Predictive biomarkers of ICU and mechanical ventilation duration in critically-ill COVID19 patients

Background: Detection of early metabolic changes in critically-ill COVID-19 patients under invasive mechanical ventilation (IMV) at the intensive care unit (ICU) could predict recovery patterns and help in disease management. Methods: Targeted metabolomics of serum samples from 39 COVID-19 patients under IMV in ICU was performed within 48 hours of intubation and a week later. A generalized linear model (GLM) was used to identify, at both time points, metabolites and clinical traits that predict the length of stay (LOS) at ICU (short ≤ 14 days /long >14 days) as well as the duration under IMV. All models were initially trained on a set of randomly selected individuals and validated on the remaining individuals in the cohort. Further validation in recently published Metabolomics data of COVID-19 severity was performed. Results: A model based on hypoxanthine and betaine measured at rst time point was best at predicting whether a patient is likely to experience a short or long stay at ICU (AUC=0.92). A further model based on kynurenine, 3-methylhistidine, Ornithine, p-Cresol sulfate and C24.0 sphingomyelin, measured one-week later, accurately predicted the duration of IMV (Pearson correlation=0.94). Both predictive models outperformed APACHE II scores and differentiated COVID-19 severity in published data. Conclusion: This study has identied specic metabolites that can predict in advance LOS and IMV, which could help in the management of COVID-19 cases at ICU.


Abstract
Background: Detection of early metabolic changes in critically-ill COVID-19 patients under invasive mechanical ventilation (IMV) at the intensive care unit (ICU) could predict recovery patterns and help in disease management.
Methods: Targeted metabolomics of serum samples from 39 COVID-19 patients under IMV in ICU was performed within 48 hours of intubation and a week later. A generalized linear model (GLM) was used to identify, at both time points, metabolites and clinical traits that predict the length of stay (LOS) at ICU (short ≤14 days /long >14 days) as well as the duration under IMV. All models were initially trained on a set of randomly selected individuals and validated on the remaining individuals in the cohort. Further validation in recently published Metabolomics data of COVID-19 severity was performed.
Results: A model based on hypoxanthine and betaine measured at rst time point was best at predicting whether a patient is likely to experience a short or long stay at ICU (AUC=0.92). A further model based on kynurenine, 3-methylhistidine, Ornithine, p-Cresol sulfate and C24.0 sphingomyelin, measured one-week later, accurately predicted the duration of IMV (Pearson correlation=0.94). Both predictive models outperformed APACHE II scores and differentiated COVID-19 severity in published data.
Conclusion: This study has identi ed speci c metabolites that can predict in advance LOS and IMV, which could help in the management of COVID-19 cases at ICU.

Full Text
Due to technical limitations, full-text HTML conversion of this manuscript could not be completed.
However, the manuscript can be downloaded and accessed as a PDF. Figure 1 Study design. Day one represents the day of inclusion and rst sample collection when all participants were already under mechanical ventilation. Patients' intubation started two days before to four days after ICU admission (window of intubation). Blood samples were collected from ICU patients one day before ICU admission to ve days after ICU admission (window of ICU admission), then seven days later. Clinical and metabolic pro les were measured at day one and day seven and were correlated with four phenotypes: two continuous (days at ICU and days under mechanical ventilation) and two categorical (short (≤14 days) or long (>14 days) stay at ICU and progression to ECMO. Clinical outcomes were recorded at days one, seven, fourteen, twenty-one and sixty. Participants' data for age, BMI, days under mechanical ventilation and days at ICU are presented as mean ± standard deviation (SD). Hypoxanthine and Betaine with independent effects (d). Validation of the model using the prediction set (n=16) and assuming a hypothetical separation line (dashed line in red), the model only misclassi ed one ICU long stay patient (e). The AUC value from ROC curve analysis was 0.92 (f). Although the APACHE II score is signi cantly higher at day 1 in patients that remain at ICU for longer than 14 days (p=0.01, table1), in terms of discriminatory power it is inferior to our model (AUC=0.71, n=39) (g). Testing the model on published metabolomics data (28 healthy subjects, 25 non-COVID-19 patients, 25 non-severe COVID-19 patients, and 28 severe COVID-19 patients) revealed that the predicted scores from COVID19 patients are lower than controls, and similar to the lower predicted scores by ICU long stay patients when compared to short stay (* p value <0.001) (h). Data points were slightly scattered across the x-axis for ease of visualization in all boxplots.  . The predictive model was trained on metabolites and clinical traits measured from the training set (n=17) on day one (c), then validated on the prediction set (n=16) (d). The model comprised of three metabolites and one clinical trait (e) that together showed a better predictive power compared to APACHE II score (f). Using the model to predict the highly correlated number of days at ICU produced a correlation level of 0.66 with their observed counterparts (g). Figure 4