Interpretable machine learning to predict NOAF in ICU patients with CKD: validation in US and Chinese cohorts

Le résumé fourni par la source

Objective Critically ill patients with chronic kidney disease (CKD) are at high risk for New-Onset Atrial Fibrillation (NOAF), which significantly increases their risk of adverse events. Early detection of NOAF is crucial for prompt intervention and better outcomes. This study aims to develop and validate predictive models for the early identification and stratification of NOAF risk in this vulnerable population. Methods We developed and validated a predictive model using a cohort of 6,476 critically ill patients with CKD from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database. After performing feature selection via least absolute shrinkage and selection operator (Lasso) logistic regression, we trained six machine learning (ML) models. These algorithms included: Random Forest, Gradient Boosting, eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), Multi-layer Perceptron (MLP), and Light Gradient Boosting Machine (LightGBM). The best-performing model was interpreted using SHAP to provide clinical insights. Its robustness and generalizability were confirmed through rigorous external validation on two distinct international cohorts: the US-based eICU-CRD (eICU Collaborative Research Database) ( n = 12,509) and a Chinese ICU database from Weifang People’s Hospital ( n = 880). Results Ultimately, 12 predictive features were ultimately selected: age, SOFA score, minimum heart rate, congestive heart failure, average heart rate, minimum systolic blood pressure (SBP), mechanical ventilation use, minimum oxygen saturation (SpO 2 ), average respiratory rate, minimum magnesium, SAPS II score, and maximum white blood cell (WBC) count. The Random Forest model demonstrated the best overall performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.855 in internal validation. The model’s robustness was confirmed through external validation on two independent cohorts, yielding an AUC of 0.760 on the eICU-CRD and 0.740 on the Weifang People’s Hospital database. According to the SHAP analysis, age, average heart rate, minimum heart rate, SOFA score, and SAPS II were the top five most influential predictors for NOAF. Conclusion We developed an interpretable machine learning model to predict NOAF in critically ill CKD patients, demonstrating strong generalizability through external validation on both a large US cohort (eICU-CRD) and a single-center Chinese cohort (Weifang People’s Hospital). SHAP analysis enhanced model interpretability.

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