Aim of the present study was to establish easy predictors for a fatal outcome in critically ill COVID-19 patients by investigating the first two weeks of treatment at a high-care ICU of a university medical center. For this purpose, we first screened the current literature for any parameters to be used as possible predictors of good or bad outcome in patients with COVID-19; furthermore, we added parameters that we considered crucial in every day intensive care treatment. This way, a panel of 883 distinct parameters was generated for each patient if a complete dataset was available. The number of patients to be included was rather low because we focused on those COVID-19 patients who had required the highest-level ICU therapy (41 out of 59 patients included in the present study had been transferred from an ICU of a non-tertiary hospital to an ICU at our university medical center). Despite this special selection, the cohort of patients included in the present study did not differ to the extent that could have been expected in terms of baseline and demographic data as well as pre-existing comorbidities (1, 5).
Univariate analysis
In a first step, we checked all parameters suitable for predicting patient outcome. Some results were in line with published data, and some results were unexpected. For example, in contrast to data published by Choron et al., fever was no predictor of higher mortality in our study (9). A possible reason for this seeming discrepancy may be different definitions of fever. We classified each day with a peak temperature of ≥ 38°C to be a ‘fever day’, Choron et al. had considered the absolute peak temperature levels for comparing survivors and non-survivors.
In the context of vital signs, MAP was a highly significant predictor of patient outcome with lower values for non-survivors in our study. This finding is per se not surprising and in line with an obviously higher need for cardiocirculatory support of non-survivors who also required higher dosages of norepinephrine.
Serious courses of COVID-19 seem to be associated with a higher risk of venous thrombosis (13, 14). In our cohort, higher prophylactical dosages of unfractionated or low molecular weight heparin were more often administered to survivors than to non-survivors. However, the heterogeneity of different dosages of UFH and LMWH in context with the overall small number of patients in our cohort does not allow any further conclusions on specific therapy recommendations regarding the intensity of anticoagulation.
According to the initial recommendations (15, 16), COVID-19 patients should not be treated with steroids. This guideline was applied when generating our initial in-house standard for the ICU treatment of COVID-19 patients during the first wave of the disease (Supplement 1). As the patients included in the present study had been treated according to this initial standard and had therefore hardly received any steroids, we did not consider treatment with steroids to be a reasonable parameter for predicting outcome in our cohort. Meanwhile, the efficacy of dexamethasone in the treatment of COVID-19 patients with respiratory support has been demonstrated (17).
Analysis of the laboratory diagnostics showed lower blood pH, BE, and bicarbonate values for non-survivors. Lactate, chloride, and paCO2, however, did not significantly differ in a relevant matter between the two groups. Choron et al. found acidosis to be a predictor of mortality in mechanically ventilated COVID-19 patients (9).
Multivariable analysis and development of a model for predicting fatal outcome
Calculation of the daily medians of metric parameters and the presentation as box plots for the 14-day observartion period gives a visual impression of the question whether there is a difference between survivors and non-survivors. Such daily comparisons, however, may be falsified because patients were admitted to our ICUs at different stages of their COVID-19 course of disease: ‘day 1’ may be the day at which a patient first required ICU therapy but could also be the first day at our ICU after many days of ICU treatment in an external hospital. Therefore, mean or extreme values for each patient should be more conveniently considered in a multivariable analysis. In addition, the use of mean or extreme values does not implicate that patients who die earlier than 14 days after ICU admission have to be excluded from further calculation.
Because the number of 59 patients in our cohort is rather small for multivariable analysis and to avoid overfitting, the number of parameters selected for multivariable analysis was deliberately kept very low. We therefore only considered parameters that showed highly significant differences between the two groups for further examination. The derived simple formula correctly predicts fatal outcome with a very high model accuracy of 84.7%, which underlines the practicality of the chosen approach. As mean MAP and minimum blood pH values can easily be extracted from patient data management systems, the formula provides a simple option for predicting outcome in critically ill COVID-19 patients. Figure 2 (for MAP) and Fig. 3 (for pH) show the remarkable difference in both parameters between survivors and non-survivors over the entire observation period, indicating the high stability of the model. MAPmean and pHmin do not reflect any progressive divergence of the two parameters over time caused by proceeding degeneration of the patient status in non-survivors and improvement of the patient status in survivors.
Calculation of the linear regression model with only one parameter for pHmin, MAPmean and pHmean showed a highly predictive value for pHmean that almost reached the level of the above model including MAPmean and pHmin (AUC 0.901 for pHmean alone vs. AUC 0.945 for MAPmean and pHmin). This finding was not expected and underlines the particular strength of the pH value; lower pH values are a highly predisposing factor for a fatal outcome in critically ill COVID-19 ICU patients.
Nine patients from the non-survivors had died during the 2-week observation period. One could have assumed that a relevant drop in blood pH value and MAP could have occurred within the last days before death due to multiple organ failure, thus causing explicit differences in pHmin and MAPmean between both groups and falsification of the model. However, exclusion of the last 3 days before death in all non-survivors for calculation does not result in relevant differences regarding the prediction of fatal outcomes (Supplement 12).
Our multivariate model considering pHmin and MAPmean would have predicted death in three patients who, however, had survived. When taking a closer look at these patients, it can be noticed that in all cases the lowest blood pH value which is crucial for calculation had occurred at the very beginning of the observation period thus suggesting that these patients had been admitted in a very bad condition. In these three cases a longer interhospital transfer with limited options for extended intensive care and diagnostics was necessary prior to admission. It can be assumed that treatment could be optimized promptly after admission to the ICU thus leading to fast improvement of patients´ condition and that the initial values did not represent the status that would have been observed when extended ICU treatment could have been performed permanently.
Limitations
The overall number of patients included in the present study is rather low, thus making multivariable analysis difficult. To avoid overfitting, we considered only parameters for calculation that had clearly yielded highly significant differences between the two groups over time and restricted the number of parameters included in the model.
We could not provide any model for calculating of the probability for fatal outcome in critically ill COVID-19 patients by single values obtained at the very beginning of ICU treatment (e.g., certain blood values at admission to ICU) as crucial parameters for predicting the further course of disease. In fact, our static model can only be used two weeks after admission to the ICU. Nevertheless, being able to estimate the further course with a very high probability even after a period of two weeks of ICU treatment could be very helpful for daily clinical practice because this stage is typically the point in time to decide whether continuation of therapy is reasonable or not. However, a multimodal prognostication with not only considering the results of statistical calculations but also critical evaluation of the previous course of ICU treatment and the results of different kinds of diagnostics will be indispensable prior to withdrawal of life-sustaining treatment.
Our study cohort consisted exclusively of patients from the first wave of the COVID-19 pandemic. Patients from the second and third wave may have different characteristics to those of the patients included in the present study. We have already started a follow-up study that additionally includes the patients of the second and third wave in Germany to re-evaluate the established parameters.