Nutritional determinants of non-alcoholic fatty liver disease: a machine learning approach with nutritional scoring

Background and aims Non-alcoholic fatty liver disease (NAFLD) affects roughly a quarter of adults worldwide, and diet is considered a modifiable determinant. Models that include the variables defining NAFLD may overstate the contribution of diet, an effect that has not been quantified. We examined associations of diet, fast-food and soda consumption, and physical inactivity with NAFLD in Korean adults, compared machine learning classifiers under leakage-controlled conditions, and evaluated a Nutritional Liver Risk Score (NLRS). Methods We analysed pooled 2008–2011 data from the Korea National Health and Nutrition Examination Survey (KNHANES). After removing 12,604 duplicated records and applying exclusions, 18,506 adults remained. NAFLD was defined as a hepatic steatosis index (HSI) above 36; alanine aminotransferase (ALT) served as a second, BMI-independent outcome. Diet was assessed by one 24-h recall. Analyses comprised non-parametric tests, correlations, survey-weighted regression, effect modification, E-values, measurement-error scenarios, and eight tuned classifiers evaluated by nested 10-fold cross-validation grouped by primary sampling unit across feature sets progressively excluding the HSI components. The NLRS was validated temporally. Results NAFLD prevalence was 20.9%. Models including the HSI components reached an area under the receiver operating characteristic curve (AUROC) of 0.998, reflecting circularity; excluding them gave 0.897, and 0.747 without waist circumference and glucose. Diet with age and lifestyle gave an AUROC of 0.595 versus 0.597 for age and lifestyle alone. Fibre, sodium and potassium showed small positive associations (adjusted odds ratio [aOR] 1.03–1.06); fat and carbohydrate showed none. Soda consumption was consistently associated with NAFLD (aOR 1.11 per SD) and physical inactivity with elevated ALT (aOR 1.20, E-value 1.4), confined to men; fast food showed a weak association. Continuous ALT differed by 3% or less. The NLRS did not discriminate NAFLD in temporal validation (AUROC 0.512). Conclusion Diet showed small associations with HSI-defined NAFLD and added little to prediction beyond age and lifestyle; near-perfect accuracy of models containing the HSI components is circular. Soda consumption and physical inactivity were the most consistent associations, but the cross-sectional design cannot establish causation. The feature-set framework and open code offer a template for leakage-controlled evaluation.