Machine learning identifies clinical sepsis phenotypes that translate to the plasma proteome

Area of Science:

• Sepsis research
• Machine learning in medicine
• Proteomics

Background:

• Current sepsis therapy focuses on infection and support, lacking targeted molecular treatments.
• Previous sepsis subtype classifications did not lead to identified therapeutic options.
• Individualized therapy for sepsis requires understanding molecular changes and clinical phenotypes.

Purpose of the Study:

• To utilize machine learning to identify clinical sepsis phenotypes and their temporal development.
• To characterize identified sepsis phenotypes using plasma proteomics.
• To pave the way for future individualized sepsis therapy.

Main Methods:

• Collected clinical data and blood samples from 384 sepsis patients.
• Identified sepsis phenotypes using clinical measurements and analyzed plasma proteomes of 301 patients via mass spectrometry.
• Developed supervised machine learning models for prospective phenotype classification.

Main Results:

• Identified three distinct sepsis phenotypes (Clusters A, B, C) with varying disease severity and organ failure (liver, renal).
• Cluster C associated with higher mortality; Cluster B showed potential for cluster transition.
• Plasma proteome reflected clinical phenotypes, showing complement and coagulation factor consumption with severity.
• Supervised ML models accurately classified patients using seven common clinical features.

Conclusions:

• Identified clinical sepsis phenotypes correlate with disease severity and are mirrored in plasma proteomic profiles.
• Proteomic profiling provides novel insights into sepsis molecular mechanisms and phenotype characterization.
• This approach supports the development of more precise and individualized sepsis treatments.