Clinical subphenotypes of sepsis based on mixed continuous and categorical data and differences in treatment effects: a cluster analysis of multicenter observational studies.

backgroundSepsis is a heterogeneous syndrome; therefore, identifying subphenotypes is essential to determine the optimal treatment for each patient. Previous studies have mainly analyzed continuous variables and focused on the treatment effects in patient cohorts receiving a single intervention. In this study, we explored subphenotypes using mixed continuous and categorical data and examined multiple treatment effects for each subphenotype.

methodsWe used Fifty-two admission variables from two multicenter registries—the Tohoku Sepsis Registry and Focused Outcomes Research in Emergency Care for Acute Respiratory Distress Syndrome, Sepsis, and Trauma study—which enrolled patients aged 16 years or older with sepsis admitted to intensive care units in Japan. Cluster analysis was conducted using methods for mixed data (i.e., datasets containing both continuous and categorical variables): k-UMAP, which applies k-means clustering after dimensionality reduction via Uniform Manifold Approximation and Projection (UMAP); k-prototype; and KAMILA (KAymeans for MIxed Large data). We evaluated the stability of clustering methods and determined the optimal number of clusters using consensus clustering. Based on stable clustering results, we examined the interaction effects between treatments and subphenotypes after applying inverse probability of treatment weighting in logistic regression analysis. The outcome measure was in-hospital mortality.

resultsThe analysis included 1,756 patients. We selected the three most stable clustering results: two from k-UMAP, and one from KAMILA. Hospital mortality, patient characteristics, and treatment effectiveness varied among these subphenotypes. Significant treatment effects were observed in the k-UMAP models: the odds ratio (OR) for recombinant thrombomodulin was 0.37 (95%CI 0.13–0.80) in subphenotype 1 when classifying three subphenotypes. The OR for antithrombin III was 0.49 (95%CI 0.20–0.93) in subphenotype 2 when classifying five subphenotypes. No KAMILA-derived subphenotype showed a significant treatment effect. No adjustment for multiple comparisons was applied because these were exploratory analyses.

conclusionsUsing three clustering methods on mixed-data sepsis cohorts, we identified three to five clinically distinct subphenotypes with heterogeneous treatment responses. Our findings suggest that k-UMAP may identify novel subphenotypes that could help reveal potential treatment targets not yet proven effective.