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

Le résumé fourni par la source

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.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.