Machine Learning Models Flag Nonobese MASLD Using Routine Blood Tests

Lean patients with metabolic dysfunction-associated steatotic liver disease (MASLD) are often missed because clinicians still associate fatty liver with obesity, but 2 machine learning models built from 8 routine clinical measurements identified nonobese MASLD with discrimination at least as strong as existing clinical scores, according to a study published in Frontiers in Public Health.¹

Lean MASLD Remains an Under-Recognized Diagnosis

MASLD, formerly known as nonalcoholic fatty liver disease, affects more than 30% of adults worldwide, making it the most common chronic liver disease. While strongly tied to obesity, MASLD also occurs in an estimated 15% to 20% of individuals with a normal body mass index (BMI), where it often presents with more pronounced insulin resistance, more severe dyslipidemia, and a higher risk of progressing to metabolic dysfunction-associated steatohepatitis (MASH) and significant fibrosis. Liver biopsy, the histological gold standard for diagnosis, is too invasive and costly for population-level screening, and standard abdominal ultrasonography has limited sensitivity for mild steatosis; even a widely used noninvasive score, the Fatty Liver Index, has shown a sensitivity as low as 6.3% for detecting imaging-confirmed steatosis in lean individuals.

A separate systematic review of 95 studies across 6 databases found that lean individuals make up about 19% of the overall MASLD population and nonobese individuals about 41%, with rates in the general population ranging from under 25% in some countries to more than 50% in others. That review also found lean MASLD carries risks of liver decompensation, cancer, and cardiovascular disease similar to those in patients with obesity, and some data suggest even higher mortality.²

The researchers analyzed data from 2223 adults recruited through community health screenings at a hospital in Qingdao, China, excluding those with a BMI of 25 kg/m² or higher, other liver diseases, significant alcohol use, or incomplete data, leaving 768 participants with a BMI under 25 for the final analysis. MASLD was defined per the 2023 multisociety Delphi consensus as hepatic steatosis, measured with a FibroScan-derived controlled attenuation parameter of 248 dB/m or higher, plus at least 1 metabolic risk factor; 226 participants (29.4%) met that definition.

The cohort, 74.5% female with a mean age of 50.6 years, was split into a training set (n = 615) and an independent testing set (n = 153). After univariable screening and LASSO regression narrowed dozens of candidate variables down to 8, the researchers built 8 machine learning models—linear discriminant analysis (LDA), gradient boosting machine (GBM), neural network, logistic regression, support vector machine, naive Bayes, random forest, and decision tree—using apolipoprotein B, triglycerides, fasting insulin, fasting plasma glucose, waist circumference, age, alkaline phosphatase, and systolic blood pressure as predictors.

Gradient Boosting, LDA Models Best Flag Nonobese MASLD

On the independent test set, GBM achieved the highest discrimination (area under the curve [AUC], 0.779; 95% CI, 0.695-0.863), followed closely by the neural network (0.777) and LDA (0.771). GBM and LDA also showed acceptable calibration (Hosmer-Lemeshow P = .502 and P = .609, respectively) and the strongest clinical utility on decision curve analysis, while the random forest model, despite a perfect AUC of 1.000 on the training set, dropped to 0.742 on testing, a sign of the overfitting that more complex models can be prone to in modest-sized clinical datasets.

SHAP (Shapley additive explanations) analysis showed fasting insulin was the strongest driver of the LDA model's identifications, while triglycerides dominated the GBM model, highlighting “the central roles of insulin resistance and dyslipidaemia” in non-obese MASLD, the researchers wrote. Waist circumference ranked among the top features in both models, reinforcing that visceral fat, not overall body weight, appears to drive metabolic risk in this population.

An Imperfect Screening Tool for Nonobese MASLD

At their default thresholds, the top-performing models correctly flagged only 33% to 40% of confirmed MASLD cases in the test set, meaning a majority of cases would still be missed without further threshold adjustment. The study was also conducted at a single center in China, diagnosed MASLD using a noninvasive ultrasound-based elastography measurement rather than liver biopsy, and included some participants already being treated for diabetes or hypertension, medications that could have blunted the very biomarkers the models rely on.

The researchers said external validation in larger, more diverse cohorts, along with threshold optimization, would be needed before the models could move toward clinical or population-screening use. They also noted that future work should incorporate dietary, activity, and genetic data to further sharpen accuracy.

For clinicians, the findings reinforce that a normal BMI should not rule out fatty liver disease, particularly in patients with elevated triglycerides, insulin resistance, or a larger waist circumference despite a lean frame. For health systems and payers, a screening tool built entirely from data already collected during routine blood work and vital signs, rather than imaging or biopsy, could offer a low-cost way to flag non-obese patients for further liver evaluation, though the models' modest sensitivity means they are better suited to supplementing clinical judgment than replacing it.

References

1. Yu H, Han X, Zhang N, et al. The application of explainable machine learning for the identification of MASLD in non-obese populations. Front Public Health. 2026;14:1851486. doi:10.3389/fpubh.2026.1851486
2. Njei B, Ameyaw P, Al-Ajlouni Y, Njei LP, Boateng S. Diagnosis and management of lean metabolic dysfunction-associated steatotic liver disease (MASLD): a systematic review. Cureus. 2024;16(10):e71451. doi:10.7759/cureus.71451