Identifying Clinical Predictors of Diabetes and Prediabetes: An Explainable AI Approach Using Primary Care Electronic...

Area of Science:

• Computational medicine
• Metabolic disease prediction
• Primary care informatics

Background:

• Prediabetes (PD) is a precursor to type 2 diabetes (T2DM), posing a significant health risk.
• Effective management of PD involves lifestyle modifications to normalize blood glucose levels.
• Early identification of individuals at risk for T2DM is crucial for preventive strategies.

Purpose of the Study:

• To develop and compare machine learning models for predicting normoglycemia, PD, and T2DM.
• To leverage routinely collected Canadian primary care Electronic Medical Records (EMRs) for predictive modeling.
• To enhance understanding of PD to T2DM progression using explainability techniques.

Main Methods:

• Utilized a large dataset of 21,023 patients from Canadian primary care EMRs.
• Expanded existing clinical data with drug prescriptions and smoking status information.
• Trained and compared various machine learning models for predictive accuracy.

Main Results:

• Identified key predictors of T2DM risk, including blood sugar levels, BMI, cholesterol, and blood pressure.
• Demonstrated the potential of EMR data for identifying individuals at high risk of T2DM.
• Explainability techniques provided insights into the progression pathways from PD to T2DM.

Conclusions:

• Routinely collected clinical data can effectively support early identification of high-risk individuals for T2DM.
• Machine learning models show promise in predicting metabolic disease trajectories.
• Further validation in prospective cohorts and feature expansion are recommended for improved generalizability.