Development of Various Diabetes Prediction Models Using Machine Learning Techniques.

Area of Science:

• Medical Informatics
• Public Health
• Machine Learning in Healthcare

Background:

• Existing diabetes mellitus (DM) prediction models often lack clear clinical applicability.
• There is a need for practical DM prediction tools utilizing readily available health screening data.

Purpose of the Study:

• To develop and evaluate DM prediction models using accessible health screening parameters.
• To assess the performance of models based on comprehensive and simplified variable sets.

Main Methods:

• Eight DM prediction models were developed using gradient boosting and random forest algorithms.
• Two variable sets were utilized: 62 variables from a tertiary hospital and 27 from national checkups.
• Internal validation was conducted using stratified 10-fold cross-validation.

Main Results:

• The 62-variable model achieved the highest prediction performance (ROC-AUC of 0.928) for 12-month DM prediction.
• Models using the simplified 27-variable set demonstrated good performance (ROC-AUCs 0.842-0.880).
• Including fasting glucose improved accuracy by up to 11.5%.

Conclusions:

• Easily applicable DM prediction models with good performance were created using common health screening parameters.
• The developed models are suitable for use in tertiary university hospitals and national health checkups.
• Prospective external validation is planned to facilitate widespread clinical adoption.