Early Type 2 diabetes risk prediction using explainable machine learning in a two-stage approach

Area of Science:

• Medical Informatics
• Machine Learning
• Public Health

Background:

• Diabetes mellitus is a growing global health concern with significant complications and rising healthcare costs.
• Early detection and effective management are crucial for mitigating diabetes' impact.
• A substantial number of diabetes cases remain undiagnosed worldwide.

Purpose of the Study:

• To develop and validate an explainable, two-stage machine learning (ML) framework for predicting diabetes risk.
• To enhance the early identification and management of undiagnosed diabetes cases.
• To integrate ML models into mobile health applications for improved community health worker capacity.

Main Methods:

• A two-stage ML approach was employed, utilizing public and Rwandan datasets.
• Stage one involved feature selection using Shapley Additive exPlanations (SHAP) and Multi-Layer Perceptron (MLP) weights on 520 samples.
• Stage two applied five ML models (MLP, SVM, KNN, XGBoost, Naïve Bayes) to a 270,943-sample dataset from Rwanda, with SHAP for output explanation.

Main Results:

• The MLP model achieved 95.19% accuracy in stage one, identifying key diabetes risk predictors.
• The Extreme Gradient Boosting (XGBoost) model demonstrated superior performance in stage two, reaching 97.14% accuracy.
• Identified influential predictors align with clinical recommendations for diabetes care.

Conclusions:

• A novel two-stage, explainable ML framework for systematic type 2 diabetes screening has been established.
• The framework integrates symptom-based risk evaluation with demographic and clinical data for refined assessment.
• Integration with mUzima mobile app can empower community health workers for early diabetes detection and referral, potentially reducing undiagnosed cases and improving disease management.