Predicting Helicobacter pylori positivity in a male health screening cohort using interpretable machine learning

Routinely collected health-examination variables may contain information associated with Helicobacter pylori ( H. pylori ) positivity, but any prediction model must demonstrate improvement over simple alternatives and must not be presented as a substitute for direct testing. We evaluated multivariate adaptive regression splines (MARS) and conventional logistic models in a male health-screening cohort. H. pylori positivity was defined as an observed 13 C-urea breath test value of at least 4.0. Age-only, age-plus-white-blood-cell-count (WBC), full logistic, LASSO logistic, and MARS models were evaluated using nested five-fold cross-validation with fold-specific imputation, scaling, and tuning. Performance was estimated from pooled out-of-fold predictions using area under the receiver operating characteristic curve (AUC), precision-recall AUC (PR-AUC), Brier score, log loss, calibration, and threshold-specific metrics. Descriptive comparisons used non-parametric tests, standardized mean differences, age-adjusted regressions, and false-discovery-rate correction. Complete-case and multiple-imputation analyses assessed missing-data sensitivity. Among 4,583 men, 1,059 (23.1%) were H. pylori -positive. Out-of-fold AUCs were 0.638 for age only, 0.647 for age plus WBC, 0.654 for full logistic regression, 0.650 for LASSO, and 0.646 (95% bootstrap CI 0.625–0.665) for MARS. MARS had a PR-AUC of 0.329 and Brier score of 0.170 and did not significantly outperform any comparator by paired DeLong testing after false-discovery-rate correction. At the pooled out-of-fold Youden threshold of 0.217, MARS sensitivity was 0.669, specificity 0.561, and positive predictive value 0.314. The final additive MARS model retained age, WBC, and drinking score. Complete-case and multiple-imputation analyses did not materially improve predictive performance. Exploratory prediction of the continuous breath-test value was weak (log-scale R-squared 0.054). Routinely collected variables showed only modest ability to discriminate H. pylori positivity. MARS did not improve on simpler models and is not suitable as a stand-alone screening or diagnostic tool. The findings should be considered hypothesis-generating and require external evaluation before any clinical application. Not applicable.