In this single-center prospective observational study we identified biomarkers related to inflammation and the antiviral immune response that reflect the disease course regarding organ injury and clinical outcome. Two patient characteristics are often associated with increased risk of a critical course and higher mortality in COVID-19: age [4, 5, 21] and male sex [21, 23]. In our study, age had no influence on mortality, which could however be explained by the relatively low median age of 59 years and small IQR of 11 years. We still included age into our multivariate models as part of the CCI, which accounts for comorbidities—another risk factor [4, 21, 23] and age. A higher CCI showed a non-significant trend to higher mortality and impaired QoL in survivors. As in previous studies [21], significantly more patients in our ICU were male than expected by gender distribution in society. However, within this cohort, there was no obvious sex-associated risk.
Our findings concerning general markers of inflammation are in line with previous reports. Leukocytes are known to be elevated in COVID-19 patients to healthy controls and in more severe cases [6, 10]. 1st leukocytes in our study strongly correlated with maximum and mean leukocyte counts, as well as with 1st PCT, showing a connection between inflammation at admission and during ICU treatment. 1st leukocytes mainly correlated to impaired lung function. Multivariate correlations however suggest that the first leukocytes value is no accurate predictor of disease severity.
The first PCT value on the other hand, was a good predictor of disease severity—being a significant risk factor for ECMO/ECLS and dialysis—and, in combination with sex and CCI, formed the best predictive model of QoL after 6 months. Hu et al. showed a temporary PCT increase at the beginning of COVID-19-associated hospitalization and a second increase in non-survivors until death [24]. Combined with our findings this could suggest a correlation of temporary PCT increase with disease control, while higher PCT values are associated with bacterial superinfection [25], resulting in increased lung and secondary organ injury as well as impaired QoL. However, missing influence of bacterial superinfection at admission on 1st PCT or bacterial superinfection at any time on QoL suggests a self-contained predictive value of PCT. Judgment of bacterial superinfection before admission to our ICU was hindered as many lower-level wards started calculated antibiotic therapies without attempting pathogen detection.
C-reactive protein is also often quoted as predictive factor [5, 6, 10], however it was only measured in 20 of our patients—compared to 45 patients with PCT measurements—and therefore not considered.
As TLR3 is responsible for recognizing viral RNA [9], it is surprising that we did not find an increased systemic expression compared to healthy controls. Nevertheless, increased TLR3 expression in blood showed to be a strong predictor for a less complicated course with less inflammation, less impaired lung function, better neurological short- and long-time outcome and a better QoL after 6 months. Menezes MC et al. reported a similar result; however, they were limited to TLR3 expression in blood [10]. Even though TLR3 expression in BAL was higher than in blood, it did not correlate to the clinical course. This is surprising, as a correlation between the expression in the primary spot of viral infection and the clinical outcome would seem to be likely and TLR3 induction in lung is known to help lung recovery in COVID-19 [26]. The reason for this finding needs further investigation. Potential explanations could be a discrepancy between mRNA and protein expression and TLR3 induction in the lung that was not sufficient for local virus control -hence the complicated COVID-19 course- but strong enough for systemic containment of disease.
The discrepancy in prognostic power between TLR3 expression in blood and BAL as well as the importance of blood TLR3 expression in our model for systemic damage highlights the relevance of TLR3 for a successful systemic rather than a local immunoreaction. The correlation of TLR3 expression in the blood with mRS after 14 days, at discharge and after 6 months but not at admission could hint at the relevance of systemic TLR3 for eradicating the virus during ICU treatment. However, this might be bought at the cost of thromboembolic events, as the latter could be caused by TLR3-induced endothelial dysfunction [27].
Systemic IL-8 was elevated in COVID-19 and correlated with impaired lung function and neurological impairment. However, these correlations were not significant in multivariate models, despite IL-8 blood being a significant predictor of ECMO/ECLS in ROC analyses. IL-8 in BAL was significantly higher than IL-8 in blood, which highlights the function of IL-8 to attract neutrophiles to the lungs [7, 28, 29], where they are found to be elevated in severe COVID-19 [30]. Consistently, IL-8 BAL was relevant in our model for local damage.
The correlation of IL-8 in blood and BAL with mRS at admission and systemic IL-8 with mRS after 14 days but no correlation of IL-8 with mRS at discharge or after 6 months suggests a more prominent role in the initial state of COVID-19 disease [8]. Whether IL-8 or IL-6 has better predictive relevance is still being discussed [4, 7, 8]. Our results suggest that this should be looked at with a differentiation between local and systemic damage rather than differentiating the use of IL-6 or IL-8 according to the phase of disease [8], especially as both parameters strongly correlated in blood and BAL.
Elevated IL-6 has been acknowledged as important marker for complicated courses of COVID-19 [4,5,6]. However, elevation of cytokine concentration in blood was shown to be less distinct than in other severe illnesses, potentially making the description as ‘cytokine storm’ exaggerated [31,32,33]. Our study backs the hypothesis of IL-6 measurement in BAL being more complicated but also more specific than measurement in blood [34]. IL-6 in BAL showed more correlations to clinical outcome, is part of our model for systemic damage and correlates to CO-Hb max and mean, unlike IL-6 in blood.
Correlation analyses within this study were intended to give indications about the importance of a biomarker for “local” and “systemic” damage as well as QoL. As interference with other biomarkers and demographic markers is likely, they were not intended to be used as predictive models for themselves. Likewise, ROC analyses were a further step towards developing our final three models.
Our models for local and systemic damage as well as QoL after 6 months generate useful tools to predict the most relevant clinical outcomes after assessing just four biomarkers and two easy to assess patient characteristics. For most of the questions, the models possessed an AUC > 90% and a Tjur’s R2 > 0.5, highlighting their accuracy. Nevertheless, there are also limitations to our models. IL-6 and IL-8 show an Odds Ratio of approximately 1 despite being often-proven risk factors and giving the models higher accuracy. Using cytokine secretion in BAL rather than in blood results in higher AUC and higher R2 (Additional file 1: Table S4). One theory could be, that patients for which BAL was taken shortly after admission had more critical courses and therefore secretion in BAL is likely to be different to the other patients. This argument does not exactly fit the situation in our ICU, where the regime as to when to collect a BAL changed during the study recruitment phase and was not always based on the clinical situation alone. Furthermore, even when comparing only those patients for which secretion in both, BAL and blood, was assessed, models using cytokine secretion in BAL had a higher AUC and R2 (Additional file 1: Table S4). Assumptions about the function of a specific marker must be taken with care, since the only significant markers are intercept and 1st PCT for QoL after 6 months. Since most of the markers, especially systemic TLR3 expression, are no part of routine laboratories in hospitals, implementation into clinical practice is not without difficulties.
Our models differ from most models published so far (for example [4, 6, 7, 11]) as our models are based on biomarker measurement after admission to ICU and therefore measurements were done at a later point in time with an already more severe disease course. We deliberately chose a different approach in a patient population already requiring maximum-level intensive care in a cross-regional center for ARDS treatment. Many of the patients showed a prolonged LOS ICU with phases of improving and phases of stagnating or even deteriorating clinical status. Medical personnel were often looking for prognostic markers to assess the likely outcome in times of missing clinical improvement. By providing such markers, our models could possibly not only help with ICU resource allocation but could also guide treatment decisions on an individual basis. While models targeting an early stage of disease and allowing for prediction of necessary ICU treatment are very helpful, our models could complement them by predicting clinical course in ICU and long-term QoL after severe COVID-19.
Within our models, outcome was categorized into “local” and “systemic” damage as well as QoL. We believe they all answer different questions and therefore recommend establishing the whole set of predictive markers for each case. This approach seems feasible as all models have several markers in common. If establishing the whole set of markers is not possible, adaptions to the most pressing clinical questions can be made. While the model for local damage could predict the need for resource intensive ECMO/ECLS and dialysis, it does also provide the markers required for predicting long-term QoL. The model for systemic damage, however, could predict the clinically important questions of potential thromboembolic events and mortality.
Considering the characteristic CO-Hb dynamic, lacking correlation between HO-1 and clinical course as well as literature findings so far [14, 35], predictive power of CO-Hb at admission seems very limited.
We could however show that the increase, which was also observed in previous studies [36, 37], was not just associated with a more complicated clinical course but could be actively used as a predictor. CO-Hb mean > 2% seems to be the more accurate predictor for death in ROC analysis than CO-Hb max \(\ge\) 3%, however both strongly correlate and in contingency analysis there was no difference in RR. As it is easier to assess whether a marker exceeded a threshold at any time than to assess whether the mean measurements exceeded a threshold, use of CO-Hb max \(\ge\) 3% should be considered for re-evaluating risk of death in ICU as well as indicating a prolonged stay in ICU. The increase is likely to be a sign of stress and dysregulated immune response [15]. Unlike observations in other diseases [12, 38], CO-Hb min does not seem to be a predictor in COVID-19.
There are further limitations to our study. Limiting clinical assessment to the first 14 days in ICU might improve differentiation between consequences of inflammation caused by COVID-19 and consequences of prolonged ICU-stay. However, it might also hide some clinical phenomena caused by COVID-19. Because of missing matched controls, comparisons between controls and COVID-19 patients must be taken with care. The very homogenous measurements among our controls, low age in our patient cohort and apart from TLR3 in blood missing influence of age on biomarker expression could hint at a reduced influence of age and sex on the baseline. Therefore, it is unlikely that the observed differences result from differences in age and sex alone. Since our study was a prospective single-center study at a maximum-level ICU with ECMO/ECLS therapy, patients are likely to show more severe clinical courses while not having contraindications to ECMO therapy such as high age. However, this makes our findings potentially even more important for very critical courses of COVID-19. Additionally, only 21 BALs were available at admission, which is limiting the power of our study to show differences in biomarker expression in BAL and their implications for clinical course. Since there is only a small number of cells and therefore limited mRNA in BAL, incorrect measurement in qPCR is possible. Lastly, there was no validation cohort for our prognostic models, which could lead to overestimating their predictive power. It must be pointed out that most of the correlations shown in this paper are only weak correlations (r < 0.5). Despite these limitations, there are also strengths to our study. We focused on finding biomarkers, which could differentiate between different courses of COVID-19-patients in ICU. These markers enable the medical personnel to allocate resources to the places where they are likely to be needed and to differentiate treatment. Our findings also advocate a more regular use of cytokine secretion in BAL as predictive marker and the use of CO-Hb as easy to assess parameter for re-evaluating clinical course while in ICU. Despite not finding new biomarkers, we tried to generate self-contained models explaining the mechanisms leading to multiple complications often seen at ICU in COVID-19 treatment.