CKM stage redefines long-term risk stratification after cardiac arrest: a multicenter machine learning analysis

Background Cardiac arrest (CA) is one of the leading causes of death among cardiovascular emergencies, with poor long-term prognosis. Currently, there is a lack of research accurately predicting the prognosis of CA patients, and existing studies mostly focus on acute factors, underestimating the impact of chronic factors, especially cardiovascular-renal-metabolic (CKM) burden, on long-term outcomes. Purpose This study aims to establish an interpretable machine learning framework for accurately predicting the risk of death in CA patients at different time points based on the same indicators, thereby enabling the identification of high-risk patient clinical phenotypes and early intervention. Methods We conducted a retrospective cohort study, including patients diagnosed with CA from the MIMIC-IV and eICU databases. We identified high-risk factors for death in CA patients through survival analysis. Subsequently, we used three methods (LASSO, Boruta algorithm and RFE) to screen for five core features: CKM stage, 24-hour urine output, GCS score, SOFA score, and mean systolic blood pressure. Based on the above variables, we constructed models using five machine learning methods and plotted ROCs to compare model predictive power and stability. We achieved model interpretability by quantifying feature importance using Shapley Additive Explanation (SHAP). K-means clustering helped us achieve high-risk phenotype identification and risk stratification. Results Our study ultimately included 4,679 patients admitted to the ICU due to CA. 1,901 patients were included in the MIMIC-IV extrapolated cohort and randomly assigned in a 7:3 ratio to the training set (n = 1332) and the internal validation set (n = 569). A cohort of 2778 independent patients from the eICU database served as the external validation set. Survival analysis showed a significant difference in 365-day survival between CKM groups (χ²=93.36, P<0.0001). Multivariate Cox regression analysis confirmed that CKM stage was a significant independent risk factor for mortality (HR = 1.622, 95% CI: 1.414–1.860). Among the five machine learning methods, the logistic regression model achieved robust and accurate predictions. The highest AUC was 0.876 for 30-day mortality (95% CI 0.862–0.890). SHAP analysis revealed a key temporal shift in prognostic factors: short-term mortality was primarily influenced by acute indicators (urine output), while long-term survival was primarily influenced by baseline CKM stage. Cluster analysis identified a high-risk phenotype (progressive CKM, severe organ failure) with a 1-year mortality rate of 87.6% (compared to 50.8% in the low-risk group, P < 0.001). Conclusion Prognostic factors for CA survivors shift from acute injury to chronic burden over time. Integrating CKM staging significantly improves long-term risk stratification, suggesting that future post-resuscitation treatment must incorporate chronic metabolic management strategies.