How Gradient Boosting Predicts Gallstone Disease Using Routine Health Data

Abdominal ultrasonography remains the first-line diagnostic modality for gallstone disease, yet researchers at Ankara VM Medical Park Hospital evaluated 319 patients to test whether routine health parameters could classify gallstone status without immediate imaging. A retrospective clinical dataset of 319 patients (161 individuals with confirmed gallstones and 158 healthy controls) was analyzed.

Key Clinical Takeaways:

• CatBoost outperformed XGBoost, LightGBM, and HistGBM with an AUC of 0.862 and an accuracy of 0.794.
• Bioimpedance parameters alone achieved an AUC of 0.797, approaching the discriminatory power of routine laboratory markers at 0.829.
• C-reactive protein, Vitamin D, and bone mass emerged as the most influential variables via SHAP analysis.

Gradient Boosting Performance Against Traditional Diagnostic Baselines

Four gradient boosting machine learning algorithms were examined during the retrospective study: XGBoost, LightGBM, CatBoost, and HistGBM. Model performance relied on repeated stratified 10-fold cross-validation, measuring parameters including AUC-ROC, accuracy, F1-score, sensitivity, specificity, Brier score, and decision curve analysis. CatBoost secured the highest discriminatory performance alongside the lowest Brier score of 0.149, establishing statistically significant superiority over competing architectures via the Friedman test at p < 0.001.

Bioimpedance and Laboratory Markers Provide Independent Predictive Signals

Feature ablation testing separated demographic attributes, bioimpedance measurements, and laboratory markers to isolate their individual predictive contributions. Demographic attributes alone proved insufficient for reliable classification, registering an AUC of just 0.476. Conversely, bioimpedance parameters provided substantial independent discriminative signal with an AUC of 0.797, closely trailing laboratory markers which achieved an AUC of 0.829.

SHAP analyses pinpointed specific physiological metrics driving the classifications. C-reactive protein, Vitamin D levels, and bone mass ranked among the most influential variables. The evaluation also revealed non-linear risk associations tied directly to visceral fat area and body mass index across the patient cohort.

Gradient Boosting Models Offer Potential for Pre-Imaging Triage

Decision curve analysis confirmed positive net clinical benefit across threshold probabilities ranging from 0.10 to 0.80. The findings demonstrate that gradient boosting models offer proof-of-concept potential for opportunistic risk stratification in metabolic care settings. Rather than replacing established abdominal ultrasonography, the models provide supplementary utility for pre-imaging triage during ambiguous outpatient presentations.