An Explainable Machine Learning Approach for Early Dengue Detection Using Routine Hematological Parameters

Abstract

Gallstone disease remains a common biliary disorder with substantial diagnostic and clinical consequences, creating a need for practical tools that can support early risk identification. This study developed and internally validated an artificial intelligence-based clinical decision-support model for gallstone disease using a publicly available de-identified dataset of 319 patients and 38 candidate predictors. Demographic, comorbidity, anthropometric, laboratory, and body-composition variables were evaluated through elastic-net logistic regression, random forest, gradient boosting, and radial-basis-function support vector machine models using repeated stratified five-fold cross-validation. Model performance was assessed through receiver operating characteristic area under the curve, accuracy, sensitivity, specificity, precision, F1 score, Brier score, calibration measures, risk stratification, permutation importance, and decision-curve analysis. The extended elastic-net model was selected as the final model because it combined interpretability with good discrimination and calibration, achieving an area under the curve of 0.848, accuracy of 78.1%, sensitivity of 72.8%, specificity of 83.2%, and Brier score of 0.161. At an optimized threshold of 0.393, sensitivity increased to 85.4% while specificity remained 72.0%. Vitamin D, C-reactive protein, bone mass, extracellular fluid distribution, and visceral fat area were influential predictors. The model showed meaningful risk separation and potential net clinical benefit, but external multicenter and prospective validation remain necessary before routine implementation across diverse clinical settings and patient populations.