Machine learning-based prediction of diabetic patients using blood routine data.

Area of Science:

• Computational biology
• Medical informatics
• Data science in healthcare

Background:

• Diabetes is a widespread chronic disease often diagnosed late.
• Conventional diagnostic methods may miss early-stage diabetes.
• There is a need for improved, timely diabetes detection methods.

Purpose of the Study:

• To develop and validate a computational framework for early diabetes prediction.
• To identify key hematological indicators associated with diabetes.
• To create a tool for preliminary diabetes risk assessment.

Main Methods:

• Collected blood routine data from 1000 diabetes patients and 1000 healthy individuals.
• Employed machine learning algorithms: eXtreme Gradient Boosting (XGBoost), random forest, support vector machine, and elastic net.
• Utilized Shapley additive explanations (SHAP) for model interpretability and logistic regression for nomogram creation.

Main Results:

• The XGBoost model demonstrated superior predictive performance (AUC 99.90% training, 98.51% testing).
• External validation achieved an overall accuracy of 81.54%.
• Identified key predictors: MCHC, LY%, RDW-SD, and MCH.

Conclusions:

• The developed XGBoost model offers a highly effective tool for diabetes prediction.
• SHAP analysis revealed crucial hematological markers for diabetes risk.
• A logistic regression-based nomogram can aid in preliminary clinical diabetes assessment.