Chapter 7 212 participants with and without PASC at 12 and 24 weeks after illness onset. Median (IQR) log-concentrations were compared using the Mann-Whitney U-test with correction for multiple testing using the Bonferroni method. Comparisons were first performed using data from the whole cohort (mild to critical COVID-19) and subsequently restricted to participants with initially mild or moderate COVID-19. In order to additionally assess differences in cytokine levels using an objective marker of PASC, we compared median (IQR) log-concentrations of cytokines measured at 21-24 weeks between participants with and without an impaired diffusion capacity (DLCO) at 6 months after illness onset. To assess factors cross-sectionally associated with cytokine levels after illness onset, we applied two linear mixed-effects models: the first at 3 months (serum collected at 9-12 weeks) and the second at 6 months (serum collected at 21-24 weeks) after illness onset. We modelled log concentrations of cytokines with time since illness onset as a random effect and as fixed effects: sex, age and clinical characteristics (i.e. body mass index [BMI], comorbidities – defined at illness onset), PASC status (at 12 and 24 weeks after illness onset in each model, respectively), and recent [<4 weeks] vaccination). In order to incorporate an objective measure of abnormal pathology following COVID-19, we substituted PASC status (based on self-reported symptoms) for impaired diffusion capacity (DLCO) in an additional model at 6 months after illness onset. Condition indices were computed to ensure that there was no collinearity among the variables (i.e., condition index<10). A correlation matrix of cytokines at 0-4, 9-12 and 21-24 weeks were used to help interpret the complexity of inter-marker associations. We then performed random forest regression, a model-free machine-learning approach, to identify the early predictors of: (1) PASC at 24 weeks and (2) higher levels, at 21-24 weeks after illness onset, of CRP and IL-6. We chose to investigate predictors of raised pro-inflammatory cytokines at 6 months in order to include more objective outcome measures than the current PASC definition. We performed k-fold cross validation (k=5) to tune the hyperparameters of each random forest regressor. We used F1 scores and mean squared error as scoring functions of the random forest regressors used in (1) and (2)/(3) respectively. We then computed Shapley additive explanation values as measures of importance for the different predictors[19]. CRP and IL6 at 0-4 weeks were not individually included as predictors of their measurements at 21-24 weeks. All statistical analyses were performed in Python using the statsmodels package (v. 0.13.2) [20], whilst the random forest regression analyses were performed in Python using the scikit-learn package (v. 1.1.3).
RkJQdWJsaXNoZXIy MTk4NDMw