Population scale proteomics enables adaptive digital twin modelling in sepsis

Sepsis is one of the leading causes of mortality in the world. Currently, the heterogeneity of sepsis makes it challenging to determine the molecular mechanisms that define the syndrome. Here, we leverage population scale proteomics to analyze a well-defined cohort of 1364 blood samples taken at time-of-admission to the emergency department from patients suspected of sepsis. We identified panels of proteins using explainable artificial intelligence that predict clinical outcomes and applied these panels to reduce high-dimensional proteomics data to a low-dimensional interpretable latent space (ILS). Using the ILS, we constructed an adaptive digital twin model that accurately predicted organ dysfunction, mortality, and early-mortality-risk patients using only data available at time-of-admission. In addition to being highly effective for investigating sepsis, this approach supports the flexible incorporation of new data and can generalize to other diseases to aid in translational research and the development of precision medicine.