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

Area of Science:

• Critical Care Medicine
• Data Science
• Medical Informatics

Background:

• Sepsis is a major global health threat with diverse patient characteristics impacting treatment outcomes.
• Identifying distinct sepsis patient subpopulations is crucial for developing targeted interventions.
• Previous clustering analyses on the MIMIC-III database showed promise in stratifying sepsis patients.

Purpose of the Study:

• To extend sepsis subpopulation identification to the larger, multi-hospital eICU database.
• To validate previous findings from the MIMIC-III analysis in a new dataset.
• To uncover novel sepsis subpopulations within a more heterogeneous patient cohort.

Main Methods:

• Applied dimensionality reduction techniques: t-distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and Variational Autoencoders (VAE).
• Utilized density-based clustering (DBSCAN) for subpopulation identification.
• Employed Self-Organizing Maps (SOM) for topological validation.

Main Results:

• Successfully identified recognizable sepsis subpopulations with shared clinical characteristics.
• Validated several findings from the prior MIMIC-III study.
• Discovered new subpopulations, reflecting the increased heterogeneity of the eICU database.

Conclusions:

• Advanced data analysis techniques can effectively stratify sepsis patients into distinct subpopulations.
• Findings support the potential for personalized treatment strategies in sepsis care.
• The eICU database analysis confirms and expands upon previous sepsis subtyping research.