An interpretable progressive residual network for automated multiclass diabetes diagnosis

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Chronic Disease Management

Background:

• Diabetes mellitus is a global health burden, with early detection hindered by invasive methods and high costs.
• Current machine learning models often simplify diabetes diagnosis to a binary task, neglecting the crucial pre-diabetic stage and lacking interpretability.

Purpose of the Study:

• To develop an interpretable deep learning model, ProgMDD, for multiclass diabetes diagnosis, including pre-diabetes, using routine clinical biomarkers.
• To address the limitations of existing "black box" models by enhancing transparency and clinical applicability.

Main Methods:

• Implemented a progressive residual network (ProgMDD) with channel attention and multi-level regularization for enhanced feature learning.
• Utilized a strict, leakage-free pipeline with LASSO for feature selection and UMAP for data visualization.
• Compared ProgMDD against multiple baseline models using 5-fold cross-validation.

Main Results:

• ProgMDD achieved a mean accuracy of 97.02% in cross-validation and 97.59% on an imbalanced hold-out test set.
• Feature importance analysis using LASSO and SHAP demonstrated biological plausibility and model transparency.
• Ablation studies validated the effectiveness of ProgMDD's architecture and components.

Conclusions:

• ProgMDD provides a feasible, interpretable, and accurate approach for early diabetes and pre-diabetes screening using low-cost clinical data.
• The model's transparency and performance support its application in primary care for risk stratification.
• The methodological paradigm is transferable to other chronic disease prediction tasks.