Non-invasive opportunistic screening for diabetes mellitus: an interpretable stacking ensemble framework.

Area of Science:

• Medical Informatics
• Public Health
• Diabetes Research

Background:

• Opportunistic screening for type 2 diabetes (T2D) is crucial for early detection.
• Non-invasive methods are desirable to increase accessibility and reduce patient burden.
• Existing screening often relies on invasive tests or complex risk scores.

Purpose of the Study:

• To develop and evaluate a non-invasive screening model for T2D using a Stacking ensemble approach.
• To assess the model's performance in terms of sensitivity, specificity, and discriminative ability.
• To explore the impact of modifiable behavioral factors on T2D screening probability.

Main Methods:

• Development of a heterogeneous Stacking ensemble model (Task C) utilizing non-invasive features.
• Features included demographic, lifestyle, medical history, and symptom-based data.
• Model performance was evaluated using Recall, AUC, specificity, and Brier Score.

Main Results:

• The model achieved high sensitivity (Recall = 0.9267) but modest overall discriminative performance (AUC = 0.5515) and low specificity (0.1106).
• Simulation analyses indicated that adjusting the top three modifiable behavioral factors accounted for 85.5% of the risk reduction.
• Decision curve analysis suggested utility primarily at lower screening thresholds.

Conclusions:

• The proposed non-invasive ensemble model is a technically feasible tool for preliminary T2D case-finding.
• The model provides insights into the significant impact of modifiable behavioral factors on T2D risk.
• Further validation is needed, but the approach shows promise for accessible diabetes screening.