Authors: Pureti Anusha, M. Radhika
Abstract: Gallstone disease (GSD) is one of the most frequently encountered digestive disorders in adults, and it often stays silent until it produces pain, infection or blockage of the bile ducts. Because its early symptoms overlap with many other abdominal conditions, a timely and confident diagnosis is not always possible, and imaging is not available at every point of care. This work presents a machine learning frameworsk that estimates gallstone risk from routinely recorded demographic, bioelectrical impedance and blood test parameters, and that explains every prediction it makes. A public dataset of 319 patients from the UCI repository is cleaned, scaled and encoded, after which four feature configurations are prepared: all attributes, attributes chosen through a correlation heatmap, attributes ranked by the ANOVA F-test, and the overlap of the last two. Seven classifiers, namely Random Forest, Support Vector Machine, Gradient Boosting, Logistic Regression, AdaBoost, Decision Tree and a Bagging ensemble, are trained under stratified five-fold cross-validation, giving 28 model–configuration pairs in total. The Support Vector Machine trained on correlation-selected features gives the strongest result, with 83.33% accuracy, 85.42% sensitivity and an area under the ROC curve of 88.80%. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are then attached to this model. Both methods point to Vitamin D, C-reactive protein, bone mass, aspartate aminotransferase and hepatic fat grade as the dominant drivers of risk. The outcome is a transparent, low-cost screening aid that can help physicians flag likely gallstone cases early and understand why the system reached each decision.
DOI: https://doi.org/10.5281/zenodo.23595845
