Predicting Diabetes in Patients with Metabolic Syndrome Using Machine-Learning Model Based on Multiple Years' Data.

Area of Science:

• Medical Informatics
• Machine Learning
• Endocrinology

Background:

• Metabolic syndrome (MetS) is a precursor to type 2 diabetes.
• Accurate prediction of incident diabetes in MetS patients is crucial for timely intervention.
• Existing prediction models may not fully leverage longitudinal health data.

Purpose of the Study:

• To evaluate machine learning models for predicting incident diabetes in MetS patients.
• To assess the impact of using multiple years of continuous health data on prediction performance.
• To identify key predictive features for diabetes onset in this population.

Main Methods:

• Utilized health records from 4510 nondiabetic MetS patients (2008-2020).
• Developed logistic regression, random forest, and Xgboost models using single-year and multiple-year data (1-3 years).
• Defined MetS using International Diabetes Federation (IDF) criteria; tracked incident diabetes over 7±1.4 years.

Main Results:

• Model performance, measured by AUROC, improved with increased longitudinal data duration.
• Random forest (1-3 years: 0.893) and Xgboost (1-3 years: 0.897) showed superior predictive power with multi-year data.
• Key predictors included fasting plasma glucose, HbA1c, BMI, and "delta weight" (yearly weight change).

Conclusions:

• Longitudinal data enhances machine learning model performance for diabetes prediction in MetS.
• Parameter variation trends, not just levels, are vital for personalized diabetes risk assessment.
• Multi-year data-driven models offer improved tools for evaluating future diabetes risk.