ANOVA–Mutual Information Feature Selection with CatBoost for Gallstone Disease Classification

Abstract

Gallstone disease is a common biliary disorder for
which non-invasive, ultrasonography-independent risk assessment
may support clinical decision-making. This study proposes
an ANOVA–Mutual Information (ANOVA–MI) feature-selection
framework combined with CatBoost, evaluated on 319 patients
and 38 candidate predictors from the UCI Gallstone dataset. The
framework fuses min–max normalized ANOVA F-statistics and
mutual-information scores through a tunable weight α, ranks
features, and selects compact subsets inside a nested repeated
stratified cross-validation protocol (5×5 outer folds, 3-fold inner
Optuna tuning) to prevent information leakage. At α = 0.3,
the method retains 15 features (60.5% dimensionality reduction)
and achieves an F1-score of 0.800 ± 0.021 and ROC-AUC of
0.861±0.032, slightly outperforming the full 38-feature CatBoost
baseline as well as ANOVA-only and MI-only selection. The
corresponding sensitivity and specificity are 78.2% and 82.6%,
respectively. Compared with three recent studies on the same
dataset, the proposed approach attains the highest recall and
F1-score while using fewer or a comparable number of features.
SHAP analysis identifies C-reactive protein and vitamin D as the
dominant predictors. These results demonstrate that a modelagnostic
hybrid of statistical discrimination and informationtheoretic
relevance can substantially reduce feature dimensionality
while preserving competitive predictive performance under
rigorous nested evaluation.