Risk Prediction of Celiac Disease from Routine Clinical Variables: Internal Validation of a Pre-Biopsy Model and Implications for Gluten-Related Food Safety Management

Background: Celiac disease requires lifelong gluten avoidance, and gluten is a regulated food hazard; accurate risk stratification before biopsy could support timely diagnosis. Public clinical datasets are increasingly used to build prediction models, but data quality and validation are often poorly reported. We developed and internally validated a pre-biopsy prediction model and audited a widely used public dataset. Methods: A public dataset of 2,206 records was reduced to 649 unique records after removing rows identical on all variables. Eleven pre-biopsy predictors (age, total immunoglobulins A, G and M, diabetes, symptoms, short stature and sex) were used; histological stage and clinical subtype were excluded as outcome proxies. Logistic regression, decision tree, random forest and gradient boosting were evaluated with 10 × 5-fold stratified cross-validation, with tuning, calibration and threshold selection nested within training folds. Discrimination, calibration, Brier score, decision-curve net benefit and subgroup performance were assessed in Python. Results: Logistic regression showed a ROC-AUC of 0.82 (95% CI 0.78–0.86), an area under the precision–recall curve of 0.93 (prevalence 0.78), a Brier score of 0.101 (skill score 0.42) and a calibration slope of 0.98; gradient boosting performed better (ROC-AUC 0.87). A sensitivity-targeted threshold gave sensitivity 0.91 and specificity 0.61. Diabetes accounted for most of the discrimination (ROC-AUC 0.76 alone, 0.65 without it); discrimination among diabetic patients was modest (0.58). Class weighting and synthetic oversampling did not improve discrimination and worsened calibration. Outcome-proxy variables produced near-perfect but uninformative performance, and retaining duplicated records inflated ROC-AUC to 0.91. Conclusion: Routine variables provide moderate, well-calibrated pre-biopsy discrimination that is largely driven by diabetes. The findings are exploratory, based on a dataset of unverifiable provenance, and require external validation; they do not constitute a quantitative risk assessment of gluten exposure. This is a preprint: it has not been peer reviewed. Files: the preprint (PDF with title page, manuscript, tables and figures); Supplementary File S1 (Tables S1–S6); Supplementary File S2 (code and numerical outputs, ZIP); Supplementary Files S3 and S4 (STROBE and TRIPOD+AI checklists). The source dataset is a public Kaggle file (https://www.kaggle.com/datasets/jackwin07/celiac-disease-coeliac-disease) with no stated licence and unverified provenance; neither the source file nor the de-duplicated analytic file is redistributed here, and the README inside the ZIP explains how to obtain the data and reproduce the analysis.