Improving Diabetes Prediction Accuracy Using Ensemble Machine Learning Models

Abstract

This study investigates prediction of HbA1c level which is a principal biomarker of diabetes control based on patient biographical and health data from a publicly accessible dataset [1]. I tried regression models like Linear Regression [2], Decision Tree Regressor [3], and Random Forest Regressor [4] to predict accurate HbA1c levels. Upon facing poorly performing models, most likely because of data bias and feature insufficiency, I restructured the task as a classification problem by approximating the ranges of HbA1c levels into significant categories. I implemented models including Random Forest Classifier [5], Decision Tree Classifier [6],  K-Nearest Neighbors [7], and an ensemble Voting Classifier  [8]. The Voting Classifier increased the best accuracy to 72.5%, improving over Random Forest’s standalone accuracy of 68.1% [5]. Model tuning focused on parameters such as the number of trees and maximum depth. Variance Inflation Factor analysis was executed to evaluate feature multicollinearity and it confirmed that multicollinearity was not a major issue. Results show that classification models are more suitable for this dataset and confirm the importance of feature engineering and hyperparameter adjustment. This finding demonstrates that classification models better suit this dataset, showing how predictive instruments can assist medical personnel in approximating HbA1c values without resorting to decisions purely based on costly or time-consuming laboratory testing.