Machine learning derivation of two cardiac arrest subphenotypes with distinct responses to treatment.

introductionCardiac arrest (CA), characterized by its heterogeneity, poses challenges in patient management. This study aimed to identify clinical subphenotypes in CA patients to aid in patient classification, prognosis assessment, and treatment decision-making.

methodsFor this study, comprehensive data were extracted from the Medical Information Mart for Intensive Care IV (MIMIC-IV) 2.0 database. We excluded patients under 18 years old, those not initially admitted to the intensive care unit (ICU), or treated in the ICU for less than 72 h. A total of 57 clinical parameters relevant to CA patients were selected for analysis. These included demographic data, vital signs, and laboratory parameters. After an extensive literature review and expert consultations, key factors such as temperature (T), sodium (Na), creatinine (CR), glucose (GLU), heart rate (HR), PaO2/FiO2 ratio (P/F), hemoglobin (HB), mean arterial pressure (MAP), platelets (PLT), and white blood cell count (WBC) were identified as the most significant for cluster analysis. Consensus cluster analysis was utilized to examine the mean values of these routine clinical parameters within the first 24 h post-ICU admission to categorize patient classes. Furthermore, in-hospital and 28-day mortality rates of patients across different CA subphenotypes were assessed using multivariate logistic and Cox regression analysis.

resultsAfter applying exclusion criteria, 719 CA patients were included in the study, with a median age of 67.22 years (IQR: 55.50-79.34), of whom 63.28% were male. The analysis delineated two distinct subphenotypes: Subphenotype 1 (SP1) and Subphenotype 2 (SP2). Compared to SP1, patients in SP2 exhibited significantly higher levels of P/F, HB, MAP, PLT, and Na, but lower levels of T, HR, GLU, WBC, and CR. SP2 patients had a notably higher in-hospital mortality rate compared to SP1 (53.01% for SP2 vs. 39.36% for SP1, P < 0.001). 28-day mortality decreased continuously for both subphenotypes, with a more rapid decline in SP2. These differences remained significant after adjusting for potential covariates (adjusted OR = 1.82, 95% CI: 1.26-2.64, P = 0.002; HR = 1.84, 95% CI: 1.40-2.41, P < 0.001).

conclusionsThe study successfully identified two distinct clinical subphenotypes of CA by analyzing routine clinical data from the first 24 h following ICU admission. SP1 was characterized by a lower rate of in-hospital and 28-day mortality when compared to SP2. This differentiation could play a crucial role in tailoring patient care, assessing prognosis, and guiding more targeted treatment strategies for CA patients.