Feature importance and model performance for prediabetes prediction: A comparative study

Abstract

Objectives: Prediabetes is a significant health condition that elevates the risk of developing type 2 diabetes and other associated complications. This study aims to (1) explore the potential of machine learning models to improve the prediction of prediabetes, (2) compare the performance of various machine learning models with traditional regression methods, and (3) identify the most influential demographic, socioeconomic, and health-related factors associated with prediabetes. Methods: This study utilized data from the 2021 Behavioral Risk Factor Surveillance System (BRFSS) and employed comprehensive data preprocessing techniques. Logistic regression analysis was conducted to assess correlations between features and prediabetes risk. Feature importance was quantified using Adjusted Mutual Information values. Multiple machine learning models, including Random Forest, K Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Neural Network, and Logistic Regression, were used for prediction. The best model was selected and validated through cross-validation to ensure robustness. Results: Significant associations were observed between prediabetes and key predictors such as cholesterol levels, BMI categories, hypertension status, age groups, and income categories. Among the models tested, Random Forest demonstrated the highest accuracy and robustness, outperforming traditional regression models. Conclusions: This study highlights the potential of machine learning to enhance prediabetes prediction and underscores the importance of identifying high-risk individuals for early intervention. The findings contribute to population health strategies by integrating advanced analytical methods with public health data.

Keywords

• Adjusted mutual information
• Machine learning models
• Multivariate logistic regression
• Prediabetes
• Risk factors