From IBS Factors to Digital-Twin States: Literature-Guided Feature Grouping, Explainable Machine Learning, and the Limits of Static Prediction

Irritable bowel syndrome (IBS) is shaped by interacting symptom, psychological, behavioral, and biological factors, yet computational studies often examine these factors in isolation. This fragmentation makes it difficult to decide which variables should be prioritized for prediction and which should become state variables in a patient-specific digital twin. We address this problem in three stages. First, we construct a targeted 14-study parameter corpus by combining six diagnostic-AI studies identified in a recent systematic review with eight recent human studies covering treatment response, psychological features, lifestyle, brain measures, and disease burden. Second, we group the reported variables into clinically interpretable domains and reanalyze an open 77-participant IBS dataset using leakage-aware repeated cross-validation. Physical fatigue showed the strongest univariate association with IBS status (r=0.626), followed by mental fatigue (r=0.526), anxiety (r=0.507), depression (r=0.489), and sleep quality (r=0.462). A logistic-regression model using non-severity features achieved mean ROC-AUC 0.893 ± 0.072 across 100 repeated five-fold test folds. Fatigue alone reached AUC 0.878, anxiety/depression 0.848, sleep 0.765, cognition 0.680, and demographics 0.507. Third, we tested whether the same static factors could predict IBS severity among the 49 IBS participants. They could not: the best cross-validated regression remained below the mean baseline (R²=-0.328; MAE=58.9). This contrast is the central result. Factors that discriminate IBS status are not sufficient to forecast an individual's changing symptom burden. We therefore translate the literature and model rankings into a digital-twin design in which symptoms, bowel state, stress, sleep, diet, treatment exposure, and optional biological measurements are updated longitudinally with explicit uncertainty. The study provides an evidence-guided bridge from heterogeneous IBS factor studies to a testable longitudinal patient model without treating cross-sectional associations as causal effects.