Comparison of Machine Learning Models and the Fatty Liver Index in Predicting Lean Fatty Liver

Area of Science:

• Hepatology
• Machine Learning
• Predictive Analytics

Background:

• Non-alcoholic fatty liver disease (NAFLD) affects lean populations, with prevalence ranging from 7.6% to 19.3% in reported studies.
• Accurate prediction of NAFLD in lean individuals is crucial for early intervention and management.
• Existing predictive tools may have limitations in this specific demographic.

Purpose of the Study:

• To develop and evaluate machine-learning models for predicting fatty liver disease in lean individuals.
• To compare the predictive performance of a novel neural network model against the established Fatty Liver Index (FLI).

Main Methods:

• A retrospective study analyzed 12,191 lean subjects (BMI < 23 kg/m²).
• Data from 27 clinical features were used to train and test models, excluding medical and substance use history.
• A two-class neural network was developed and compared with the FLI using the area under the receiver operating characteristic curve (AUROC).

Main Results:

• The study identified 741 (6.1%) lean individuals with fatty liver.
• The two-class neural network model, utilizing 10 features, achieved the highest AUROC (0.885) during development.
• In the testing group, the neural network (AUROC: 0.868) outperformed the FLI (AUROC: 0.852) in predicting fatty liver.

Conclusions:

• A two-class neural network demonstrates superior predictive value for fatty liver disease in lean individuals compared to the FLI.
• Machine learning offers a promising approach for non-invasive NAFLD prediction in specific populations.
• Further validation of this model in diverse lean cohorts is warranted.