Predictive Value of the Albumin-Bilirubin Score for All-Cause Mortality in Individuals with Stroke: Machine Learning Prediction Models with SHAP-Based Interpretation.

Background

The albumin-bilirubin (ALBI) score, an objective liver function and prognostic indicator, has shown predictive value in various diseases, but its long-term prognostic significance in stroke survivors remains unclear.

Objective

To evaluate the ALBI score's predictive value for all-cause mortality (ACM) in individuals with stroke.

Methods

Data from the National Health and Nutrition Examination Survey (1999-2018) were analyzed. The Boruta algorithm selected key variables associated with ACM. Kaplan-Meier curves, Cox regression, and restricted cubic splines examined associations and nonlinearity between ALBI and ACM risk. Five machine learning (ML) models were built with 10-fold cross-validation; performance was assessed via C-index, integrated Brier score, time-dependent AUC, and calibration curves. SHapley Additive exPlanations (SHAP) interpreted the best model.

Results

Among 1,838 stroke survivors (median follow-up: 78 months), Boruta-selected variables included age, sex, race, marital status, body mass index, cardiovascular disease, lymphocyte count, aspartate/alanine aminotransferase, uric acid (UA), polyunsaturated fatty acids, basophil percentage, and creatinine. Kaplan-Meier curves showed significant survival differences across ALBI quartiles (log-rank P<0.001), with higher ALBI linked to lower survival. Cox regression confirmed ALBI as an independent risk factor (Model 3: HR=2.297, 95% CI: 1.662-3.174, P<0.001). Among ML models, gradient boosting machine performed best (C-index=0.749, 95% CI: 0.733-0.764), with 1-, 3-, and 5-year td-AUCs of 0.820, 0.798, and 0.771, respectively. SHAP identified age as the top predictor, followed by UA, creatinine, ALBI, marital status, and others.

Conclusion

In stroke survivors, the ALBI score independently predicts ACM and may aid risk stratification and individualized management.