Predicting Metabolic Dysfunction-Associated Steatotic Liver Disease using Machine Learning Methods | JoVE Visualize

Area of Science:

Medical Informatics
Public Health
Machine Learning in Healthcare

Background:

Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD) affects a significant portion of U.S. adults, representing the most prevalent chronic liver disease.
Early detection of MASLD is crucial for intervention and preventing progression to cirrhosis, highlighting the need for effective prediction models.
Existing prediction methods require evaluation for fairness and reproducibility in large-scale clinical data.

Purpose of the Study:

To develop and validate a fair, rigorous, and reproducible prediction model for MASLD using electronic health records.
To compare the performance of various machine learning models, including LASSO logistic regression, random forest, XGBoost, and neural networks.
To assess and improve model fairness across different racial and ethnic subgroups using postprocessing techniques.

Main Methods:

Evaluated multiple machine learning models (LASSO, random forest, XGBoost, neural network) using clinical features, focusing on top SHAP-ranked features.
Employed an equal opportunity postprocessing method to mitigate disparities in true positive rates among racial and ethnic groups.
Utilized a large electronic health record database comprising training, validation, and testing datasets for model development and evaluation.

Main Results:

A LASSO logistic regression model (MASER) with the top 10 features demonstrated competitive performance (AUROC 0.836, accuracy 77.6%).
Before fairness adjustments, the model achieved 78% accuracy and 72% sensitivity.
After applying equal opportunity postprocessing, accuracy increased to 81% and specificity to 94%, with a noted trade-off in sensitivity (41%) and F1-score (0.515).

Conclusions:

The MASER model, a LASSO logistic regression approach, provides a balance of predictive performance and fairness for MASLD prediction.
Interpretable models can achieve comparable performance to complex ensemble or tree-based models while addressing fairness concerns.
This study demonstrates the feasibility of developing fair and accurate MASLD prediction tools for diverse patient populations using EHR data.