Early Prediction of Diabetes Using an Ensemble of Machine Learning Models.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health Data Science

Background:

• Diabetes is a rapidly spreading global disease with severe complications, increasing morbidity and mortality.
• Early diagnosis of diabetes is crucial for reducing severity and risk factors.
• Challenges in diabetes prediction include a shortage of labeled data and issues like outliers and missing data.

Purpose of the Study:

• To introduce a newly labeled diabetes dataset from Bangladesh.
• To propose an automated machine learning classification pipeline for early diabetes prediction.
• To enhance the reliability and effectiveness of diabetes prediction models.

Main Methods:

• Development of a weighted ensemble of machine learning classifiers (Naive Bayes, Random Forest, Decision Tree, XGBoost, LightGBM).
• Implementation of grid search for hyperparameter optimization.
• Inclusion of missing value imputation, feature selection, and K-fold cross-validation.

Main Results:

• The weighted ensemble model (Decision Tree + Random Forest + XGBoost + LightGBM) combined with preprocessing achieved an accuracy of 0.735 and an AUC of 0.832.
• Statistical imputation and Random Forest-based feature selection yielded optimal results for early diabetes prediction.
• The proposed pipeline significantly improved diabetes prediction performance.

Conclusions:

• The newly introduced dataset from Bangladesh can aid in developing robust machine learning models for diabetes prediction.
• The automated classification pipeline, particularly the weighted ensemble, demonstrates significant potential for early and accurate diabetes detection.
• The study highlights the effectiveness of integrated preprocessing techniques and ensemble methods in overcoming data challenges for diabetes prediction.