BENCHMARKING MACHINE LEARNING CLASSIFIERS FOR HOSPITAL READMISSION PREDICTION IN DIABETES CARE
TL;DR
A comprehensive benchmarking analysis of supervised machine learning classifiers for diabetic patients suggests that no one classifier was superior in all the evaluation measures, which illustrates the need for multi-metric evaluation in designing predictive models for imbalanced healthcare datasets.
Abstract
Hospital readmission is a serious problem in healthcare systems as it leads to higher treatment costs, resource use and patient morbidity. Identifying patients who are likely to be readmitted can help guide prompt action by clinical staff and enhance health outcomes. In this study, a comprehensive benchmarking analysis of supervised machine learning classifiers for diabetic patients is presented. The data pre-processing steps comprised of missing value handling, drop of identifier attributes, categorical feature encoding, and binary target generation. In order to overcome class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was used only for the training set, and the original test set was kept imbalanced to evaluate the algorithm in an unbiased manner. The performance of eight machine learning classifiers Logistic Regression, Decision Tree, Random Forest, Extra Trees, AdaBoost, Gradient Boosting, XGBoost and LightGBM was assessed. The experimental results show the superiority of ensemble learning approaches over the conventional classifiers in predicting hospital readmission. Extra Trees with 86.54% classification accuracy obtained the best ROC AUC 0.6079 among models and with the training time of 1.87 s, it shows a balance between the predictive power and computational efficiency. Random Forest had the highest precision (21.58%), AdaBoost had the highest recall (33.42%) and F1 score (0.1919), Logistic Regression had the highest balanced accuracy (53.35%) and XGBoost had the highest Matthews correlation coefficient (0.0759). All the findings suggest that no one classifier was superior in all the evaluation measures, which illustrates the need for multi-metric evaluation in designing predictive models for imbalanced healthcare datasets. This benchmark study offers an experimental framework and hands-on guidance on selecting appropriate machine learning models for hospital-readmission predictions in diabetes management.