Evaluation of repeated machine learning-based phenotyping in patients with cardiogenic shock

Abstract Background Cardiogenic shock (CS) is a highly mortal disease with variable clinical manifestations, severity, and in-hospital trajectories. Multiple clinical trials have failed to show an impact of therapies on outcomes, partly because they focus on etiology as a classifier, failing to capture the clinical heterogeneity. A phenotype-based approach to CS allows for a dynamic risk stratification based on disease severity. We have previously presented three CS phenotypes at baseline using machine learning-based algorithms: phenotype I (non-congested), II (cardiorenal), and III (cardiometabolic). However, little is known about the clinical course of these CS phenotypes during hospitalization and whether repeated phenotyping may provide additional insights. Purpose To assess relevance and prognostic impact of repetitive reassessment of CS phenotypes and their trajectories during the first 72h after CS diagnosis. Methods We included all-cause CS patients from the multi-center, multinational Cardiogenic Shock Working Group (CSWG) V3 and V4 registries (Yr. 2022-2023). Phenotypes were assigned by nearest centroid classification based on the centroids from the initial CSWG V1 (Yr. 2019) registry derivation cohort. Missing values required for phenotype assignment were imputed for all patients, but imputed values were only used for phenotype assignment and not reported as outcomes. Results A total of 4308 patients were included, of which 928 had CS due to myocardial infarction (MI-CS) and 2164 had CS due to heart failure (HF-CS). At time of diagnosis/ baseline, 1627 (37.8%), 1832 (42.5%), and 849 (19.7%) of patients were in phenotype I (non-congested), II (cardiorenal) and III (cardiometabolic), respectively. In MI-CS, patients were evenly distributed between the three phenotypes (35.5%, 34.4%, and 30.2%, respectively), whereas in HF-CS, non-congested and cardiorenal CS were more common than cardiometabolic shock (38.8%, 48.9%, and 12.3%, respectively). Within the first 12 hours, most patients in cardiorenal and cardiometabolic CS shifted to non-congested CS (Figure 1). In-hospital mortality in admission phenotype I, II, and III was 21.7%, 32.1%, and 50.1% (p<0.0001 for difference), respectively, in concordance with previous reports. When assessed at different time points, phenotypes II and III were associated with higher odds of mortality than phenotype I at 24h, 48h, and 72h after initial shock diagnosis (p≤0.03 for all, Figure 2). Conclusion Beyond baseline, re-assessment of machine learning-based CS phenotypes within the first 72h after CS diagnosis may provide additional prognostic insights. These data may inform future clinical trial design to tailor interventions to specific CS phenotypes.Figure 1Figure 2