Unsupervised identification of sepsis subpopulations in the eICU database: A multi-method clustering approach with validation

Sepsis remains one of the leading causes of death worldwide, and despite extensive research, uncertainties persist regarding its treatment outcomes due to the diversity of the condition and characteristics across patients. Identifying subpopulations of sepsis patients with distinct clinical behaviors can be instrumental in developing more targeted and effective interventions. In this study, we build on previous work that applied clustering techniques to the large cohort single-hospital MIMIC-III intensive care unit (ICU) database by extending the analysis to the larger cohort multi-hospital eICU database. We employ multiple-dimensional reduction methods such as t-distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and Variational Autoencoders (VAE) in combination with density-based clustering (DBSCAN) and use Self-Organizing Maps (SOM) as an extra topological validation. Our approach was able to uncover recognizable subpopulations of sepsis with some shared characteristics, both validating many results from the previous MIMIC-III analysis and identifying new results that appear indicative of the more heterogeneous eICU database.