Risk Assessment of Hospitalized Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2)–Infected Patients Using Laboratory Data and Immune Cell Morphologic Assessment

In this observational study, we sought to validate reported data relating to common clinical laboratory abnormalities in COVID-19 patients admitted or presenting to our hospital system. By tracking outcomes, we also aimed to identify what laboratory parameters might be outcome predictive. We also sought to devise and validate a laboratory parameter–based outcome prediction model, incorporating peripheral smear morphology, which could serve to assist with the triaging of hospital COVID-19 patients alongside common clinical decision tools.

Research ethics board review and approval, guided by the principles of the Declaration of Helsinki, was obtained (CHREB20-1164). This study was designed as an observational cohort study; however, we obtained waiver of patient consent from our research ethics board, given that all data and materials described herein were collected for clinical purposes and no specific clinical intervention was undertaken. Immediately after data collection, all unique patient identifiers were removed.

Beginning in March 2020, SARS-CoV-2 testing was routinely performed in our laboratory network, allowing for the rapid identification of clinical and laboratory data from COVID-19–infected patients. Adult patients (ie, age at least 18 years) were considered for inclusion in this study on the basis of available peripheral smear(s) and laboratory data obtained within 14 days of a positive SARS-CoV-2 test result. All patients considered for inclusion in the study were presenting to or already admitted to hospital with confirmed SARS-CoV-2 infection (including hospital-acquired illness). There were no specific exclusion criteria. Patients diagnosed during the months of March 2020 to June 2020 were assigned to an initial discovery set, with a second independent validation set collected from August 2020 to November 2020. These data periods correspond to our region's first and second waves of the pandemic, respectively. Of note, regional SARS-CoV-2 variant identification data were not consistently available, although public health data suggest that the wild type was dominant in our region until April 2021 (written personal communication; July 5, 2021).

Electronic medical records were reviewed, with an extensive list of demographic, clinical, radiologic, and interventional parameters recorded. COVID-19 disease severity at presentation and discharge was scored in accordance with the World Health Organization (WHO) COVID-19 risk stratification tool.¹⁹  Available laboratory data were retrieved from our regional electronic records system. A synopsis of the queried clinical, laboratory, and hematology data is presented in Supplemental Table 1 (see supplemental digital content at https://meridian.allenpress.com/aplm in the January 2022 table of contents).

SARS-CoV-2 infection testing was performed by using a laboratory-developed reverse transcription–polymerase chain reaction (RT-PCR) assay (as described in detail by Pabbaraju et al²⁰ ) and validated against Canadian National Microbiology Laboratory Standards. In a small subset of COVID-19 patients, blood SARS-CoV-2 viral copy numbers were also derived from droplet digital PCR (ddPCR). For ddPCR analysis, viral nucleotides were extracted with the QIAamp Viral RNA Mini Kit (QIAGEN Inc, Germantown, Maryland), followed by analysis with the Bio-Rad SARS-CoV-2 Droplet Digital PCR Kit (Bio-Rad, Pleasanton, California). By way of fluorescence intensity, specimens were quantified for each of the N1 and N2 viral nucleocapsid protein RNA and presence of the human RNAseP gene, with analysis performed with the QuantaSoft 1.7 and QuantaSoft Analysis Pro 1.0 Software. Specimens with at least 1 replicate with 0.1 copies/μL or more and 2 or more positive droplets for N1 and N2, with 5000 or more droplets total were considered positive.

Routine laboratory data were generated on, or through a combination of, the COBAS 8000 (Hoffman-La Roche, Basel, Switzerland), GEM5000 (Instrumentation Laboratory, Bedford, Massachusetts), ACL TOP (Instrumentation Laboratory, Bedford, Massachusetts), and/or Sysmex XN 9000 (Sysmex Corp, Kobe, Japan) analyzers.

Peripheral smears were digitally scanned with the CellaVision DC-1 digital morphology analyzer (CellaVision AB, Lund, Sweden), with settings for a 500-cell differential count. Occasional specimens with fewer than 500 detectable cells were not specifically excluded. Automated cell identification and preclassification was then performed with the CellaVision Peripheral Blood Application Software, followed by manual verification. This latter step was undertaken independently by local hematopathologists, with validity assessed by Bland-Altman analysis of independent reviews by external hematopathologists using the CellaVision Remote Review Software. Informed by recent literature, we separated atypical lymphocytes (henceforth referred to as “covidocytes,” and defined in this study to consist of plasmacytoid lymphocytes, immunoblastic cells, and plasma cells) and large granular lymphocytes from the larger lymphocyte differential. Other special morphologic findings, including platelet clumping, apoptotic/necrobiotic forms, smudge cells, and red cell morphologic changes, were also recorded where identified. Where possible in the discovery set, analysis of multiple samples from patients during the course of their hospital stay was also performed.

The duration of patient follow-up was limited to discharge from the presenting hospital admission. Clinical classification into critical and noncritical outcomes was performed. Critical outcomes included death, admission to intensive care unit, mechanical ventilation, vasopressor support, or extracorporeal membrane oxygenation support required during the course of the index hospital admission.

Statistical analyses were performed in SPSS (v. 26; IBM Corp, Armonk, New York) and figures generated in GraphPad Prism (v. 9; GraphPad Software, LLC, San Diego, California). Advanced data modeling was performed in MATLAB (v. R2018a; The MathWorks, Inc, Natick, Massachusetts). In general, proportions were compared by using the binomial test. Differences in categorical variables were assessed by using the χ² test; differences between continuous variables were assessed by using the Student t test. Classifier performance was assessed by using receiver operating characteristic (ROC) area under the curve (AUC) analyses, with ROC-AUC values compared between different models by the method of Hanley and McNeil.²¹  Parameters such as sensitivity, specificity, and likelihood ratios were calculated where applicable. Ninety-five percent confidence intervals were reported for relevant parameter estimates; P values of less than .05 were considered statistically significant.

The study participant flow diagram is outlined in Figure 1. In the discovery set, 107 patients were initially considered for inclusion; of these, 59 patients had peripheral smear and laboratory data procured within a 14-day window from RT-PCR detection of SARS-CoV-2 infection. Within the discovery set, furthermore, 17 critical outcomes occurred. Subsequently, in the validation set, 112 patients were considered, with 54 patients meeting inclusion criteria; within the validation set, 18 critical outcomes occurred. Across the entire cohort, ages at presentation ranged from 20 to 98 years (mean, 58 years), and patients with male biological sex predominated (68 of 113, or 60%). Although ages were not seen to differ significantly between the 2 subsets, more males were present in the validation set than in the discovery set (38 of 59 or 70% versus 30 of 54 or 51%, respectively). Significant demographic differences by outcome were not seen, however.

Patients presented to hospital most frequently with cough (66 of 113 or 58%), dyspnea (66 of 113 or 58%), malaise/fatigue (63/113 or 56%), and fever/chills (61 of 113 or 54%). Anosmia/ageusia were described in only 9 of 113 patients (8%). Patient history parameters noted to be more frequent in critical outcome patients included cardiac history (any type, including coronary artery disease and rhythm disorders; χ² = 5.9, P = .02), history of chronic kidney disease (χ² = 18.0, P < .001), and thrombotic disease history (χ² = 5.1, P = .03). Patients with critical outcomes were also more likely to present with acute chest x-ray findings (of any type; χ² = 11.2, P = .001), altered level of consciousness (of any type; χ² = 18.7, P < .001), and with higher COVID-19 triage screening scores (χ² = 35.9, P < .001).

For each evaluated clinical laboratory parameter, upper and lower limits of normal (ULN and LLN, respectively) were retrieved. A parameter was deemed to be significantly deranged in the setting of COVID-19 if, by the univariate binomial test, the proportion of cases with parameter values higher than the ULN or lower than the LLN was significantly different from an expected proportion in this range of .05, or if the proportion of cases with parameter values outside of either ULN and LLN was significantly different from an expected proportion of 0.1. The laboratory parameters noted to fall outside of local normal ranges in a significant proportion of cases are highlighted in Supplemental Figure 1, A through H. Parameters such as creatinine (Cr), urea, and the international normalized ratio of prothrombin time, in particular, were noted to be frequently elevated in our cohort. CBC parameters, including hemoglobin, hematocrit, and red blood cell (RBC) and platelet counts, were reduced in a significant proportion of cases (see Supplemental Figure 2, A through E). In addition, the proportion of cases with abnormal white blood cell (WBC) counts (either <4000/μL or >11 000/μL) was significant. Not surprisingly, the proportion of cases with abnormal absolute neutrophil counts (either <2000/μL or >9000/μL) was significant as well (see Supplemental Figure 3, A through F). The proportion of our cohort presenting with absolute lymphopenia and monocytosis was significant; however, significant abnormalities of eosinophil or basophil counts were not seen.

Interestingly, the proportion of cases presenting with absolute large granular lymphocyte counts in excess of 300/μL (the ULN in our system²² ) was also significant, and these seemed to be enriched in cases without lymphopenia (absolute count t = 5.0; P < .001). We identified a significant proportion of cases (53 of 113 or 47%) with extremes of smudge cells (ie, >15% of WBCs). We also noted that platelet aggregates (defined to account for at least 2% of WBCs, with at least 3 platelets per aggregate) were identifiable in 45 of 113 cases (40%). Figure 2, A through D, highlights a selection of covidocytes (defined in this study to consist of plasmacytoid lymphocytes, immunoblastic cells, and plasma cells); the distribution of covidocyte counts is highlighted in Figure 3, A and B (percentage of WBCs in Figure 3, A, and absolute counts in Figure 3, B). Supplemental Table 2 lists the laboratory parameters in which a significant difference was identified by univariate testing between the groups of cases stratified by outcome. To ensure that our covidocyte counts were reproducible, we compared covidocyte counts by using a Bland-Altman analysis (see Supplemental Figure 4, A and B) between internal and external assessors, and concluded that our assessments were generally within tolerance.

From the totality of available laboratory data in the discovery set, we then derived a coarse decision tree classifier, calculated against critical/noncritical outcome (Figure 4). These results suggested that a simple decision tree approach, using a combination of RBC, Cr, urea, and covidocyte percentage, could inform a compelling risk stratification. More specifically, by this decision tree, patients with RBC parameters less than 3.15 × 10⁶/ μL, low Cr (ie, <1.02 mg/dL) but high urea (≥39.9 mg/dL), or high Cr (ie, ≥1.02 mg/dL) but with low covidocyte counts (ie, <0.3%) were classified as “high risk” for a critical outcome.

The Table lists the test parameters obtained when the decision tree is applied to both the discovery and validation sets. Figure 5, A through C, highlights the ROC analyses performed over the discovery, validation, and composite patient sets. While an element of overfitting by our model is noted in the validation set relative to the discovery set, the AUC confidence intervals of the appertaining ROC analyses are seen to overlap. We also compared the COVID-19 clinical classification scheme proposed by the WHO¹⁹  to our model and found that the 2 classifiers performed comparably (ie, no significant difference in paired AUC comparisons was observed over our validation set: WHO versus Lab-Value Classifier z = 1.7, P = .08). Similarly, in deference to other reports citing its significance,^14,15  we compared the neutrophil-lymphocyte ratio (NLR) classifier to our own, and found that the 2 performed comparably well (ie, no significant difference in paired AUC comparisons was observed over our validation set: Lab-Value Classifier versus NLR z = −0.86, P = .39).

We further explored the potential association of covidocyte counts with duration of hospitalization. For nondeceased patients in the discovery set with multiple available peripheral smears (N = 19), we derived best-fit lines of covidocyte percentage versus day of admission. In this manner, best-fit lines with positive slopes indicated a trend to increasing covidocytes during the course of a given patient's admission (and vice versa). The resulting slope values were then clustered into 2 groups by unsupervised k-means clustering (ie, relative low [N = 15] and high [N = 4] values). Membership into these high and low categories was seen to separate nondeceased patients into relatively shorter and longer hospital stays, respectively (t = −2.2, P = .045; see Figure 6).

We also investigated whether there might be an association between the degree of viremia and number of covidocytes. Blood SARS-CoV-2 viral load data by ddPCR were available from 10 cases in our cohort. All specimens were positive for SARS-CoV-2, with a range of viral copy number from 0.61 to 12.11 copies/μL of extracted RNA. Corresponding covidocyte counts ranged from 0% to 7% but without an obvious association to viral load.

Despite strident public health measures, improved supportive care, and the availability of effective vaccines, COVID-19 remains a significant global public health concern. Given the ongoing impact that COVID-19 has on developing countries especially, simple and effective tools to risk-stratify COVID-19 patients are still invaluable. To assist in this effort, we undertook a detailed observational study, incorporating independent discovery and validation sets, to identify what laboratory parameters might be clinically useful. In contrast to other recently published studies,^12,17,23  we undertook a combined analysis of the clinical findings, laboratory changes, and peripheral blood morphology features in a cohort of COVID-19 patients of significant size, with incorporation of outcome data.

Our data validate the reports of several other studies, inasmuch as a variety of laboratory parameters are deranged in the setting of COVID-19 infection. Our data support other publications that identified derangement of CBC parameters, as well as biomarkers of renal and liver disease, as commonplace in the setting of COVID-19 infection. In line with the substantive published evidence suggesting that these findings are typical and may be adversely prognostic in COVID-19, we did note significant proportions of patients presenting with thrombocytopenia and lymphopenia.^5,6,16  Interestingly, when compared to other laboratory factors and by way of multivariate analyses, we did not see strong evidence to support thrombocytopenia or lymphopenia as adversely prognostic. This may relate to differences in our study population of interest, as well as differing cutoffs and analytical methods relative to other studies.

By way of a comprehensive analysis of numerous laboratory parameters, furthermore, we derived and validated a simple yet compelling clinical risk prediction model. This model is easy to use, and when compared to the WHO COVID-19 risk stratification tool that incorporates clinical and radiologic data,¹⁹  as well as the NLR model,^14,15  our model performs comparably. Clinicians may benefit from the availability of an alternative laboratory-based model when other triaging models are equivocal.

Despite the numerous laboratory parameters noted to be significantly deranged in COVID-19, we arrived at a fairly restrictive subset of parameters that appears to be outcome predictive. In particular, while much attention is often paid to hemoglobin as an important predictor of outcomes in a variety of disease states, relatively recent data suggest that the volume-standardized RBC parameter might be more predictive of cardiovascular reserve in some circumstances.^24,25  Similarly, while a number of liver biomarkers were clearly deranged in our cohort, it is the renal biomarkers of creatinine and urea that appear to be the most robust predictors of good outcome in COVID-19.

Our data suggest that “covidocytes” play a role in driving good outcomes, both inasmuch as they were relatively increased in our good-outcome patients and that they trended upwards in patients with relatively shorter hospital stays. These results reflect the similar observations made by others.^26,27  Given that our blood viral load data do not support the idea that covidocytes represent peripheralized SARS-CoV-2–infected cells, the presence of increased covidocytes might instead serve to represent the elaboration of a more appropriate immune response.

Important limitations of our study should be highlighted. Most notably, we incorporated automated differential methods and cell morphology analysis; these technologies may not be available to some health systems. To this end, others wishing to apply similar morphology-informed clinical prediction models might need to establish locally specific cutoffs. It also bears noting that SARS-CoV-2 variant subtyping was not available to us during the initial period of patient accrual. Indeed, the variant forms of SARS-CoV-2 have become the dominant strains only well after patient accrual to our study ceased. As such, we were unable to account for the impact of variant-specific transmissibility and infectivity on clinical features, laboratory parameters, and outcomes. Finally, despite the relatively recent high uptake of vaccination in our and other regions, we were unable to explore the relevance of our laboratory risk model on vaccinated patients diagnosed with COVID-19, as our study was conducted before widespread vaccine availability.

In summary, we performed a systematic assessment of laboratory parameters in hospitalized COVID-19 patients, arriving at a simple outcome-predictive model. By way of our model and the other results described in our study, we provide additional evidence to support the clinical significance of atypical lymphocytes in the systemic immunologic response to infection with SARS-CoV-2.

The authors gratefully acknowledge the hematology laboratory staff for their collection and collation of laboratory data and peripheral smears. We are also grateful for the use of the DC-1 digital morphology analyzer, kindly provided by CellaVision AB for this study.