Implementation of Machine Learning Algorithms to Screen for Advanced Liver Fibrosis in Metabolic...

Area of Science:

• Hepatology
• Machine Learning in Medicine
• Diagnostic Tools

Background:

• Metabolic dysfunction-associated steatotic liver disease (MASLD) requires accurate fibrosis staging.
• Primary care settings need accessible tools for advanced fibrosis (AF) detection.
• Explainability is crucial for clinical adoption of machine learning algorithms (MLAs).

Purpose of the Study:

• To develop and validate MLAs for detecting AF in MASLD patients within a primary care context.
• To ensure the interpretability of MLA predictions for responsible clinical implementation.
• To compare MLA performance against established fibrosis scoring systems.

Main Methods:

• Trained five MLAs using readily available clinical features from 618 MASLD patients.
• Defined advanced fibrosis (AF) based on liver stiffness (≥9.3 kPa) or biopsy (≥F3).
• Validated MLAs against Fibrosis-4 index (FIB-4) and NAFLD fibrosis score (NFS) in primary care data, employing feature importance and SHAP values for explainability.

Main Results:

• Extreme gradient boosting (XGBoost) achieved an AUC of 0.91, significantly outperforming FIB-4 (0.78) and NFS (0.81).
• XGBoost demonstrated superior sensitivity (91% vs. 79%) and specificity (76% vs. 59%/48%) for AF detection.
• Key predictors included HbA1c and gamma-glutamyl transpeptidase (GGT); SHAP values aided in reclassifying false positives.

Conclusions:

• An explainable XGBoost algorithm is a superior screening tool for AF in MASLD patients compared to FIB-4 and NFS in primary care.
• The algorithm's interpretability enhances its safety and suitability for clinical decision-making.
• This approach facilitates responsible integration of AI in managing MASLD.