An interpretable model combining brain imaging and clinical indicators for predicting overt hepatic encephalopathy

Area of Science:

• Medical Imaging
• Hepatology
• Machine Learning

Background:

• Distinguishing overt hepatic encephalopathy (OHE) from covert hepatic encephalopathy (CHE) is difficult due to similar symptoms and grading tool limitations.
• Quantitative susceptibility mapping (QSM) offers potential for improved diagnostic accuracy.

Purpose of the Study:

• To integrate QSM features with clinical biomarkers for an OHE predictive model.
• To develop an interpretable, web-based tool for OHE risk assessment.

Main Methods:

• Sixty-eight cirrhotic patients (31 OHE, 37 CHE) were analyzed.
• LASSO regression identified key predictors (MBP, LRN, RCA, RBC, Hgb, Fib).
• QSM-enhanced logistic, conventional logistic, and random forest models were compared using AUC, PR-AUC, and DCA.

Main Results:

• The QSM-enhanced logistic model demonstrated superior performance (AUC=0.83, PR-AUC=0.83).
• Internal validation confirmed model stability (bootstrap AUC=0.831, cross-validation AUC=0.823).
• SHAP analysis provided insights into variable importance and interactions, informing an interactive prediction tool.

Conclusions:

• Integrating QSM imaging with routine blood tests allows for accurate, explainable OHE risk prediction.
• The SHAP-based platform facilitates early detection and personalized management strategies for hepatic encephalopathy.