Critically ill COVID-19 phenotypes identified by unsupervised clustering: A multicenter retrospective ICU study

Critically ill COVID-19 patients are commonly managed as a homogeneous population, yet unsupervised clustering analyses have revealed distinct phenotypes with potentially different treatment responses.

This study aimed to identify COVID-19 clusters and investigate therapeutic responses within these clusters.

A multicenter observational retrospective study was conducted between March 2020 and December 2021. Hierarchical clustering on principal components (HCPC) was used to identify COVID-19 profiles based on standard laboratory tests. Demographic data, ICU management, mortality, and therapeutic responses were compared between clusters.

Among the 329 patients analyzed, HCPC identified two clusters (246 and 83 patients). Cluster 2 patients were older (69 [61-74] vs. 64 [55-71], p = 0.001) and had more comorbidities (3.5^2-5 vs. 2^1-4, p < 0.001). These patients experienced more severe organ failure, with higher rates of invasive mechanical ventilation (69.9%vs. 56.5%, p = 0.033) and renal replacement therapy (39.8%vs. 14.2%, p < 0.001). Correspondingly, ICU mortality was higher in Cluster 2 (44.6%vs. 27.2%, p = 0.004). In total, 218 patients (66.3%) received corticosteroids, including 170 (69.1%) in Cluster 1 and 48 (57.8%) in Cluster 2. Corticosteroid therapy was associated with improved survival in Cluster 1 (adjusted hazard ratio = 0.62, 95% CI [0.36, 1.07], p = 0.06; Log-Rank test, p = 0.035), but not in Cluster 2 (adjusted hazard ratio = 1.09, 95% CI [0.67, 1.76], p = 0.738). Additionally, in Cluster 1, corticosteroids were linked to fewer thrombotic events (14.8%vs. 20%, p = 0.012).

Based on HCPC, two distinct profiles with different severity and different outcome were identified. In patients of Cluster 1, corticosteroids were associated with improved survival and reduced thrombotic events.