Objectives The use of machine learning to identify patient ‘clusters’ using post-return of spontaneous circulation (ROSC) vital signs may facilitate the identification of patient subgroups at high risk of rearrest and mortality. Our objective was to use k-means clustering to identify post-ROSC vital sign clusters and determine whether these clusters were associated with rearrest and mortality. Methods The ESO Data Collaborative 2018–2022 datasets were used for this study. We included adult, non-traumatic OHCA patients with >2 post-ROSC vital sign sets. Patients were excluded if they had an EMS-witnessed OHCA or were encountered during an interfacility transfer. Unsupervised (k-means) clustering was performed using minimum, maximum, and delta (last minus first) systolic blood pressure (BP), heart rate, SpO2, shock index, and pulse pressure. The assessed outcomes were mortality and rearrest. To explore the association between rearrest, mortality, and cluster, multivariable logistic regression modeling was used. Results Within our cohort of 12,320 patients, five clusters were identified. Patients in cluster 1 were hypertensive, patients in cluster 2 were normotensive, patients in cluster 3 were hypotensive and tachycardic (n = 2164; 17.6%), patients in cluster 4 were hypoxemic and exhibited increasing systolic BP, and patients in cluster 5 were severely hypoxemic and exhibited a declining systolic BP. The overall proportion of patients who experienced mortality stratified by cluster was 63.4% (c1), 68.1% (c2), 78.8% (c3), 84.8% (c4), and 86.6% (c5). In comparison to the cluster with the lowest mortality (c1), each other cluster was associated with greater odds of mortality and rearrest. Conclusions Unsupervised k-means clustering yielded 5 post-ROSC vital sign clusters that were associated with rearrest and mortality.
Read full abstract