Analysis of composite immune-inflam matory indices using interpretable machine learning for identifying prevalent metabolic dysfunction-associated steatotic liver disease in patients with diabetes
Chunyao Li, Yaotan Li, Yuchen Fu, Huijuan Zheng, Jiale Zhang, Xinghua Zhang, Shunxuan Xue, Rui Zhou, Haijiao Zhao, Weiwei SunBackground
Diabetes mellitus (DM) and metabolic dysfunction-associated steatotic liver disease (MASLD) share insulin resistance, chronic inflammation, and dysregulated lipid metabolism. However, associations of composite immune-inflammatory indices with prevalent MASLD in DM remain incompletely characterized.
Objectives
To assess 13 immune-inflammatory indices in relation to United States Fatty Liver Index (USFLI)-defined prevalent MASLD in adults with DM and their contributions to an interpretable classification framework.
Design
Cross-sectional analysis of 2007-2018 National Health and Nutrition Examination Survey data.
Methods
We included 951 adults with DM. Survey-weighted logistic regression, false-discovery-rate-corrected quartile analyses, and restricted cubic splines were performed; propensity score matching (PSM) served as a sensitivity analysis. Exploratory unweighted classification compared generalized linear model, random forest, support vector machine, and extreme gradient boosting (XGBoost). The final XGBoost model was interpreted using Shapley additive explanations (SHAP), and a simplified nomogram was internally evaluated.
Results
Neutrophil-to-high-density lipoprotein cholesterol ratio (NHR), monocyte-to-high-density lipoprotein cholesterol ratio (MHR), lymphocyte-to-high-density lipoprotein cholesterol ratio (LHR), and platelet-to-high-density lipoprotein cholesterol ratio (PHR) showed the most consistent evidence, with odds ratios of 1.921, 1.926, 1.930, and 1.811, respectively. NHR and LHR showed nonlinear associations, whereas MHR and PHR were consistent with linear relationships. PSM yielded 247 matched pairs with directionally consistent associations. XGBoost showed the highest mean cross-validated discrimination; in the held-out test set, in which MASLD prevalence was 72.0%, the areas under the receiver operating characteristic and precision-recall curves were 0.801 and 0.905, respectively. SHAP identified body mass index (BMI) and triglycerides as the largest contributors, with BMI × NHR and BMI × MHR as leading candidate non-additive patterns within the fitted model. The nomogram achieved corresponding values of 0.828 and 0.921 and positive net benefit across a broad range of thresholds.
Conclusion
NHR, MHR, LHR, and PHR were consistently associated with USFLI-defined prevalent MASLD among adults with DM. These accessible indices and interpretable tools may complement characterization and identification of MASLD.