Explainable machine learning models predict liver fibrosis risk and outcome in the general population: Development...

Area of Science:

• Computational Medicine
• Data Science in Healthcare
• Liver Disease Diagnostics

Background:

• Significant liver fibrosis often presents asymptomatically but is a predictor of adverse health outcomes.
• Current screening methods may not be universally accessible or efficient for population-level pre-screening.
• Development of interpretable machine learning (ML) frameworks using accessible data is crucial for early fibrosis detection.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) framework for population-level pre-screening of significant liver fibrosis.
• To utilize readily obtainable demographic, anthropometric, and clinical variables for the ML model.
• To assess the model's accuracy, interpretability, and prognostic relevance.

Main Methods:

• Trained and evaluated ten ML algorithms on 9424 European participants from the UK Biobank.
• Utilized XGBoost, validated in the NHANES (n=15,270) and an independent real-world cohort (n=694).
• Assessed discrimination via AUC, interpretability using SHAP and LIME, and prognostic relevance via all-cause mortality.

Main Results:

• XGBoost achieved high AUCs in internal validation (0.818) and test cohorts (0.816).
• External validation showed AUCs of 0.746 (NHANES) and 0.793 (real-world cohort).
• Key predictors included weight, age, height, hypertension, and waist circumference; high-risk status correlated with increased mortality.

Conclusions:

• The XGBoost model, using accessible data, shows significant potential for identifying high-risk individuals for significant liver fibrosis in primary care.
• This interpretable ML approach can enhance early detection strategies for liver fibrosis.
• The model facilitates timely interventions, improving patient outcomes.