Machine Learning-Based Prediction of Ultrasound-Detected Hepatic Steatosis Within the Metabolic...

Area of Science:

• Hepatology
• Machine Learning
• Metabolic Disease

Background:

• Metabolic dysfunction-associated steatotic liver disease (MASLD) is the leading cause of chronic liver disease globally.
• Early detection of hepatic steatosis is crucial for cardiometabolic risk assessment but conventional imaging is impractical for screening.

Purpose of the Study:

• Develop interpretable machine learning (ML) models to predict ultrasound-detected hepatic steatosis.
• Utilize routinely available clinical and biochemical data for MASLD screening.

Main Methods:

• Analyzed data from 644 adults, with 50% having ultrasound-detected hepatic steatosis.
• Employed a scikit-learn pipeline for data preprocessing, imputation, and feature selection.
• Evaluated nine supervised ML classifiers using cross-validation and SHAP for interpretability.

Main Results:

• Logistic Regression and Gradient Boosting models achieved the best performance (accuracy=0.65, AUROC=0.71).
• Key predictors included weight, Ponderal Index, FIB-4, BUN/Creatinine ratio, APRI, and Visceral Adiposity Index.
• ML models outperformed traditional indices like FLI and HSI.

Conclusions:

• Routinely available clinical and biochemical data can predict hepatic steatosis with moderate accuracy.
• Interpretable ML models, particularly Logistic Regression and Gradient Boosting, offer a pragmatic, low-cost approach for early MASLD identification.
• These models can be valuable tools in primary and metabolic care settings.