Development and evaluation of explainable machine learning models for predicting prognosis in patients with primary...

Area of Science:

• Hepatology
• Medical Informatics
• Machine Learning

Background:

• Primary biliary cholangitis (PBC) is a chronic liver disease requiring accurate prognostic prediction.
• Current prognostic models may benefit from advanced machine learning approaches.
• Interpretable models are crucial for clinical trust and application in PBC management.

Purpose of the Study:

• To develop and validate interpretable machine learning models for predicting prognosis in primary biliary cholangitis (PBC) patients.
• To compare the performance of various machine learning algorithms for PBC prognosis.
• To identify key clinical and laboratory features influencing PBC prognosis through model interpretation.

Main Methods:

• A large cohort of 7905 PBC patients was used, with data split into training and testing sets.
• External validation was performed on an additional 2372 patients.
• Six machine learning models (LR, RF, ET, XGBoost, LightGBM, MLP) were evaluated using SHapley Additive exPlanations (SHAP) for interpretation.

Main Results:

• Ensemble tree models, particularly Extreme Random Trees (ET), outperformed linear models and shallow neural networks.
• The ET model achieved high predictive efficacy (AUC = 0.9898 training, 0.9681 external validation).
• The bilirubin-albumin ratio was identified as the core prognostic feature, supported by bilirubin, days, and prothrombin, aligning with clinical understanding.

Conclusions:

• The developed ET model offers precise and interpretable prognosis prediction for PBC patients.
• SHAP analysis confirmed model interpretability and identified key prognostic indicators.
• Findings provide quantitative evidence for clinical assessment and support individualized treatment strategies in PBC.