Optimized prediction of diabetes complications using ensemble learning with Bayesian optimization: a cost-efficient...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Diabetology

Background:

• Global rise in diabetes prevalence necessitates improved complication prediction.
• Current models often overlook fundamental laboratory parameters.
• Need for cost-effective predictive tools in diabetes management.

Purpose of the Study:

• Develop an optimized predictive model for diabetes complications.
• Utilize 12 frequently tested laboratory indicators.
• Enhance early intervention strategies through accurate prediction.

Main Methods:

• Dataset compilation from a tertiary hospital with rigorous cleaning.
• Training and evaluation of multiple machine learning classifiers (Random Forest, XGBoost, SVM, MLP).
• Implementation of an ensemble learning model with Bayesian optimization and feature importance analysis.

Main Results:

• Ensemble model achieved >90% accuracy for predicting diabetic complications.
• Exceptional performance for diabetic nephropathy prediction (98.50% accuracy, 99.76% AUC).
• Cost reduction of 2.5% achieved through feature elimination without accuracy loss.

Conclusions:

• High-quality dataset of laboratory indicators developed.
• Accurate and cost-efficient predictive model for diabetes complications created.
• Model offers a foundation for clinical application and reduces testing expenses.