Identifying subpopulations of septic patients: A temporal data-driven approach

Area of Science:

• Critical Care Medicine
• Data Science in Healthcare
• Computational Biology

Background:

• Sepsis is a leading cause of mortality, with treatment outcomes often uncertain despite extensive research.
• Personalized medicine approaches are needed to address the heterogeneity of sepsis patient populations.

Purpose of the Study:

• To evaluate the utility of temporal electronic health records (EHR) for stratifying sepsis patients.
• To identify distinct patient subpopulations with similar clinical trajectories and needs using data-driven methods.

Main Methods:

• Utilized hierarchical clustering and DBSCAN on intensive care unit (ICU) data from the MIMIC III database.
• Employed t-Distributed Stochastic Neighbor Embedding (t-SNE) for patient visualization in a 2D space.
• Assessed cluster validity using silhouette index and resampling-based stability analysis.

Main Results:

• Hierarchical clustering with a Euclidean metric identified twelve clinically relevant subgroups of septic patients.
• These subgroups exhibited diverse characteristics despite sharing common sepsis-related conditions.
• Demonstrated the effectiveness of data-driven clustering in revealing patient heterogeneity.

Conclusions:

• Temporal EHR data can be effectively used to stratify sepsis patients into distinct subgroups.
• Data-driven approaches facilitate the identification of clinically relevant patient groups for tailored care.
• Findings support the development of customized care platforms for improved sepsis management.