Hepatic Steatosis Severity Prediction in Nonobese Individuals: Machine Learning Model Development and Validation

Steatotic liver disease affects 40% of nonobese individuals, but existing screening tools inadequately detect and stage disease severity in this population because of the limited sensitivity of conventional ultrasound and the lack of dedicated prediction models.

This study aimed to develop and validate an interpretable machine learning model specifically for multiclass hepatic steatosis severity prediction in nonobese individuals to support early risk stratification in this underrecognized group.

Health examination data from 215,145 nonobese participants (BMI <28 kg/m²) were randomly divided into training (n=150,601, 70%) and test (n=64,544, 30%) sets. Hepatic steatosis was diagnosed and graded using the controlled attenuation parameter with established thresholds (none: <248 dB/m; mild: 248-268 dB/m; and moderate to severe: >268 dB/m). From 42 candidate variables, 14 predictors were selected using Least Absolute Shrinkage and Selection Operator regression and Recursive Feature Elimination based on Random Forest importance. Six machine learning algorithms-k-nearest neighbors, naive Bayes, multilayer perceptron, random forest, support vector machine, and Extreme Gradient Boosting (XGBoost)-were developed using 10-fold cross-validation, with hyperparameters optimized for maximal area under the receiver operating characteristic curve (ROC-AUC). Model interpretability was assessed using Shapley Additive Explanations analysis. External validation was conducted in non-Hispanic Asian participants from the National Health and Nutrition Examination Survey (n=726). Model performance was evaluated using accuracy, Cohen κ, ROC-AUC, area under the precision-recall curve, F1-score, precision, sensitivity, and specificity.

The final cohort included 215,145 participants, with steatosis severity classified as none (n=92,944, 43.2%), mild (n=54,121, 25.2%), and moderate to severe (n=68,080, 31.6%). Among the 6 machine learning models, XGBoost achieved the best discrimination on the test set, with an accuracy of 0.824 and a macro-average ROC-AUC of 0.941. In external validation, the model maintained strong performance (macro-average ROC-AUC=0.874). Shapley Additive Explanations analysis identified BMI, waist circumference, liver enzymes (alanine aminotransferase and aspartate aminotransferase), renal function indicators (uric acid and serum creatinine), and metabolic indices (triglycerides, continuous metabolic syndrome score, and triglyceride-glucose index) as key contributors to model predictions. The model has been implemented as an online prediction platform to facilitate clinical use.

This interpretable XGBoost model accurately predicts controlled attenuation parameter-defined hepatic steatosis severity in nonobese individuals and demonstrates robust performance in both internal and external validation cohorts, providing a practical tool for early risk stratification in this underrecognized population.