Towards Machine Learning-Based Detection of Endotoxemia Using Plasma Protein Denaturation Profiles and Metabolomics

Sammanfattning

Sepsis is a serious condition, being one of the most common causes of death in the intensive care unit. Millions of people worldwide are affected every year, and a critical step for reducing both morbidity and mortality for the patients is to start the right antibiotic treatment as soon as possible. However, diagnosing sepsis and characterizing the pathogens in the blood is time-consuming and requires efforts outside the emergency department. Meanwhile, machine learning has shown promise in the field of diagnostics and healthcare. This master's thesis has, therefore, investigated the possibility of using machine learning models, namely random forest, adaptive boosting, logistic regression, and support vector machine, for diagnostic purposes and for gaining more knowledge about the physiological processes behind the condition. The dataset originated from a study performed by the Emergency Medicine Research Center at Region Östergötland, Linköping, Sweden, in which healthy participants were given an intravenous dose of lipopolysaccharide (LPS) to experimentally induce endotoxemia. The same participants were also, at another occasion, given saline solution as placebo control. Plasma samples were collected at different time points, and analyzed using nano differential scanning fluorimetry (nanoDSF) and nuclear magnetic resonance (NMR), generating plasma protein denaturation profiles and metabolomic profiles, respectively. The machine learning modeling of the plasma protein denaturation profiles showed potential for detecting endotoxemia, with a maximum mean accuracy of 72.5% during the mid stages of endotoxemia. However, the models faced issues that likely originated from the dataset consisting of few individuals and the LPS group receiving an insufficient dose of LPS, which compromised the reliability of the results. The machine learning modeling of the metabolomic profiles suggested that metabolomic profiles can be used for detecting endotoxemia, although with limited clinical applications. The machine learning modeling could detect endotoxemia with a maximum mean accuracy of 85.0% for the mid stages of endotoxemia and 91.7% for the late stages of endotoxemia. The findings also confirmed the importance of lactate in sepsis diagnostics, as well as exploring the role of amino acids in inflammation contexts. While this thesis showed great potential for machine learning applications of plasma protein denaturation profiles and, especially, metabolomic profiles, further research in the field of machine learning and sepsis must be conducted before any clinical implementation and confirmation of new biomarkers for sepsis.