METABOLOMIC PROFILING AND PROGNOSTICATION IN COVID-19-RELATED ARDS

SESSION TITLE: Late Breaking Posters in Critical Care SESSION TYPE: Original Investigation Posters PRESENTED ON: 10/18/2022 01:30 pm - 02:30 pm PURPOSE: The majority of deaths in COVID-19 are due to acute respiratory distress syndrome (ARDS). We recently identified two subphenotypes among patients with COVID-19 related ARDS (C-ARDS) with divergent outcomes and responses to therapies. However, the precise biological processes that distinguish the subphenotypes, remain to be fully elucidated. High-resolution profiling of the metabolome can be used to gain precise insights into disease pathogenesis. The purpose of this study was to use precise, metabolomic profiling at the onset of C-ARDS to identify metabolic alterations and predict hospital mortality. METHODS: This was a retrospective, matched cohort study. Participants were adults with COVID-19 who met Berlin criteria for ARDS on the initial day of mechanical ventilation. All participants had prospectively banked plasma samples collected within one week of intubation. Twenty-five survivors to 90-days were matched on age, sex, and ethnicity to 25 patients who died within 28 days of intubation. Untargeted and targeted metabolomic analysis was performed using mass spectrometry and compared between survivors and non-survivors. Statistical analyses were performed with conditional logistic regression modeling with Bayesian inference. Compounds associated with mortality were identified using a cut-off of Bayes Factor (BF) > 3. Biological clustering analysis was performed using ChemRICH. Competitive modeling by four machine learning models—LASSO, adaptive LASSO, Random Forest, and XGBoost—was used to predict mortality. Three sets of predictors were explored: all metabolites, metabolites with BF > 1, and metabolites with BF > 3. RESULTS: Targeted and untargeted metabolomics of metabolic analytes yielded data for 30 bile acids, 340 biogenic amines, 522 complex lipids, 83 oxylipins, and 133 primary metabolites. Twenty-five compounds were identified with significant differences between survivors and non-survivors. Five compounds had increased levels associated with mortality, and 20 had decreased levels associated with mortality. Biological clustering analysis on these compounds identified four key clusters of compounds—unsaturated and saturated lysophosphatidylcholines, plasmalogens, and saturated ceramides—that were decreased amongst non-survivors. A machine learning-derived signature reflecting these metabolites showed excellent discrimination in predicting mortality, with the best model demonstrating area-under-the-receiver-operating-characteristic curve of 0.91. CONCLUSIONS: Metabolomic analysis identified differential enrichment of lipid metabolites in C-ARDS survivors compared to non-survivors. A machine learning model was able to accurately predict mortality from C-ARDS based on metabolomic profiles. CLINICAL IMPLICATIONS: Improved characterization of the metabolomic derangements in COVID-19 ARDS may lead to an enhanced understanding of drivers of mortality and improve prognostication and precision therapy. DISCLOSURES: No relevant relationships by Thomas Briese No relevant relationships by Xiaoyu Che No relevant relationships by Matthew Cummings No relevant relationships by Oliver Fiehn No relevant relationships by David Furfaro No relevant relationships by Wenhao Gou no disclosure on file for Walter Lipkin;no disclosure on file for Nischay Mishra;No relevant relationships by Max O'Donnell

Detection of Novel Coronavirus Pneumonia Based on CT Image with Convolutional Neural Network Processing

To improve the ability to distinguish novel coronavirus pneumonia from common pneumonia and assist medical staff in chest CT examination of pneumonia patients, a detection method using convolution neural network and CT image based on artificial intelligence image analysis was proposed. First, a convolution neural network model was built, and the influence of model depth on detection results was evaluated to select the best network structure. Second, a tabu genetic algorithm was proposed to obtain the optimal hyper-parameter combination of the network model and to enhance the performance of the model. Finally, the best network model was employed to distinguish novel coronavirus pneumonia from common pneumonia. Experimental results show that the accuracy, MCC, and F1Score of the proposed detection algorithm are 93.89%, 93.32% and 91.40%, respectively, which has higher detection accuracy than other algorithms. © 2021, Acta Metrologica Sinica Press. All right reserved.