Association between Helicobacter pylori infection and MASLD: a cross-sectional study with interpretable machine...

Helicobacter pylori (HP) infection has been increasingly linked to metabolic dysfunction-associated steatotic liver disease (MASLD). However, evidence remains inconsistent, and the predictive value of HP status within integrated risk models is unclear. This study aimed to evaluate the association between HP infection and MASLD and to develop interpretable machine learning models for MASLD risk stratification.

A total of 1,021 participants from a Chinese hospital-based cohort and 4,870 participants from the U.S. National Health and Nutrition Examination Survey (NHANES 1999-2000) were included. In the Chinese cohort, active HP infection was assessed using the 13C or 14C-urea breath test, and MASLD was diagnosed by ultrasonography. In NHANES, HP exposure was defined by serum IgG seropositivity, and MASLD was approximated using the Fatty Liver Index (FLI). Logistic regression was used to assess the association between HP and MASLD. Mediation analysis was conducted to explore potential metabolic and inflammatory pathways. Eleven machine learning models were developed in the Chinese cohort and evaluated using ROC, calibration, and SHAP analysis, with external comparison in NHANES using a reduced feature set to minimize potential bias.

In the Chinese cohort, HP infection was significantly associated with MASLD in crude (OR = 7.11), age- and sex-adjusted (OR = 6.48), and further adjusted models (including diabetes, cholesterolemia, and BMI; OR = 4.83, 95% CI: 3.07-7.23, p < 0.001). This elevated OR should be interpreted as cohort-specific and not generalizable. A positive association was also observed in NHANES (adjusted OR = 1.22, 95% CI: 1.14-1.59, p = 0.017). Machine learning models demonstrated strong internal performance (AUC up to 0.8876), while external comparison in NHANES showed moderate discrimination (best AUC = 0.6536). SHAP analysis identified HP and LHR as the top contributors to MASLD prediction, with TyG-related indices showing secondary importance. Mediation analysis suggested potential involvement of metabolic and inflammatory indices, although these findings should be interpreted as exploratory.

HP infection was associated with MASLD in both Chinese and U.S. populations. Interpretable machine learning models demonstrated potential for MASLD risk stratification, with consistent identification of key predictors across cohorts. However, due to cross-sectional design, differences in diagnostic definitions, and the use of FLI in NHANES, findings should be interpreted as associative rather than causal, and external validation results should be considered exploratory.