Non-Obese Hepatic Steatosis Severity Prediction: Machine Learning Model Development and Validation (Preprint)

<sec> <title>BACKGROUND</title> Non-obese individuals account for 40% of global steatotic liver disease (SLD) cases, yet lack dedicated targeted screening tools. Current ultrasound-based methods exhibit low detection rates for mild steatosis, delaying intervention. </sec> <sec> <title>OBJECTIVE</title> This study aimed to develop and validate a non-invasive, multi-class machine learning model using the Ultrasound Attenuation Parameter to predict hepatic steatosis severity grades (none/mild/moderate-severe) in non-obese populations. It sought to address the critical diagnostic gap in early detection and risk stratification for this under-recognized group, thereby enabling timely intervention. </sec> <sec> <title>METHODS</title> A cohort of 215,145 participants enrolled from 2018 to 2024 was analyzed. UAP thresholds defined steatosis severity: &lt;244 dB/m (none), 244–269 (mild), &gt;269 (moderate to severe). Least Absolute Shrinkage and Selection Operator (LASSO) regression identified 14 predictors. Six ML models were trained (70% of the dataset) and validated (30%) using 10-fold cross-validation. Performance metrics included accuracy, Cohen’s kappa, area under the receiver operating characteristic curve (AUROC), F1-score, and SHapley Additive exPlanations (SHAP) analysis for interpretability. </sec> <sec> <title>RESULTS</title> The ML models were developed and validated using 150,602 participants in the training set and 64,543 in the test set, comprising non-SLD (n=92,944), mild SLD (n=54,121), and moderate-to-severe SLD (n=68,080). The Extreme Gradient Boosting (XGBoost) model demonstrated superior performance compared to other models. On the training set, it achieved a macro-average AUROC of 0.929, a macro-average precision-recall (PR) AUC of 0.878, and an accuracy of 0.788. On the test set, performance remained strong, with a macro-average AUROC of 0.908, a macro-average PR AUC of 0.842, and an accuracy of 0.759, surpassing other models. </sec> <sec> <title>CONCLUSIONS</title> The XGBoost model enables timely severity assessment, reduces risks of delayed diagnoses, and supports data-driven individualized interventions, demonstrating significant translational potential of this AI-driven approach for SLD management. </sec>