Beyond Binary Cutoffs: An Explainable Machine Learning Framework for Individualized Diagnostic Reasoning in Suspected...

Background: Emergency department evaluation of suspected urolithiasis increasingly relies on non-contrast CT, yet not all patients require imaging. Existing clinical prediction rules help stratify stone probability, but by converting continuous measurements into fixed binary indicators, they offer little insight into why a particular patient is at risk or how much uncertainty remains after each testing stage-questions that bear directly on individualized diagnostic decisions. Methods: We retrospectively analyzed 1000 ED patients with suspected urolithiasis who underwent non-contrast CT (stone prevalence 85.0%). A gradient boosting classifier was trained on 17 continuous clinical and laboratory features and compared against binary-thresholded counterparts and an established scoring system; the 17-feature model achieved AUC 0.771 (95% CI 0.726-0.813) versus 0.723 (95% CI 0.675-0.771) for the reference score on this cohort (DeLong p = 0.001). Individual predictions were explained using an interventional Shapley value approach, and a Shannon entropy-based framework was applied to quantify the marginal diagnostic contribution of each sequential testing stage. Results: Held-out permutation importance identified red blood cell count on microscopy, age, pain duration, and prior stone history as the most influential predictors. Several features showed non-linear contributions that diverged from conventional binary thresholds: creatinine effect crossed zero near 0.90 mg/dL and pain duration peaked between 2 and 5 h. C-reactive protein, absent from existing scoring systems, emerged as a meaningful negative predictor. Sequential entropy analysis showed that dipstick urinalysis provided the largest marginal information gain among non-history stages (6.1% of prior entropy), while physical examination contributed 2.3%. A prevalence sensitivity analysis projected that the framework's threshold behavior would differ substantially in lower-prevalence populations, underscoring that the cohort-specific cut-points are not portable decision rules. We therefore position the framework as a reasoning aid that complements clinical judgment and imaging, not as a stand-alone triage tool. Conclusions: Explainable machine learning can address questions that aggregate discrimination metrics cannot: which features drive risk for a given patient, how those effects behave across the continuous measurement range, and how much diagnostic uncertainty each testing stage resolves. The Shapley-based explanations and entropy framework developed here offer a structured approach to individualized diagnostic reasoning in the ED evaluation of suspected urolithiasis, functioning as an interpretive adjunct to, rather than a replacement for, existing clinical tools and CT imaging.