New AI explained and validated deep learning approaches to accurately predict diabetes.

Area of Science:

• * Computational biology and machine learning applications in healthcare.
• * Development of advanced artificial intelligence (AI) algorithms for disease prediction.

Background:

• * Diabetes mellitus is a critical metabolic condition requiring early detection to prevent severe chronic complications and organ failure.
• * Existing predictive models for diabetes often suffer from low accuracy, class imbalance issues, and a lack of transparency in their decision-making processes.
• * Deep learning (DL) models show potential but require enhancements for practical clinical application.

Purpose of the Study:

• * To introduce two novel deep learning models, LeDNet and HiDenNet, for enhanced early and accurate diabetes prediction.
• * To address the challenges of class imbalance and poor interpretability in current diabetes prediction models.
• * To improve the reliability and transparency of AI-driven tools for clinical decision support in diabetes diagnosis.

Main Methods:

• * Development and training of two novel deep learning architectures: LeDNet (LeNet + Dual Attention Network) and HiDenNet (Highway Network + DenseNet).
• * Utilization of the Diabetes Health Indicators dataset, with mitigation of class imbalance via majority-weighted minority over-sampling.
• * Application of K-fold cross-validation for model stability assessment and integration of explainable AI (XAI) techniques (LIME, SHAP) for interpretability.

Main Results:

• * LeDNet achieved an F1-score of 85%, recall of 84%, accuracy of 85%, and precision of 86%.
• * HiDenNet demonstrated comparable performance with accuracy, F1-score, recall, and precision all at 85% or 86%.
• * Both models outperformed existing state-of-the-art deep learning models and provided interpretable feature insights through XAI.

Conclusions:

• * LeDNet and HiDenNet offer significant improvements in accuracy and reliability for early diabetes prediction compared to current DL models.
• * The proposed models effectively handle class imbalance and provide crucial interpretability, addressing key limitations of previous approaches.
• * These explainable AI-enhanced models represent promising tools for clinical decision-making and early diabetes diagnosis, enhancing transparency and trust.