Clinical subphenotypes of sepsis based on mixed data and differences in treatment effects: a cluster analysis of multicentre observational studies

Background Sepsis is a heterogeneous syndrome, and the treatment effects may vary depending on its subphenotype. Previous studies have not fully used mixed clinical data, nor have investigated the effects of these multiple treatments across subphenotypes. In this study, we aimed to classify patients with sepsis into subphenotypes based on mixed clinical data and examined the differences in treatment effectiveness by subphenotype. Methods This study was a secondary analysis of multicenter registries that enrolled patients with sepsis admitted to intensive care units in Japan. The patients aged 16 years or older admitted to the ICU due to a diagnosis of sepsis were included and fifty-two variables at admission were used in the cluster analysis. We applied the state-of-the-art clustering method named k-UMAP, which uses uniform manifold approximation and projection for dimensionality reduction, followed by clustering using k-means and the previous clustering methods k-prototype and KAMILA. To examine differences in the effectiveness of the six treatments by subphenotype, a logistic regression model was used for propensity score-based weighted data to calculate the odds ratios for the interaction of the treatments and subphenotype in each clustering method. The primary outcome was the in-hospital mortality rate. Results The analysis included 1,756 patients. The results of three clustering methods using mixed clinical data led to the three conditions with high robustness being selected: k-UMAP [number of clusters (k) = 3], k-UMAP [k = 5], and KAMILA [k = 3]. Hospital mortality, patient characteristics, and treatment effectiveness varied by subphenotype. Recombinant thrombomodulin was significantly effective for subphenotype 1 in the k-UMAP [k = 3] and in the k-UMAP [k = 5]. Antithrombin III was significantly effective for subphenotype 2 in the k-UMAP [k = 5]. On the other hand, in KAMILA [k = 3], no significant treatment differences were observed between subphenotypes. Conclusions Using state-of-the-art clustering methods for mixed data were identified three to five subphenotypes which were associated with clinical information and outcomes. The effects of treatments differed for each subphenotype, and k-UMAP may reveal appropriate treatment targets that have not been proven effective.