(154) Machine Learning-Guided Subphenotyping of Critically Ill Patients After Cardiac Surgery

Introduction: Post cardiac surgery patients are complex and require high acuity of care. Multiple subphenotypes have been identified in patients with sepsis and ARDS, however, it is unclear if such heterogeneity exists in patients early after cardiac surgery. Identification of such subphenotypes early on would allow for optimization of allocation of resources. In this study we aimed to identify subphenotypes among critically ill patients who have undergone cardiac surgery using routinely available data in electronic health records.

Methods: We identified cardiac surgery patients from the MIMIC-IV database who went coronary artery bypass graft or valve repair/replacement surgeries. We derived features from available demographics, comorbidities, vital signs, labs, and vasopressor doses within the first 24 hours of ICU stay after cardiac surgery. Principal component analysis was used for dimensionality reduction followed by agglomerative hierarchical clustering for identification of subphenotypes. Chi-squared and Kruskal-Wallis tests were used for inter-subphenotype comparisons.

Results: There were 6624 post cardiac surgery patients. Among them we identified three subphenotypes characterized by increasing levels of complexity. Subphenotype 1 (n=2315) included patients with median age of 64y, Elixhauser comorbidity score of 3, median 1.29 day ICU stay, median ventilator time of 5.4 hours, and 0.13% in-hospital mortality. Subphenotype 2 (n=3165) included patients with median age 71y, higher comorbidity burden (Elixhauser score of 4), slightly longer ICU stays (median 1.4 days), longer ventilator time (7.2 hours), and higher mortality rate (0.32%). Subphenotype 3 (n=1144) included the most complex patients with highest age (median 72y), highest comorbidity burden (Elixhauser score 6), longest ICU stays (2.48 days), most ventilator hours (16 hours), and highest mortality rate (5.33%). All results were significant at p < 0.05.

Conclusions: We identified three distinct subphenotypes of critically ill patients within 24 hours after cardiac surgery. These subphenotypes had differing characteristics and outcomes. Early recognition of these subphenotypes will allow for personalized management strategies for these patients and enrichment of population for enrollment in clinical trials.