Machine Learning Models Overstate Dietary Liver Disease Risk Through Circular Calculations

A comprehensive computational evaluation of dietary patterns and liver health among 18,506 Korean adults has demonstrated that machine learning algorithms predicting non-alcoholic fatty liver disease achieve near-perfect statistical accuracy primarily through circular mathematical reasoning rather than genuine nutritional insight. Publishing in Frontiers in Nutrition, an international research team investigated how machine learning models evaluate dietary intakes, lifestyle habits, and metabolic indicators against the hepatic steatosis index, a widely used diagnostic formula. While standard models incorporating full clinical panels registered an extraordinary receiver operating characteristic curve of 0.998, this apparent predictive perfection collapsed systematically once investigators removed the biological measurements that define the index itself. When stripped of these defining components and secondary proxies like waist circumference and fasting glucose, the addition of dietary recall data yielded an area under the curve of 0.595, failing to improve upon a baseline model of age and lifestyle alone at 0.597.

To examine nutritional determinants without falling prey to analytical artefacts, the investigators pooled multi-wave epidemiological records from the 2008 through 2011 cycles of the Korea National Health and Nutrition Examination Survey, administered by the Korea Disease Control and Prevention Agency. The merged file encompassed 39,306 records, from which researchers identified eighteen corrupted entries spanning fourteen participant identifiers that could not be read reliably. Crucially, the examination uncovered that every participant record from the 2010 and 2011 survey waves appeared twice, creating 12,604 duplicate rows that would have distorted cross-validation if uncorrected. After deduplication, 26,684 unique adult respondents remained. Sequential exclusions removed 2,979 individuals lacking required clinical measurements for index calculation, 2,698 lacking dietary recall, 749 testing positive for hepatitis B surface antigen, 1,015 flagged for heavy alcohol use, 63 pregnant individuals, 239 reporting implausible caloric intake under 500 or over 5,000 kilocalories daily, and 435 missing survey design variables, leaving 18,506 adults.

Across the analytic cohort of 18,506 adults, the unweighted prevalence of non-alcoholic fatty liver disease was 20.9 percent with 3,877 identified cases, while survey-weighted population prevalence reached 22.2 percent. The epidemiological patterns exhibited sharp divergence across biological sexes, registering 26.3 percent among men versus 18.7 percent among women. Male prevalence peaked in early adulthood at 35 percent among individuals aged thirty to thirty-nine before tapering steadily, whereas female prevalence ascended progressively to peak at 28 percent among those aged sixty to sixty-nine years. Because liver ultrasound and biopsy were absent from the survey, primary ascertainment utilized the hepatic steatosis index exceeding 36. This formula calculates risk from eight times the alanine aminotransferase to aspartate aminotransferase ratio, body mass index, and diabetes and sex adjustments. To ensure findings were not solely artefacts of body mass calculations, the team also analyzed elevated alanine aminotransferase as a body-mass-independent secondary outcome.

The investigators benchmarked eight machine learning classifiers using nested ten-fold cross-validation grouped by primary sampling unit to prevent data leakage between training and testing folds. Within inner validation loops, hyperparameters were optimized for logistic regression, random forests, XGBoost, LightGBM, multilayer perceptrons, k-nearest neighbours, Gaussian naïve Bayes, and decision trees. Comparing five nested feature sets demonstrated how algorithmic leakage operates. Feature Set A, incorporating full clinical, demographic, and nutritional variables, achieved an area under the curve of 0.998 across gradient boosting and logistic regression because the algorithms directly reconstructed the index formula. Removing the five formula components in Feature Set B reduced discrimination to 0.897, which remained high because waist circumference and blood glucose acted as proxies for adiposity and diabetes. Eliminating waist circumference and glucose in Feature Set C dropped performance to 0.747. Feature Set D, combining diet with age and lifestyle, reached only 0.595, matching Feature Set E containing age and lifestyle alone at 0.597.

To test whether a streamlined dietary scoring tool could identify risk, the researchers developed the Nutritional Liver Risk Score. Using LASSO logistic regression on 9,562 participants from 2008 and 2009, the algorithm evaluated eighteen non-energy nutrient densities standardised per one thousand kilocalories. The model retained all eighteen nutrients, assigning scaled integer weights from zero to ten based on multivariable coefficients, with protein, phosphorus, and carotene receiving the largest weights. When tested on the independent temporal validation cohort of 8,944 participants from 2010 and 2011, the score failed to provide discriminative utility. The validation area under the curve reached only 0.512, underperforming a basic model of age, sex, smoking, and physical activity at 0.524. Across score quintiles, adjusted odds ratios showed no significant gradient, rising non-significantly from 1.05 in quintile two to 1.14 in quintile five, leading the authors to conclude that the dietary score cannot serve as an effective screening instrument.

Multivariable logistic regression models demonstrated that individual nutrient intakes had minimal and often paradoxical associations with liver disease. After adjusting for demographic factors, socioeconomic status, physical activity, and total energy intake, small positive associations emerged for dietary fibre, sodium, and potassium. Specifically, each five-gram daily increment in dietary fibre was associated with an adjusted odds ratio of 1.06, while sodium and potassium intakes showed odds ratios of 1.03 per 1,000 milligrams and 500 milligrams, respectively. Total fat, carbohydrate, vitamin C, water, and baseline caloric intake showed no significant relationships with liver disease. In sensitivity models adjusting for body mass index, the associations for fibre, sodium, and potassium attenuated toward the null, while carbohydrate intake displayed an odds ratio of 1.08. The researchers emphasized that these modest positive associations do not imply biological harm, but rather reflect confounding by total food volume and body mass, as individuals with larger body size consume greater absolute food quantities.

Examining specific dietary habits and physical activity patterns on a standardised scale highlighted soda consumption and physical inactivity as the most consistent behavioural signals. Per standard deviation increase, soda consumption was associated with an eleven percent increase in the odds of hepatic steatosis and a nine percent increase in elevated alanine aminotransferase. Physical inactivity was reported by 47.7 percent of participants and was linked to a borderline nine percent elevation in liver disease odds alongside a twenty percent increase in elevated aminotransferase odds. Interaction analyses revealed pronounced effect modification by biological sex: inactivity was strongly associated with liver disease in men, showing odds ratios of 1.32 for steatosis and 1.35 for elevated aminotransferase, whereas women exhibited no heightened odds at 0.90 and 1.04. Total fat intake as a percentage of energy was unrelated to steatosis at an odds ratio of 0.98, but showed a slight inverse association with elevated aminotransferase at 0.94.

Associations with fast-food intake proved fragile and sensitive to analytical methodology. A composite fast-food index aggregating frequencies of hamburgers, pizza, and fried foods showed a weak positive association with steatosis in weighted models at an odds ratio of 1.09 per standard deviation, but attenuated to 1.05 and crossed unity in unweighted analyses. Crude Spearman correlations between fast-food items and the steatosis index were unexpectedly negative, a reflection of strong age confounding because fast-food consumption was concentrated among young adults. Furthermore, machine learning explainability analyses using Shapley Additive Explanations demonstrated that dietary features accounted for only 7.7 percent of model attribution in the index-free model. Metabolic and anthropometric measures drove 75.1 percent of model predictions, with waist circumference exhibiting dominant influence across decision trees while individual contributions from dietary fibre, sodium, and potassium remained confined within a negligible margin of plus or minus 0.01 log-odds across intake quintiles.

The study exposed substantial exposure measurement errors and reporting biases that complicate population-level nutritional assessments. Comparing reported energy intake against basal metabolic rates calculated using the Mifflin-St Jeor formula revealed that 11.4 percent of participants reported implausibly low caloric intakes below 0.9 times basal requirements. This under-reporting bias differed markedly by health status, occurring in 15.8 percent of participants with liver disease compared to 10.3 percent among unaffected individuals. When researchers restricted analyses to the 15,588 plausible energy reporters, total caloric intake emerged as a clear risk factor with an odds ratio of 1.18 per 500 kilocalories, contrasting with the inert 1.01 observed across the uncorrected population. In this plausible reporting subset, the predictive discrimination of dietary factors combined with age and lifestyle reached an area under the curve of 0.607, slightly higher than the 0.595 attained by age and lifestyle alone, confirming that systematic reporting bias masks dietary effects.

Acknowledging key methodological limitations, the authors cautioned that the cross-sectional survey design cannot determine causality, as reverse causation may prompt individuals aware of liver conditions to alter dietary habits. The public survey database provided only nineteen broad nutrient totals and lacked granular information on saturated fatty acids, omega-3 fatty acids, and added industrial fructose, preventing evaluation of specific lipotoxic pathways. Relying on a single twenty-four-hour dietary recall also introduces within-person measurement error that attenuates observed associations. Ultimately, the investigation demonstrates that exceptional headline accuracies in machine learning models predicting formula-based liver disease often stem from circular evaluations of defining components rather than nutritional insights. The authors concluded that future research must utilize longitudinal cohort designs, detailed fatty acid profiling, and objective imaging techniques rather than formulaic clinical surrogates that generate artificial algorithmic performance.

Subject of Research: Machine learning evaluation of dietary determinants and algorithmic circularity in non-alcoholic fatty liver disease risk prediction

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

Article References: Ghosh, J., Pal, P. P., Dasgupta, P., Chattopadhyay, D., & Nayak, D. (2026). Nutritional determinants of non-alcoholic fatty liver disease: a machine learning approach with nutritional scoring. Frontiers in Nutrition, 13, 1982917. https://doi.org/10.3389/fnut.2026.1982917

Image Credits: AI Generated

DOI: 10.3389/fnut.2026.1982917

Keywords: non-alcoholic fatty liver disease, hepatic steatosis index, machine learning, algorithmic circularity, dietary epidemiology, KNHANES, alanine aminotransferase, physical inactivity, soda consumption, waist circumference, nutritional epidemiology, feature leakage, cross-sectional study, nested cross-validation, energy under-reporting, public health nutrition, metabolic risk factors, Korean adult health