Machine learning models flag nonobese MASLD in a study

Out of 768 participants with a BMI under 25 screened in a community health evaluation, 226 participants—or 29.4%—met the multisociety Delphi consensus definition for metabolic dysfunction-associated steatotic liver disease, according to a study published in Frontiers in Public Health.

Key Clinical Takeaways:
• Gradient boosting machine and linear discriminant analysis models achieved the highest discrimination for nonobese MASLD using eight routine variables, with areas under the curve reaching 0.779 and 0.771 respectively.
• Fasting insulin and triglycerides emerged as the strongest drivers of model performance, pointing to underlying insulin resistance and dyslipidemia in lean populations.
• At default thresholds, the top-performing models correctly identified only 33% to 40% of confirmed cases, demonstrating that the tools currently supplement rather than replace clinical judgment.

Predictive Models Confront Lean Metabolic Disease

Researchers recruited 2223 adults through community health screenings at a hospital in Qingdao, China, excluding individuals with a body mass index of 25 kg/m² or higher, significant alcohol use, or other liver conditions. After filtering out incomplete data, the final analysis evaluated 768 participants with a body mass index under 25. MASLD was diagnosed using a FibroScan-controlled attenuation parameter of 248 dB/m or higher alongside at least one metabolic risk factor. The cohort averaged 50.6 years of age and was 74.5% female. Investigators split the participants into a training set of 615 and an independent testing set of 153 to evaluate eight distinct machine learning approaches.

Variables and Performance Diverge Across Algorithms

Following univariable screening and least absolute shrinkage and selection operator regression, researchers narrowed dozens of candidate features down to eight core predictors: apolipoprotein B, triglycerides, fasting insulin, fasting plasma glucose, waist circumference, age, alkaline phosphatase, and systolic blood pressure. On the independent test set, the gradient boosting machine achieved the highest discrimination with an area under the curve of 0.779, followed by the neural network at 0.777 and linear discriminant analysis at 0.771. Shapley additive explanations analysis revealed that fasting insulin drove the linear discriminant analysis model, while triglycerides dominated the gradient boosting machine.

In contrast, the random forest model yielded a perfect area under the curve of 1.000 on the training set but dropped to 0.742 on testing, illustrating the overfitting risks common to complex algorithms on modest clinical datasets. At default classification thresholds, the top models flagged just 33% to 40% of confirmed MASLD cases. The study was limited by its single-center design in China, reliance on ultrasound-based elastography instead of liver biopsy, and the inclusion of patients taking medications for diabetes or hypertension that could blunt biomarker levels.

Medium-Term Diabetes Risk in Biopsy-Confirmed Cohorts

Among seven evaluated machine learning approaches, XGBoost demonstrated the strongest performance in the external validation cohort. Its area under the receiver operating characteristic curve reached 0.99 in training and 0.89 in external validation, accompanied by an external accuracy of 84%, a sensitivity of 86%, and a specificity of 82%. Both studies show that visceral fat distribution and metabolic dysfunction drive disease progression, whether evaluated in lean populations or among patients with confirmed hepatic steatosis.

Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice. Always consult with a qualified healthcare provider regarding any medical condition, diagnosis, or treatment plan.