Demographic and clinical characteristics
The study cohort comprised 146 patients for whom blood samples were taken within 48 hours of hospital admission. The cohort consisted of a heterogeneous group of non-acute and acutely ill patients with varying phases and degrees of inflammatory response, as well as a range of comorbidities. Of these patients, 39 had COVID-19 severity assignments of “critical”, 70 had assignments of “severe”, 27 were assigned as “moderate” and 10 as “mild” by the criteria of Table 1. Of the 146 patients in the cohort, 43 (29%) were admitted to ICU and of these, 10 (23%) were classified as severe and 33 (77%) were classified as critical. Table 2 summarizes the demographic and clinical characteristics of the patients in the study cohort, stratified by their severity assignments as adjudicated by the clinical care team at Hospital del Mar. Supplementary Table 2 summarizes the interventions provided to the patients in the cohort. The severe and critical patients required more interventions to manage their clinical trajectories as compared to the moderate and mild cases. Hospital del Mar had a well-established protocol in place involving the early administration of corticosteroids.
Table 2 Demographic and Clinical characteristics of patients stratified by COVID severity. The severity categories were defined according to the WHO COVID-19 Guidance 2021 [19]. There were 146 patients in this cohort. Recorded values are medians (interquartile ranges). p-values were calculated with one-way ANOVA.
Parameter
|
Mild
(n=10)
|
Moderate
(n=27)
|
Severe
(n=70)
|
Critical
(n=39)
|
Critical & Died
(n=13)
|
p-value
|
Age
|
76 (64-77)
|
59 (50-66)
|
64 (51-79)
|
69 (62-78)
|
79 (66-84)
|
0.002
|
Sex (M/F)
|
4/6
|
14/13
|
35/35
|
26/13
|
9/4
|
0.3
|
Race / ethnicity
|
White = 8
Asian = 1,
Unknown = 1
|
White = 21
Asian = 2
Unknown = 4
|
White = 49
Asian = 12
Black = 1
Unknown = 8
|
White = 24
Asian = 9
Black = 1
Unknown = 5
|
White = 9
Asian = 3
Black = 1
|
0.41
|
% SpO2 min
|
94 (93.25-95)
|
94 (94-95)
|
91 (88-93)
|
89 (86-91)
|
88 (86-91)
|
<0.001
|
|
|
|
|
|
|
Respiratory Rate, breaths/min (max)
|
23
|
22
|
28
|
28
|
28
|
<0.001
|
(18-26)
|
(20-24)
|
(22-32)
|
(24-32)
|
(24-32)
|
|
WBC (cells/mm3)
|
5540
|
6090
|
6230
|
6690
|
8700
|
0.5
|
(4712-6530)
|
(4995-7380)
|
(4900-7360)
|
(5250-8715)
|
(6830-8900)
|
|
Neutrophil/
|
4.9
|
3.1
|
4.3
|
6.1
|
8.9
|
0.004
|
Lymphocyte ratio
|
(2.6-5.9)
|
(2.5-5.0)
|
(2.8-6.5)
|
(3.8-9.5)
|
(5.6-12.0)
|
|
Lactate (mmol/L)
|
1.57
|
1.23
|
1.4
|
1.54
|
1.66
|
0.009
|
(1.23-2.43)
|
(1.07-1.45)
|
(1.07-1.7)
|
(1.21-2.04)
|
(1.47-2.36)
|
|
Creatinine (mg/dl)
|
1.22
|
0.81
|
0.85
|
1.08
|
1.31
|
0.005
|
(0.90-1.85)
|
(0.69-0.95)
|
(0.7-1.06)
|
(0.88-1.42)
|
(1.10-2.37)
|
|
CRP (mg/L)
|
6
|
5
|
7
|
9
|
7
|
0.020
|
(0-7)
|
(3-8)
|
(3-10)
|
(5-14)
|
(5-14)
|
|
IL-6 (pg/ml)
|
16.7*
|
35
|
29
|
37
|
40
|
0.7
|
(21-44)
|
(7-51)
|
(15-84)
|
(33-73)
|
|
D-dimer (mg/L)
|
890
|
590
|
645
|
720
|
2530
|
0.053
|
(710-1025)
|
(420-760)
|
(415-1128)
|
(430-1015)
|
(735-7642)
|
|
SeptiScore
|
4.9 (4.5-6.1)
|
6.0 (5.8-6.8)
|
6.8 (5.9-7.7)
|
7.1 (6.6-8.0)
|
7.1 (6.7-8.3)
|
<0.001
|
*IL-6 measured for only one Mild patient, so IQR could not be calculated
Principal component analysis
Principal Component Analysis (PCA) [26] is a powerful tool to help visualize information contained in complex datasets that are highly multidimensional, and to identify key variables. We conducted a PCA in which ICU admission vs. non-admission was the variable chosen as the basis for class separation. A list of 46 quantitative variables used to define the dimensions of the PCA is given in Supplementary Table 3. Fig. 1 shows a “biplot” which overlays the individual patients (represented by dots) with vectors (represented by arrows) which indicate the direction and strength of contributions of different single variables to the class separation along PCA dimensions 1 and 2. The first two dimensions of the PCA explain 30.5% of the variation in data. The length of each arrow from the origin is proportional to the contribution of its variable to the construction of the given dimension. When the angle between two arrows is small, the associated variables are highly correlated: for example, SeptiScore is correlated with SeptiCyte Band as would be expected, and is also correlated with other clinical variables that are elevated in the sepsis phenotype or in patients admitted to the ICU, such as C-reactive protein (CRP), D-dimer, interleukin-6 (IL-6), procalcitonin (PCT), and indicators of the systemic inflammatory response syndrome (SIRS). When the angle between arrows is at approximately 90 degrees there is no correlation between the variables, as can be seen e.g. in the comparison of lymphocyte/neutrophil ratio and blood pressure.
The PCA analysis provides evidence that patients admitted to the ICU have a higher SeptiScore compared to patients that were not admitted to the ICU; and that SeptiScore contributes more than other parameters such as D-dimer, IL-6, and CRP toward the construction of the principal components for identifying patients that needed ICU admission.
Stratification by clinical severity
Fig. 2A presents the distribution of SeptiScores across the four COVID-19 clinical severity categories defined in the WHO COVID-19 Guidance 2021 [19], using the assessments provided by the care team at Hospital del Mar. This figure shows that the median SeptiScores were significantly higher for the critical and severe groups than for the moderate and mild groups.
A further stratification analysis by hospital location (ICU vs. non-ICU) is presented in Fig. 2B, in which SeptiScores are stratified based on whether or not the patients were admitted to the ICU (in addition to their COVID severities). It was found that relatively few critical cases (8/41 = 19.5%) were managed outside the ICU. While the SeptiScores of the critical group do not appear to differ between ICU vs. non-ICU locations, the SeptiScores of the severe group are significantly higher for patients in ICU (n=9) vs. patients not in ICU (n=61) with AUC 0.76 and p-value 0.026 for this comparison.
As indicated in Fig. 2C, if the critical and severe patients who were admitted to the ICU (red, n=40) are compared to the mild and moderate cases none of whom were admitted to the ICU (blue, n=37), the performance of SeptiCyte RAPID has an AUC of 0.80 and p-value of 2.48 x 10-6. Also, the median SeptiScores of critical or severe COVID cases that were not admitted to the ICU (grey, n=69) are lower than those admitted to the ICU (red, n=40), but higher than the mild or moderate cases not admitted to the ICU (blue, n=37).
A ROC test comparison (Table 3) indicated that SeptiCyte RAPID was significantly better at distinguishing the critical+severe ICU patients from the moderate+mild non-ICU patients, as compared to similarly timed D-dimer, lactate, IL-6 or creatinine measurements which are typically used to assess the severity of patient clinical trajectories. The AUC of SeptiCyte RAPID (AUC = 0.81) was also found to be higher than that of CRP (AUC = 0.67), although this comparison was on the edge of statistical significance (p-value = 0.067 by bootstrap resampling).
Table 3 Performance of SeptiCyte RAPID relative to other clinical laboratory tests. Critical + severe cases in ICU (n=40) were compared to moderate + mild cases not in ICU (n=37). Values in parentheses ( ) represent the % of patients in each clinical category for whom the relevant laboratory test results were available. There were some missing values as indicated by the percentages in the table. AUC values were compared to the SeptiCyte RAPID AUC with a bootstrap method as described in [23], and p-values for the AUC comparisons are indicated.
Lab Test
|
Critical+Severe
in ICU
|
Moderate+Mild
not in ICU
|
AUC ± SE
|
p-value
|
SeptiCyte RAPID
|
n = 40 (100)
|
n = 37 (100)
|
0.81 ± 0.050
|
|
CRP
|
n = 39 (97.5)
|
n = 36 (90)
|
0.67 ± 0.063
|
0.067
|
Lactate
|
n = 35 (87.5)
|
n = 28 (70)
|
0.62 ± 0.073
|
0.024
|
Creatinine
|
n = 40 (100)
|
n = 37 (100)
|
0.60 ± 0.065
|
0.009
|
IL-6
|
n = 30 (75)
|
n = 15 (37.5)
|
0.56 ± 0.087
|
0.012
|
D-dimer
|
n = 38 (95)
|
n = 32 (80)
|
0.51 ± 0.070
|
0.0004
|
Comparative performance of SeptiCyte RAPID and IL-6
Although it is not current practice at the Hospital del Mar study center, IL-6 levels above 35 pg/ml have been used elsewhere to ascertain the need for mechanical ventilation in COVID-19 patients [27]. Fig. 3 shows the performance of SeptiCyte RAPID relative to IL-6 in a subset of 77 patients for whom both values were measured within 24 hours of hospital admission. Panels A, C show the distribution of SeptiScores across varying COVID-19 severities, whereas Panels B, D show the IL-6 levels in pg/ml for patients across the same clinical severities.
The performance of SeptiCyte RAPID for distinguishing critical (n=25) from moderate (n=14) COVID cases (Panel A) has an AUC of 0.73 (t-test p=0.06) and a Kolmogorov-Smirnov D value of 0.528 (p=0.007). In contrast, the same discrimination with IL-6 (Panel B) has AUC 0.55 (t-test p= 0.3) and a non-significant Kolmogorov-Smirnov D value of 0.275 (p=0.43).
If moderate + mild (n=15) cases are grouped together, and similarly for critical + severe cases (n=62), then the performance of SeptiCyte RAPID for distinguishing between these composite groups (Panel C) has AUC of 0.70 (t-test p=0.066), with a Kolmogorov-Smirnov D value 0.495 (p=0.003). In contrast, the performance of IL-6 (Panel D) is AUC 0.51 (t-test p=0.8), with a non-significant Kolmogorov-Smirnov D value of 0.265 (p=0.33).
It is known that IL-6 levels can be modulated by corticosteroid treatment [28-30] whereas SeptiCyte RAPID appears unaffected [31]. As the Hospital del Mar patients generally were treated with corticosteroids early during their transit through hospital, this might provide an explanation for the absence of diagnostic power for IL-6, as opposed to SeptiCyte RAPID, in this study cohort. Further consideration of this point is presented in the Discussion.
Predicting ICU admission
Figure 4 provides evidence that SeptiCyte RAPID could potentially be used to predict the need for ICU admission. Some of the critical and severe cases had “early” blood draws (i.e. before ICU admission; red, n=30). When these were compared to moderate and mild cases not in ICU (black, n=37), SeptiCyte RAPID had an AUC of 0.78 (p=0.00012). This comparison suggests that a high SeptiScore, measured early (before a patient is considered for ICU admission), might predict the need for later ICU admission.
For reference, there were also critical or severe cases which had “late” blood draws (i.e. after ICU admission; blue, n=10). When these were compared to moderate and mild cases not in ICU (black, n=37), SeptiCyte RAPID had an AUC of 0.88 (p=0.002). Here, a high SeptiScore measured after a patient was already in ICU most likely describes the patient’s current state, as opposed to being a predictor of a future state. In these ICU patients, a high SeptiScore indicated a sepsis phenotype, including caused by either a viral or bacterial infection.
SeptiCyte RAPID in relation to oxygen therapy
In two exploratory analyses, we assessed the correlation between elevated SeptiScores and the need for enhanced oxygen therapy during a patient’s clinical trajectory. We considered therapeutic escalation along the following axis: no oxygen needed, low flow oxygen therapy, high flow oxygen therapy, mechanical ventilation (MV), mechanical ventilation + intubation (MV+i), and extracorporeal membrane oxygenation (ECMO). Progression along this axis would indicate a deteriorating clinical trajectory.
In a first analysis, we asked whether elevated SeptiCyte RAPID scores correlated with the need for high flow oxygen therapy. This level of therapy falls between no/low flow oxygen, and the more extreme therapies involving mechanical ventilation, intubation, or ECMO. Figure 5 compares the group needing high flow oxygen (red, n=63) to those patients not needing supplemental oxygen (black, n=17). The performance of SeptiCyte RAPID in this comparison had AUC 0.87 (p = 1.75 x 10-8). The AUC between the groups that needed high flow (red, n=63) vs. low flow O2 (blue, n=69) was 0.62 (p = 0.008). Finally, the AUC between the groups that needed low flow oxygen (blue, n=69) vs. those that did not need oxygen supplementation (black, n=17) was 0.75 (p = 7.5 x 10-5).
In a second analysis, we conducted a multivariable linear regression to explore whether the SeptiCyte RAPID score, in combination with other clinical parameters that typically would be measured early during a patient’s transit through the hospital, correlated with the degree of oxygen therapy administered. We represented the extent of need for oxygen therapy by a “weighted oxygen therapy index” (Y) defined by the following formula:
Y = 1 (days low-flow O2) + 2 (# days high-flow O2) + 3 (# days non-invasive MV) +
4 (# days MV+intubation) + 5 (# days ECMO) (Eq. 1)
Thus, as the oxygen therapy became increasingly extreme or invasive, or drawn out in duration, it was weighted more heavily. The clinical input variables {Xi} that were considered were among those which typically would be obtained upon hospital admission or shortly thereafter, that related to COVID-19 disease presentation, and that had few if any missing data in our dataset:
X1 = age, X2 = hypertension, X3 = SeptiScore, X4 = WBC count, X5 = neutrophils, X6 = lymphocytes, X7 = CRP, X8 = D-dimer, X9 = SpO2 max (within 72h) (Eq. 2)
A linear regression analysis Y = f{Xi} returned a significant fit (right-tailed, F(1,32) = 7.429, p-value = 0.0022) to the following model: Y = b0 + b3X3 + b4X4. Since p-value < α (0.05), we reject the null hypothesis H0 and conclude that SeptiScore (X3) and WBC count (X4) contribute significantly to the need for extreme oxygen therapy in this cohort.
We note that the SeptiScore measurements were taken on either side of the time when the oxygen therapy was first administered. Because of this timing variability, it was not possible to determine if SeptiScore had predictive capabilities, with respect to the need for escalatory oxygen therapy. Thus, in these analyses, SeptiScore is most likely indicative of the current state of patients, rather than predictive of future states.
Discriminating between discharge and death
Finally, we examined whether an early SeptiCyte RAPID measurement could discriminate between patients who ultimately were discharged or who died. A modest predictive ability (AUC 0.70, p = 0.02) was observed. In the 13 patients that died, SeptiScore was measured between 1 and 52 days prior to their death. In the 133 patients that were discharged, a SeptiScore was drawn between 0 and 55 days prior to their hospital discharge. A precise interpretation of these results may be difficult, however, because of the long time gaps that occurred in some cases between blood draw and the event of interest (discharge or death).