Improving Machine Learning Diabetes Prediction Models for the Utmost Clinical Effectiveness.

Area of Science:

• Endocrinology and Metabolic Diseases
• Computational Biology and Bioinformatics
• Preventive Medicine

Background:

• Early diabetes prediction is crucial for timely interventions.
• Machine learning models can identify individuals at high risk for diabetes.
• Existing prediction models require further performance enhancement.

Purpose of the Study:

• To develop and evaluate a high-performance machine learning model for early diabetes prediction.
• To compare the efficacy of various machine learning algorithms for diabetes risk assessment.
• To investigate the utility of survival analysis models for predicting diabetes onset.

Main Methods:

• Utilized a large dataset of 38,379 subjects for model training and validation.
• Compared the performance of logistic regression, decision tree, random forest, XGBoost, Cox regression, and XGBoost Survival Embedding (XGBSE).
• Assessed model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and C-index, with threshold adjustments.

Main Results:

• XGBoost demonstrated the highest AUROC among classification algorithms.
• XGBSE achieved an AUROC > 0.9 for 2- to 9-year predictions and a C-index of 0.934.
• Adjusting the prediction threshold significantly improved sensitivity with minimal impact on specificity.

Conclusions:

• A high-performance diabetes prediction model was developed using the XGBSE algorithm with threshold adjustment.
• The developed model shows promise for real-world clinical application in diabetes prevention.
• Further external validation and simplification are planned for clinical implementation.