Development and validation of a machine learning model integrating BUN/Cr ratio for mortality prediction in critically ill atrial fibrillation patients.

Atrial fibrillation (AF), the most prevalent critical care arrhythmia, demonstrates substantial mortality associations where renal dysfunction management plays a pivotal therapeutic role. We examined the prognostic capacity of admission blood urea nitrogen-to-creatinine ratio (BUN/Cr) - a low-cost renal biomarker - for 28-/365-day mortality prediction in AF through multidimensional survival analyses leveraging the MIMIC-IV 3.1 database. Data relevant to AF patients were extracted from the publicly available MIMIC-IV 3.1 database based on predefined inclusion and exclusion criteria. Cox proportional hazards regression, Kaplan-Meier survival analysis, and Restricted Cubic Spline (RCS) models were used to assess the association between the BUN/Cr and the risk of 28-day and 365-day mortality. Subsequently, a short-term and long-term mortality risk prediction model for AF patients was developed using interpretable machine learning algorithms, incorporating the BUN/Cr and other clinical features. The MIMIC-IV analysis included 14,725 AF patients (72.9 ± 11.7 years, 60.3% male). Cox regression identified BUN/Cr as an independent predictor of 28-day and 365-day mortality, with risk quintiles showing a non-linear pattern: Q5 (> 27.8), Q4 (22.0-27.8), Q1 (≤ 15.0), Q3 (18.5-22.0), and Q2 (15.0-18.5). Kaplan-Meier curves confirmed decreasing survival with elevated BUN/Cr. Restricted cubic splines revealed U-shaped mortality relationships (P < 0.001), with inflection points at BUN/Cr = 16.49 (28-day) and 16.67 (365-day). Among machine learning models, XGBoost outperformed others in predicting mortality (28-day: AUC = 0.793 [0.776-0.810], Accuracy = 73.1%; 365-day: AUC = 0.778 [0.764-0.793], Accuracy = 69.8%). SHAP analysis ranked BUN/Cr fourth among predictors for both endpoints. The BUN/Cr emerged as a robust independent predictor of short- and long-term mortality in AF. The interpretable XGBoost model, integrating BUN/Cr with clinical variables, achieved superior predictive accuracy for 28-/365-day outcomes while maintaining generalizability. BUN/Cr constituted a fourth-ranked feature across mortality timelines. These findings underscore its clinical utility for AF risk stratification and treatment optimization, supporting biomarker-guided therapeutic interventions.