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

Background: In Germany, around 280,000 patients develop sepsis every year and more than 150 patients die every day. Currently, effective therapeutic measures are limited to the treatment of the underlying infection and supportive measures. Specific treatment options that directly address the molecular changes in sepsis have not yet been identified. As sepsis manifests itself in individual courses, a generalised therapeutic approach is ruled out. With the aim of a future individualised sepsis therapy, we used machine learning (ML) to identify clinical sepsis phenotypes in a multicentre, prospective cohort and characterised them using plasma proteomics.

Methods: Within the Sepsis- and CovidDataNet.NRW studies, routine clinical data from the patient data management systems was collected from critically ill sepsis patients. From this, clinical variables were selected from 384 patients on the day of sepsis diagnosis and a Principal Component Analysis (PCA) was carried out for dimension reduction. The sepsis phenotypes were identified by cluster analysis using the k-means algorithm on the PCA results. In addition, plasma samples from 276 patients were analysed by mass spectrometry-based plasma proteomics using an Ultimate3000 uHPLC coupled to an Orbitrap Exploris 240. Data was acquired in DIA mode and statistically analysed and interpreted in the context of the identified clinical phenotypes.

Results: Three clinical sepsis phenotypes were identified and termed A-C. Cluster C (n=25) was characterised by the highest disease severity, a mortality rate of 92% and multi-organ failure with leading liver failure. Cluster B (n=156), with a mortality rate of 45%, was also characterised by relevant organ failure, with kidney injury being particularly striking in comparison to cluster A (n=202). At the level of the plasma proteome, 613 proteins were quantified and massive differences between cluster C and clusters A and B were apparent. These mainly reflected the organ failure of patients in cluster C, and prominently showed the excessive consumption of acute phase proteins, coagulation factors and proteins of the complement system. The difference between clusters A and B was less pronounced overall, but reflected increased disease severity as well as kidney failure in cluster B. The consumption of complement and coagulation factor was already apparent in cluster B in comparison to the mildest cluster A.

Conclusions: The clinical phenotypes identified by an ML approach showed different severities of sepsis and could be translated to the plasma proteome level. The differential proteins allowed a deeper insight into the molecular processes of sepsis and open up the possibility of characterising the identified phenotypes in more detail on the molecular level. This approach may pave the way for future targeted therapeutic approaches to sepsis.