Multi-Metric Evaluation of Machine Learning Algorithms for Diabetes Prediction Using Feature Importance and ROC Analysis

Abstract

Diabetes mellitus has become a major global health threat, and many undiagnosed cases remain undetected due to some limitations of the conventional diagnostic methods. Despite the promising results of machine learning (ML) for early diabetes diagnosis, the majority of the current research assessing algorithms either uses insufficient metrics or does not follow a consistent assessment approach. This paper addresses that gap by utilising an integrated evaluation framework. The framework includes feature importance analysis, Pearson correlation assessment, confusion matrix decomposition, and ROC-AUC comparison. It applies this framework to the Pima Indians Diabetes Dataset (mde) and four popular ML classification algorithms: Naive Bayes, Decision Tree, Random Forest, and Logistic Regression. The most significant predictors, according to our feature analysis, were glucose (27.6%), body mass index (16.0%), age (12.7%), and diabetes pedigree function (12.7%). Among the classifiers, Random Forest exhibited the greatest accuracy (76.0%) and precision (68.1%), Naive Bayes the best recall (64.8%), and Logistic Regression the highest AUC-ROC (82.3%). For patients at high risk, the models' virtual projections across all three risk profiles were in agreement. Model selection should be determined by the unique clinical screening aim, since these findings suggest that there is no one better universal method. Random Forest and Logistic Regression are the most promising for assisting in preliminary diabetes prediction, although further validation on diversity datasets is needed prior to clinical deployment.

GS Cited Analysis