Development and validation of machine learning models to predict 30-day mortality in patients with cardiac arrest complicated by acute kidney injury.

BACKGROUND: Cardiac Arrest (CA) often leads to Acute Kidney Injury (AKI), presenting a major health issue. This study aims to develop and validate machine learning models predicting 30-day mortality in CA patients with AKI. METHODS: A retrospective study was conducted on 1,121 adult ICU patients diagnosed with CA and AKI using data from the MIMIC-IV database. Data from 2008‒2016 (n = 900) formed the training cohort, while 2017‒2019 data (n = 221) served as the temporal validation cohort. Feature selection utilized LASSO regression, followed by the application of six machine learning algorithms. Model performance was evaluated using ROC curves, calibration curves, and decision curve analysis, with comparisons made against the APACHE II and SOFA scores. RESULTS: Seven key features were identified, namely, APACHE II score, lactate, anion gap, RDW, oliguria, norepinephrine, and temporary pacemaker implantation. The LASSO-Logistic Regression (LASSO-LR) model demonstrated the most stable performance in validation (AUROC = 0.751), significantly outperforming APACHE II (0.583, p < 0.001) and SOFA (0.660, p = 0.009), with a Brier score of 0.209. SHAP analysis revealed oliguria and RDW as the most significant predictors. CONCLUSIONS: The LASSO-LR model offers a modest but statistically significant improvement over conventional scoring systems for predicting 30-day mortality in patients with CA and AKI, with acceptable calibration and interpretability. However, the clinical meaningfulness of this incremental benefit requires prospective validation before clinical implementation.