Optimization of diabetes prediction methods based on combinatorial balancing algorithm.

Area of Science:

• Machine Learning
• Data Science
• Medical Informatics

Background:

• Diabetes mellitus is a major public health concern requiring early detection.
• Imbalanced datasets hinder accurate diabetes prediction model performance, particularly for minority classes.

Purpose of the Study:

• To improve diabetes prediction accuracy and model efficiency using data balancing and hyperparameter optimization.
• To address limitations in current research regarding imbalanced data handling in predictive models.

Main Methods:

• Applied Synthetic Minority Over-sampling Technique (SMOTE) and Random Under-Sampling (RUS) for imbalanced diabetes dataset balancing.
• Utilized Optuna for hyperparameter optimization of the LightGBM machine learning model.
• Evaluated method effectiveness by comparing model performance before and after data balancing.

Main Results:

• The optimized LightGBM-Optuna model achieved a slight accuracy increase from 97.07% to 97.11%.
• Precision significantly improved from 97.17% to 98.99% with the proposed methods.
• Hyperparameter optimization using Optuna required only 2.5 seconds per search, demonstrating computational efficiency.

Conclusions:

• Combining SMOTE and RUS with Optuna effectively enhances machine learning models for imbalanced diabetes datasets.
• The proposed approach improves predictive performance and efficiency in diabetes prediction tasks.