Optimized feature selection for diabetes prediction using IRIME

Abstract

Early prediction of diabetes is critical for preventive care, but modern datasets often contain many features that can hinder model performance through overfitting and increased complexity. In this paper, we investigate the effectiveness of IRIME, an improved version of RIME which is a novel computation intelligence algorithm inspired by rime-ice, for optimized feature selection in diabetes prediction. We compare IRIME's performance against using all available features and against a custom ensemble feature selection model. This ensemble aggregates the outputs of several computational intelligence algorithms, including Particle Swarm Optimization, Genetic Algorithm, Differential Evolution, Ant Colony Optimization, Brain Storm Optimization, Bat algorithm, and Binary PSO. We apply these methods to a real-world Iraqi diabetes dataset with 1000 patients, using LightGBM and Multilayer Perceptron classifiers to evaluate each selected feature subset. Our results show that the IRIME-selected features achieve an F1-score of 0.990, nearly matching the performance of using all features (F1 ≈0.992) and the custom ensemble model (F1≈0.989) while dramatically reducing the feature count. We interpret the contributions of the selected features using Shapley Additive Explanations to provide insight into their impact on the model's predictions. The IRIME algorithm demonstrates its effectiveness in finding a highly informative feature subset, performing comparably to or even surpassing the custom ensemble and other individual feature selectors in predictive performance. This work demonstrates an efficient, interpretable feature selection framework that improves diabetes prediction and could reduce diagnostic tests.