AN EXPLAINABLE MACHINE LEARNING APPROACH FOR NON-INVASIVE PREDICTION OF METABOLIC DYSFUNCTION-ASSOCIATED STEATOTIC LIVER DISEASE USING ROUTINE LABORATORY BIOMARKERS

Ipek Balikci Cicek*, Zeynep Kucukakcali

ABSTRACT

Background: Metabolic dysfunction-associated steatotic liver disease (MASLD) is the leading cause of chronic liver disease worldwide, yet population-level screening relies largely on imaging unavailable in primary care. We aimed to develop and explain machine learning models that predict MASLD using routine laboratory biomarkers alone. Methods: We analyzed 8,008 adults from the National Health and Nutrition Examination Survey (NHANES) 2017-March 2020 cycle with valid vibration-controlled transient elastography (VCTE) measurements. MASLD was defined per the 2023 AASLD/EASL/ALEH consensus nomenclature as hepatic steatosis (controlled attenuation parameter [CAP] ≥248 dB/m, with ≥285 dB/m as a sensitivity threshold) plus at least one of five cardiometabolic risk criteria, after excluding participants with significant alcohol consumption. To avoid circularity, predictors were restricted to variables not used in outcome construction: age, sex, race/ethnicity, income-poverty ratio, ALT, AST, GGT, the AST/ALT ratio, albumin, creatinine, uric acid, and C-reactive protein. Five classifiers (logistic regression, random forest, XGBoost, LightGBM, CatBoost) were compared using nested cross-validation (5-fold outer, 3-fold inner) for hyperparameter tuning and unbiased performance estimation. Model behavior was explained using SHAP (global importance, dependence, and interaction values) and LIME (instance-level explanations). As a methodological contribution, we quantified (1) per-patient agreement between SHAP and LIME explanations (Spearman correlation) and (2) cross-model stability of SHAP-based feature rankings (Kendall's tau) among the four tree-based models. Results: MASLD prevalence was 56.0% (CAP≥248 dB/m) and 36.4% (CAP≥285 dB/m). CatBoost achieved the highest nested cross-validation performance (ROC-AUC 0.786 ± 0.007), closely followed by random forest, XGBoost and LightGBM (all ROC-AUC 0.78-0.79 on the held-out test set); logistic regression trailed at 0.752, indicating a modest but genuine benefit from non-linear modeling. Results were consistent, and slightly stronger, under the more stringent CAP≥285 dB/m definition (CatBoost ROC-AUC 0.798). The AST/ALT ratio was the dominant predictor across all five models, showing a sharp transition around a ratio of 1 that mirrors known hepatological patterns (ratio <1 associated with simple steatosis, >1 with more advanced disease), followed by age, CRP, uric acid, and GGT. Feature-ranking agreement among the four tree-based models was high (mean Kendall's tau = 0.86), and SHAP-LIME explanation agreement was high for most patients (mean Spearman ρ = 0.84) but diverged for a small subgroup, warranting caution in individual-level decision support. Conclusions: Routine, widely available laboratory biomarkers, combined with explainable machine learning, can identify adults with likely MASLD with fair-to-good discrimination and clinically coherent explanations, offering a low-cost pre-screening tool to prioritize patients for confirmatory imaging.

Keywords: MASLD; NAFLD; explainable AI; SHAP; LIME; NHANES; machine learning; liver steatosis.