalable and interpretable deep learning framework for the simultaneous prediction of attention deficit hyperactivity disorder (ADHD) and Diabetes, leveraging heterogeneous health data. The architecture integrates clinical metrics from electronic health records (EHRs), physiological signals from wearable devices, and behavioural assessments. A hybrid feature selection strategy combining LASSO, Recursive Feature Elimination (RFE), Mutual Information, and XGBoost was used to reduce over 400 input variables to 38 high-impact predictors. These were input into a multi-branch neural network: a CNN for static features, an LSTM for temporal data, and an MLP fusion layer producing dual-diagnosis outputs. The model achieved high predictive accuracy (93.2% for Diabetes, 92.5% for ADHD) with AUCs exceeding 0.94. Integrated explainable AI tools (SHAP, LIME, Grad-CAM) revealed clinically relevant predictors, including HRV entropy, glucose variability, BMI, and impulsivity. The framework demonstrated strong generalizability across diverse demographic groups and maintained robustness in comorbid cases. Unlike existing models, the proposed pipeline embeds explainability into the training loop, enabling both interpretability and refinement. This work advances the state-of-the-art in comorbidity-aware AI systems and offers a practical solution for early and reliable multi-disease diagnosis in clinical settings.