Abstract
Background Evidence suggests that vascular inflammation and thrombosis may be important drivers of poor clinical outcomes in patients with COVID-19. We hypothesized that a significant decrease in the percentage of blood vessels with a cross-sectional area between 1.25–5 mm2 (BV5%) on chest computed tomography (CT) in COVID-19 patients is predictive of adverse clinical outcomes.
Methods Retrospective analysis of chest CT scans from 10 hospitals across two state in 313 COVID-19 positive and 195 COVID-19 negative patients seeking acute medical care.
Results BV5% was predictive of outcomes in COVID-19 patients in a multivariate model, with a BV5% threshold below 25% associated with an odds ratio (OR) 5.58 for death, OR 3.20 for intubation, and OR 2.54 for the composite of death or intubation. A model using age and BV5% had an area under the receiver operating characteristic curve 0.85 to predict the composite of intubation or death in COVID-19 patients. BV5% was not predictive of clinical outcomes in patients without COVID-19.
Conclusion This data suggests BV5% as a novel biomarker for predicting adverse outcomes in patients with COVID-19 seeking acute medical care.
Abstract
BV5% derived from chest CT may serve as an imaging biomarker for predicting adverse outcomes in patients with COVID-19 seeking acute medical care.
Background
The severe acute respiratory syndrome coronavirus 2 (SARS-COV-2) pandemic has placed substantial pressure on many aspects of the healthcare system [1], and patients presenting to the hospital with coronavirus disease 2019 (COVID-19) have significant risk for progression to respiratory failure or death [2, 3]. Increasing evidence suggests that vascular inflammation and thrombosis may be important drivers of poor clinical outcomes in patients with COVID-19 [4]. Autopsy studies demonstrate that pulmonary endothelialitis and microangiopathy are significantly more common in COVID-19 compared to other viral respiratory illnesses [5].
Changes in the pulmonary vasculature on chest computed tomography (CT) may reflect the angiocentric injury caused by COVID-19 [6]. Quantitative analysis of chest CT scans from patients with COVID-19 demonstrates a significant decrease in the percentage of blood vessels with a cross-sectional area between 1.25–5 mm2 (BV5%) [7]. Given the effects of COVID-19 on pulmonary microcirculation, we hypothesized that BV5% reduction on chest CT in COVID-19 patients is associated with adverse outcomes of respiratory failure requiring intubation or death.
Methods
This was an institutional-review-board-approved retrospective study of patients seeking acute medical care within a large integrated healthcare network from March 1, 2020–June 30, 2020. Patients presenting to the emergency department or directly admitted to the hospital and undergoing a chest CT within 24 h of presentation and COVID-19 testing were eligible for participation (fig. 1). After identification of the COVID-19-positive cohort, a randomly selected group of COVID-19-negative patients were chosen in order to achieve a target study ratio of 60% COVID-19 positive and 40% COVID-19 negative cases for analysis.
Flow chart of eligible patients. Computed tomography (CT).
Patient demographic information, treatment course, and outcome data was obtained from the clinical information system through a combination of custom coded data retrieval and manual abstraction. All COVID-19 testing was performed using nasopharyngeal swabs, with polymerase chain reaction (PCR) assays processed in a central laboratory. In patients with a negative COVID-19 PCR test, the medical record was checked for results of an additional COVID-19 test within 14 days. Initial blood laboratory values obtained within 24 h of the CT scan were recorded.
Semi-quantitative CT analysis
A fellowship trained cardiothoracic radiologist visually scored CT scans for the severity of lung opacification, blinded to the patient's COVID status and clinical outcome. A semi-quantitative scoring method correlating with disease severity and short term outcomes in COVID-19 patients was utilized [8, 9]. Each lobe was scored as follows: 0: no opacification; 1: <5% opacification; 2: 5–25% opacification; 3: 26–50% opacification; 4: 51–75% opacification; and 5: >75% opacification. The severity score (SS) for each CT was calculated as the sum of the five lobar scores (range 0 to 25).
Quantitative CT analysis
A recent publication provides a detailed description of the quantitative CT analysis technique utilized to calculate BV5% as well as the percentage of lung opacification (LO%) [10]. Briefly, CT scans with slice thickness ≤2 mm were used to generate 3D reconstructions of the lung tissue, airways, and pulmonary vasculature. An automated algorithm segmented the vasculature by identifying and enhancing cylindrical structures, and excluded airways via Hounsfield unit thresholds. Blood vessels were then clustered into three groups by cross-sectional area. BV5 denoted the volume of blood contained in vessels between 1.25–5 mm2, BV5-10 the volume contained in vessels between 5–10 mm2, and BV10 the volume contained in vessels above 10 mm2. BV5% was calculated as the percentage of blood volume in vessels between 1.25 and 5 mm2 relative to the total pulmonary blood volume, BV5-10% as the percentage of blood volume in vessels between 5–10 mm2, and BV10% as the percentage of blood volume in vessels above 10 mm2 (fig. 2).
Volume rendered images of chest CT scans from two different patients with COVID-19, with color coded segmentation of the pulmonary vascular cross sectional area. (a) Patient with BV5% of 21%; (b) Patient with BV5% of 55%. Red color denotes blood volume in vessels between 1.25–5 mm2, yellow color is vessels between 5–10 mm2, and blue color is vessels >10 mm2.
Statistical analysis
Patient demographic and clinical characteristics were stratified by COVID-19 testing results, reported as means and standard deviations for continuous variables or frequencies and percentages for categorical variables. The Wilcoxon Rank Sum Test was used to assess differences in continuous variables, while the Fisher's Exact Test was used to compare categorical variables. For the outcomes of mortality, intubation, and its combination, multivariable logistic regression was implemented to ascertain the odds of the outcomes relative to BV5% as a continuous and categorical variable respectively. BV5% was categorized using the Lowess Smoother plot. If the relationship between BV5% and the log odds of the outcomes were not linear, the inflection point was used to create the threshold for categorisation. Multivariable linear regression was also used to estimate the mean difference in hospital length of stay and the number of intubation days relative to BV5%. Statistical models were adjusted for covariates that were prognostic for outcomes; thus, age, gender, race, BMI, tobacco use, diabetes, hyperlipidemia, hypertension, heart failure, chronic kidney disease, COPD/Emphysema, cerebrovascular disease, cancer, contrast enhanced CT scan, CT findings, anticoagulation treatment, steroid treatment, azithromycin treatment, remdesivir treatment, white blood cell count, lymphocyte count, hematocrit, platelet count, and estimated glomerular filtration rate were included in the models. Missing data were handled using multiple imputation with ten imputations. Two-sided p-values <0.05 were considered statistically significant. Data analyses were conducted using STATA version 15 (STATA Corp: College Station, TX).
A logistic regression model from scikit-learn [11] was trained on COVID-19 positive patients to predict the risk of mortality, intubation, and its combination. To prevent over-fitting, model performance was assessed in two ways: 1) accuracy of predictions in leave-one-out cross-validation, and 2) area under the receiver operating curve (AUC) and precision-recall curves after splitting data into training (80%) and testing (20%) sets [12]. Differences in AUC were compared using a one-sample Z-test of proportions. With the coefficients and intercept determined from training, predictions of risk from the logistic regression model were used to simulate patients of varying age and BV5%.
Results
A total of 508 patients from 10 hospitals across Arizona and Colorado were included in the analysis, including 313 COVID-19 positive patients and 195 COVID-19 negative patients (table 1). 55% of patients were from Arizona (p=0.04). Compared to COVID-19-negative patients, COVID-19-positive patients were younger (54.9±17 versus 58.8±18.5 yrs; p=0.02), from a racial/ethnic minority (p<0.01), had higher body mass index (33.5±12.3 versus 30.7±8; p<0.01), and more likely to be diabetic (30.9% versus 21%; p=0.01). COVID-19 positive patients had a lower incidence of tobacco use (32% versus 53.3%; p<0.01), COPD/emphysema (4.8% versus 14.4%; p<0.01), heart failure (3.8% versus 11.8%; p<0.01), coronary artery disease (6.4% versus 12.3%; p=0.02), and cancer (4.2% versus 9.2%; p=0.02). Additional demographic data is reported in table 1.
Baseline demographic, laboratory, and treatment parameters in patients with and without COVID-19
Radiology reports from chest CT scans stated findings of COVID-19 or atypical/viral pneumonia in 91.1% of COVID-19 positive patients and 32.3% of COVID-19 negative patients (p<0.01) (table 2). A normal CT scan was reported in 3.2% of COVID-19 positive patients and 24.6% of COVID-19 negative patients (p<0.01). There was no significant difference in the frequency of pulmonary embolism (p=0.45). Patients with COVID-19 were significantly less likely to have findings of non COVID-19 pneumonia (2.9% versus 13.9%; p<0.01), pulmonary edema (1.3% versus 5.1%; p=0.01), emphysema (0.6% versus 9%; p<0.01), tumor (0.3% versus 5.6%; p<0.01), aspiration/bronchitis (0.3% versus 4.6%; p<0.01), pleural effusion (0.3% versus 6.7%), or pulmonary infarct (0% versus 2.1%; p=0.02).
Computed tomography (CT) findings in patients with and without COVID-19
On quantitative CT analysis (table 2), patients with COVID-19 had significantly lower BV5% compared to COVID-19 negative patients (25.3%±7.4 versus 30.1%±9.6; p<0.01). BV5% was also significantly lower for the subset of patients with COVID-19 and CT findings of COVID-19/atypical pneumonia (23.7%±7.3 versus 27.2%±27.2; p=0.02). Patients with COVID-19 had significantly higher SS compared to patients without COVID-19 (9.7±5.4 versus 6.9±5.3; p<0.01). For patients with COVID-19, BV5% had a moderate correlation with SS (Spearman's rho −0.45, p<0.0001) (fig. 3). Differences in BV5% and Spearman's rho stratified by CT findings are reported in the supplement. There were no significant differences in total lung volume (3034±1203 mL versus 2839±1123 mL; p=0.06) and total blood volume (210±72 mL versus 215±72 mL; p=0.45) in COVID-19 positive versus COVID-19 negative patients, respectively. CT scans from COVID-19 patients had a lower peak area (9.4%±3.1 versus 11.4±9.4; p<0.01) and weight mean (17.9%±4 versus 19.4±4.3; p<0.01), and higher area under the curve (272.8±32.9 versus 264.7±35.7; p=0.02).
Spearman's rank correlation between BV5% and the visually assessed CT severity score, stratified by COVID-19 status and CT findings. Percentage of blood volume in vessels between 1.25 and 5 mm2 relative to the total pulmonary blood volume (BV5%).
There was no significant difference in BV5% obtained from contrast enhanced CT versus non-contrast CT (p=0.23). Average processing time for BV5% was 9 min and 22 s (±6 min and 3 s), with processing time dependent on scan quality. No CT scans were unable to be analyzed.
Overall clinical outcomes for patients with COVID-19 were worse than for patients without COVID-19 (table 3). COVID-19 patients were hospitalised for a longer duration (8.5±10.8 days versus 4.1±5.7 days; p<0.01), more likely to be intubated (20.8% versus 6.2%; p<0.01), and trended towards higher in-hospital mortality (11.8% versus 7.2%; p=0.09).
Outcomes in patients with and without COVID-19
In a multivariate regression analysis that did not include control for lung opacification, BV5% remained significantly associated with intubation and death in patients with COVID-19 (table 4). Specifically, BV5% as a continuous variable had an odds ratio (OR) 0.87 for mortality (95% CI 0.79, 0.96; p<0.01), OR 0.89 for intubation (95% CI 0.84, 0.95; p<0.01), and OR 0.9 for the composite mortality or intubation (95%CI 0.84, 0.96; p<0.01). A BV5% threshold of 25% had an OR 5.58 for mortality (95% CI 1.54, 20.1; p<0.01), OR 3.20 for intubation (95% CI 1.55, 6.63; p<0.01), and OR 2.54 for the composite mortality or intubation (95%CI 1.15, 5.60; p=0.02). In patients without COVID-19 there was no significant association between BV5% and intubation or mortality.
Multivariate regression analysis of outcomes relative to BV5% as a continuous and categorical variable in patients with and without COVID-19
BV5% was not predictive of length of hospitalisation or duration of intubation in patients with COVID-19. Length of stay was associated with BV5% for COVID-19 negative patients, OR 0.97 (95%CI 0.96, 0.98; p<0.01).
After adding the visually assessed CT severity score to the multivariate regression analysis, a BV5% threshold of 25% remained significantly associated with mortality in patients with COVID-19, with an OR 4.27 for mortality (95% CI 1.02, 17.8; p=0.046). In the subset of patients without COVID-19 but with CT findings of pneumonia, BV5% as a continuous variable had an OR 1.24 for the composite mortality or intubation (95%CI 1.01, 1.52; p=0.039).
For COVID-19 positive patients, a logistic regression model using age and BV5% were chosen as features for the classifier because, after excluding treatments, they had the highest coefficients in the aforementioned multivariable linear regression for the primary outcomes. A model using age and SS was similarly created. Age and BV5% predicted the composite of death or intubation with an accuracy of 0.83 (±0.02) in iterative cross-validation. Age and SS predicted the composite of death or intubation with an accuracy of 0.82 (±0.02) in iterative cross-validation. After training on 80% of the patients, the AUC 0.85 for age and BV5% and the AUC 0.87 for age and SS were not significantly different (p=0.18; fig. 4a). The model was also used to forecast risk for the composite of intubation or death using simulated data of patients between 50–80 years old and a BV5% ranging from 10%–60% (fig. 4b).
(a) Receiver operating characteristic curve from the testing set of the logistic regression model. Patient age and BV5% have AUC 0.85 to predict the composite of intubation or death in patients with COVID-19, not significantly different than AUC 0.87 for age and SS (p=0.18). (b) Probability of the composite death or intubation for simulated COVID-19 patients ages 50, 60, 70, and 80 with BV5% ranging from 10% to 60%. Percentage of blood volume in vessels between 1.25 and 5 mm2 relative to the total pulmonary blood volume (BV5%). Visually assessed CT severity score (SS).
Conclusions
In this study of 508 patients presenting to the hospital, tested for COVID-19, and undergoing chest CT within 24 h of presentation, BV5% from COVID-19 patients was significantly lower than BV5% from a heterogenous cohort of patients without COVID-19. This difference was driven mainly by patients with CT findings of COVID-19/atypical pneumonia. BV5% was predictive of outcomes in COVID-19 patients in a multivariate model that did not account for lung opacification, with a BV5% threshold <25% associated with OR 5.58 for death, OR 3.2 for intubation, and OR 2.54 for the composite of death or intubation. After including the severity of lung opacification in the multivariate analysis, a BV5% threshold of 25% remained significantly associated with mortality, with an OR 4.27.
In healthy patients BV5% constitutes the majority of pulmonary blood volume distribution [13], and is closely correlated with histologic assessment of vascular cross sectional area [14]. Alterations in BV5% are not exclusive to COVID-19, and can manifest in other diseases that diffusely affect pulmonary perfusion, such as COPD [15] and ARDS [7]. In these cohorts, chronic vascular remodeling and small vessel loss are the underlying etiology for reduced BV5% [16]. In COVID-19 the reduction in BV5% may reflect the acute sequela of microcirculatory disruption induced by SARS-CoV-2 [5], however the extent to which other processes such as ventilation/perfusion mismatch and shunting impact BV5% are unknown. The significant association between reduced BV5% and mortality in COVID-19 suggests that BV5% may offer insights into the underlying pathological processes involved in SARS-CoV2 infection, and potentially serve as a tool for quantifying the extent of these processes in an acute care setting.
Models using BV5% and SS had high AUCs for the composite of intubation or death in COVID-19 patients, supporting their role as imaging surrogates for disease severity. However, BV5% and SS are only moderately correlated, and the presence of a low BV5% in patients with a low SS may signify that changes in pulmonary vascular volume distribution represent either an early indicator of disease severity or a distinct (if low-frequency) phenotype. Since SS is a general measure of lung opacification in COVID-19, it can be influenced by alveolar disease (e.g. diffuse alveolar damage), interstitial changes (e.g. fibrosis), vascular impairment (e.g. hemorrhage), or a combination of these processes [17–19], thus posing challenges in using SS to identify targeted treatments. On the other hand, because BV5% is closely correlated with histologic assessment of vascular cross sectional area [14], we speculate that BV5% may be able to inform possible treatment pathways for COVID-19, for example by identifying patients that would benefit from anticoagulation.
Based on a pilot study of 97 patients (not published), the association between BV5% and outcomes in COVID-19 patients was significant only for CT scans performed within 24 h of presentation. Beyond this window an increasing number of variables likely obscure the prognostic impact of BV5%, especially given the absence of clear guidelines regarding hospitalisation at the time the study data was collected during the initial phase of the pandemic. This may also explain why BV5% was not predictive of duration of hospitalisation or duration of intubation in COVID-19 patients. Paradoxically, BV5% was associated with length of stay in patients without COVID-19. In addition, after including the severity of lung opacification in the multivariate analysis, BV5% was associated with the composite of mortality and intubation. These are subjects of current investigation.
There are several important limitations of this study. This was a retrospective study in which only a subset of patients with suspected COVID-19 underwent chest CT with the requisite ≤2 mm slice thickness. However, the results should be generalisable due to the geographically and demographically diverse population. Only one fellowship trained cardiothoracic radiologist performed the semi-quantitative visual severity scoring of CT scans, and it is unknown if scoring from radiologists with different levels of expertise would have the same prediction for clinical outcomes. Although only patients presenting to a hospital for acute medical care were included, these are the patients at greatest risk for poor outcome with COVID-19 [20]. It is possible that some patients with COVID-19 were miscategorised due to false negative PCR results [21], however we attempted to minimize this effect by checking for additional COVID-19 PCR testing within 14 days of the initial encounter. In addition, obtaining BV5% requires transporting the patient to a CT scanner. While limiting the movement of patients with COVID-19 throughout the hospital can mitigate exposure of the disease to healthcare personnel, the potential predictive benefit of BV5% may outweigh this risk.
As additional waves of the SARS-COV-2 pandemic are expected, there is urgent need for improved diagnostic, prognostic, and treatment tools. Although certain risk factors [22–24] and imaging findings [9] are associated with worse prognosis in COVID-19, accurately quantifying a patient's individual risk for progression to respiratory failure or death remains challenging. BV5% can be quickly derived from CT scans in patients with COVID-19 seeking acute medical care, without the need for contrast material. We hypothesize that BV5% is a unique imaging biomarker, in that it can potentially be used as a gatekeeper for identifying patients that may benefit for earlier or more aggressive therapy with anticoagulation, and we hope this study serves as a catalyst for prospective evaluation of this hypothesis.
Acknowledgements
The authors wish to thank Heather Pryzbyl RN and Corry Glade for assisting with the custom coding used for data retrieval. We thank Kathryn Olson MD for feedback on the manuscript.
Footnotes
This article has supplementary material available from erj.ersjournals.com
Support statement: No funding was received for this research.
Conflict of Interest: Dr. Glassberg reports and MKG serves on the advisory board for Bellerophon Therapeutics.
Conflict of Interest: Dr. De Backer reports other from FLUIDDA, outside the submitted work.
Conflict of Interest: Dr. Ridenour has nothing to disclose.
Conflict of Interest: Dr. Lanclus reports other from FLUIDDA, outside the submitted work.
Conflict of Interest: Dr. Kang has nothing to disclose.
Conflict of Interest: Dr. Godon reports other from FLUIDDA, outside the submitted work.
Conflict of Interest: Dr. Pershad has nothing to disclose.
Conflict of Interest: Mr. Lavon reports employment with FLUIDDA, outside the submitted work.
Conflict of Interest: Dr. Morris has nothing to disclose.
- Received November 9, 2020.
- Accepted February 3, 2021.
- ©The authors 2021.
This version is distributed under the terms of the Creative Commons Attribution Non-Commercial Licence 4.0. For commercial reproduction rights and permissions contact permissions{at}ersnet.org
