A novel method for predicting gallstones based on ensemble feature selection method

Abstract Gallstone disease is a major health problem worldwide, affecting millions of people. Although conventional diagnostic methods such as magnetic resonance imaging and endoscopic ultrasound are still in use, they are expensive as well as require significant expertise, and can produce variable results. In this study, we introduce an artifial intelligence based approach to predict gallstone disease using bioimpedance measurements and laboratory biomarkers as an alternative to imaging techniques. The dataset comprises 38 parameters, including demographic information, biochemical indicators such as glucose, lipid profiles, triglycerides, and bioimpedance data. We introduce a novel ensemble feature selection framework that combines four statistical methods of minimum redundancy maximum relevance, chi-square test, analysis of variance, and Kruskal–Wallis with four metaheuristic algorithms of grey wolf optimization, whale optimization algorithm, Harris hawk optimization, and particle swarm optimization for the purpose of enhancing predictive performance. The study findings denote that the features retained commonly by seven of the eight methods are deemed the most discriminative. Next, we evaluate the selected features using several classifiers, including logistic regression, Naive Bayes, support vector machine, decision tree, k-nearest neighbor, gradient boosting, Adaboost, random forest, and multilayer perceptron. Experimental results show that eliminating redundant features significantly improves model accuracy, where the random forest classifier achieved the highest accuracy rate of 90.62% using only 11 features instead of 38. The proposed ensemble method reduces computational complexity, enhances classification accuracy, and provides systematic evidence of how feature selection impacts different classifiers. Its superiority is also demonstrated through comparisons with state-of-the-art approaches for gallstone prediction.