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

Background: Sepsis therapy is still limited to the treatment of the underlying infection and supportive measures. Therapeutic options that address the molecular changes of sepsis have not yet been identified. With the aim of a future individualized therapy, we used unsupervised machine learning (ML) to identify clinical phenotypes in a prospective multicenter cohort of patients with sepsis and characterized them using plasma proteomics. Methods: Routine clinical data and blood samples were collected from 384 sepsis patients. Clinical phenotypes were identified based on clinical routine measurements using the k-means algorithm. In addition, plasma samples from 276 patients were analyzed using liquid chromatography-coupled tandem mass spectrometry (LC-MS/MS). The obtained data were analyzed and interpreted in the context of the phenotypes and supervised ML classifiers were developed to prospectively allocate patients to the clusters and to identify the most important features for discrimination of the phenotypes. Results: Three clinical phenotypes were identified. Cluster C was characterized by the highest disease severity and multi-organ failure with leading liver failure and a mortality rate of 92%. Cluster B, with a mortality rate of 45 %, also showed relevant organ failure, with renal damage being particularly prominent in comparison to cluster A. The plasma proteome reflected the clinical features of the phenotypes and revealed the excessive consumption of complement and coagulation factors in severe sepsis. Supervised ML and feature importance analysis underlined these findings and highlighted specific clinical measures and proteins. Conclusions: ML identified clinical phenotypes showed different degrees of sepsis severity and could be translated to the plasma proteome. Plasma proteomics provided novel insights into the molecular processes of sepsis and allowed the characterization of the phenotypes. This approach represents a blueprint to identify molecular features of sepsis subgroups and may pave the way for future targeted sepsis therapy.