Hepdecide: A standards-oriented, explainable clinical decision support prototype for the chronic liver-disease spectrum—design, guideline-grounded reporting, and a leakage-controlled retrospective evaluation

Background/Objectives: Specialist hepatology capacity is small relative to the population needing assessment, so most decisions about whom to investigate for chronic liver disease fall to non-specialists guided by static laboratory indices reported in isolation. This study presents HepDecide, a standards-oriented decision support prototype consolidating stage specific risk estimation, explainability and guideline-traceable reporting into one physician facing service, and reports a leakage-controlled retrospective evaluation of its predictive layer. Methods: HepDecide combines stage-specific tabular classifiers, an explainability layer based on SHapley Additive exPlanations (SHAP), and retrieval-augmented genera tion (RAG) anchored to EASL and AASLD guidelines; outputs are expressed as HL7 FHIR R4 resources, with inputs mapped to LOINC and diseases to ICD-10. Three endpoints were modelled on two publicly available, de-identified cohorts: steatohepatitis versus simple steatosis and histological cirrhosis (F4) in a biopsy-characterised MASLD cohort (n = 605), and one-year survival among patients with hepatocellular carcinoma (HCC; n = 165). Because patient-level predictions from the original development runs were not retained, the primary evaluation is a separate re-analysis with a locked patient-wise test partition, imputation, scaling and resampling fitted within training folds only, 25 classifier–imbalance configurations per endpoint, training-only calibration, and calibration, precision–recall, decision-curve and SHAP analyses. Results: On the locked test partitions, ROC-AUC was 0.67 (95% CI 0.50–0.84) for steatohepatitis (n = 121), 0.81 (0.52–1.00) for cirrhosis (n = 121; five events) and 0.76 (0.56–0.92) for one-year HCC survival (n = 33), with Brier scores of 0.096, 0.028 and 0.199. On the same cirrhosis test partition FIB-4 reached an ROC-AUC of 0.92 (0.82–0.99). The five highest-ranked SHAP features accounted for 58–73% of total attribution. Development-phase estimates of 0.90, 0.98 and 0.92 are reported separately as archived values; the cirrhosis estimate is documented in a retained training output, although its original input tables and individual test probabilities are unavailable. No quantitative evaluation of the language-generation layer is claimed, because the corre sponding records were not retained. Conclusions: Under leakage-controlled retrospective conditions the predictive layer showed moderate discrimination with wide confidence intervals and did not outperform FIB-4 for cirrhosis. The value of the prototype lies in its standards-oriented, explainable and guideline-grounded design and in a fully docu mented, reproducible evaluation pipeline; transportability, the text-generation layer and clinical usefulness remain to be established prospectively. HepDecide is intended as a clinician-supervised triage and reporting prototype, not a clinically ready system.