Fractional-Order Soft-Voting Ensemble Framework for Cohort-Level Gallstone Risk Prediction

This study proposes a hybrid framework for predicting gallstone presence versus absence in the UCI clinical cohort by coupling a soft-voting ensemble (Neural Network, Random Forest, and Bagging) with a doubly-pruned seven-state Caputo system. The classifiers are trained on the top-10 features selected by minimum-redundancy maximum-relevance (MRMR), and the ensemble probability is injected as an external forcing of the seven leading markers. On a frozen held-out test set (N=79) the ensemble attains AUC 0.7558, a modest increment over the best single learner (neural network, AUC 0.7500) and over equal-weight voting (AUC 0.7545). Identification of the Caputo system on the order-statistic axis of the top MRMR marker (coronary artery disease), not Age, reconstructs the cohort-level marker curves with training-weighted MSE 1.65×10−4. Existence and uniqueness of solutions are proved; positivity and an a priori bound hold only under sign restrictions that the identified couplings do not satisfy. A prototype graphical interface for exploratory use is provided; it is not a clinically validated tool. All reconstructed trajectories are cohort-level snapshots on a cross-sectional pseudo-time axis and are not individual longitudinal histories.