key: cord-0703625-g304b9i9 authors: Patel, Naomi J.; D'Silva, Kristin M.; Li, Matthew D.; Hsu, Tiffany Y‐T.; DiIorio, Michael; Fu, Xiaoqing; Cook, Claire; Prisco, Lauren; Martin, Lily; Vanni, Kathleen M.M.; Zaccardelli, Alessandra; Zhang, Yuqing; Kalpathy‐Cramer, Jayashree; Sparks, Jeffrey A.; Wallace, Zachary S. title: Assessing the Severity of COVID‐19 Lung Injury in Rheumatic Diseases versus the General Population Using Deep Learning‐Derived Chest Radiograph Scores date: 2022-03-21 journal: Arthritis Care Res (Hoboken) DOI: 10.1002/acr.24883 sha: 82dda9a6482e5e61ea28bbc1d81237ec27f108f3 doc_id: 703625 cord_uid: g304b9i9 OBJECTIVE: COVID‐19 patients with rheumatic disease have a higher risk of mechanical ventilation than the general population. We assessed lung involvement using a validated deep learning algorithm that extracts a quantitative measure of radiographic lung disease severity. METHODS: We performed a comparative cohort study of rheumatic disease patients with COVID‐19 and ≥1chest radiograph within ±2 weeks of COVID‐19 diagnosis and matched comparators. We used unadjusted and adjusted (for age, Charlson Comorbidity Index, and interstitial lung disease) quantile regression to compare the maximum Pulmonary X‐Ray Severity (PXS) score at the 10(th)‐90(th) percentiles between groups. We evaluated the association of severe PXS score (>9) with mechanical ventilation and death using Cox regression. RESULTS: We identified 70 patients with rheumatic disease and 463 general population comparators. Maximum PXS scores were similar in the rheumatic disease patients and comparators at the 10(th)‐60(th) percentiles but significantly higher among rheumatic disease patients at the 70(th)‐90(th) percentiles (90(th) percentile score of 10.2 vs. 9.2, adjusted p=0.03). Rheumatic disease patients were more likely to have a PXS score >9 (20% vs. 11%, p=0.02), indicating severe pulmonary disease. Rheumatic disease patients with PXS scores >9 vs. ≤9 had higher risk of mechanical ventilation (HR 24.1 [95% CI: 6.7, 86.9]) and death (HR 8.2 [95% CI: 0.7, 90.4]). CONCLUSIONS: Rheumatic disease patients with COVID‐19 had more severe radiographic lung involvement than comparators. Higher PXS scores were associated with mechanical ventilation and will be important for future studies leveraging big data to assess COVID‐19 outcomes in rheumatic disease patients. This article is protected by copyright. All rights reserved. Patients with rheumatic diseases and COVID-19 may have higher risk of mechanical ventilation and other poor outcomes versus general population comparators who also have COVID-19; factors driving these associations may include differences in comorbidity burden and immunosuppressive medication use. [1] [2] [3] [4] [5] [6] These associations suggest that rheumatic disease patients have more severe COVID-19 lung disease, but the pulmonary manifestations of COVID-19 remain poorly understood in this population. Among patients with symptomatic COVID-19 in the general population, radiographic severity correlates with severe COVID-19 outcomes. 7 However, there can be significant variability in the interpretation and grading of disease severity on radiographs. A convolutional Siamese neural network-based algorithm was previously validated to estimate COVID-19 radiographic severity on portable chest radiographs, yielding Pulmonary X-ray Severity (PXS) scores. 8 PXS scores correlate with manual assessment of lung severity (combined evaluation of density and extent of lung opacities) by multiple radiologists, and PXS scores from hospital admission radiographs were predictive of subsequent mechanical ventilation or death within three days. 8 Higher radiographic severity of COVID-19, as indicated by higher PXS scores, may indicate more severe subsequent pulmonary complications of COVID-19 in rheumatic disease patients, explaining some of the observed associations between these conditions and higher risk of mechanical ventilation and/or death than the general population. Additionally, since rheumatic disease patients are more likely than the general population to have baseline chronic lung disease, it is unclear whether PXS scores may reliably predict risk of severe outcomes in this population. If so, deep learning-derived x-ray scores may be leveraged for use in larger studies evaluating the relationships between rheumatic disease, immunosuppression, imaging results, and COVID-19 outcomes. In this study, we examined whether patients with rheumatic disease and COVID-19 have worse PXS scores than general population comparators with COVID-19 and aimed to validate the use of the PXS score in rheumatic disease patients. From a previously described cohort of patients at Mass General Brigham with confirmed rheumatic disease and COVID-19 1,2 , we identified patients ≥18 years of age seen at Massachusetts General Hospital or Brigham and Women's Hospital (tertiary care hospitals in Boston, Massachusetts) between January 31, 2020 and January 31, 2021 who had chest radiographs performed within ± 2 weeks of their initial positive COVID-19 test result. This study was approved by the MGB Institutional Review Board (2020P000833). Patients were not involved in the design, conduct, or reporting of this study. Each rheumatic disease patient was matched to up to 10 comparators without rheumatic disease; comparators also had confirmed COVID-19 (by polymerase chain reaction [PCR], or less commonly antigen testing) and a chest radiograph performed within ± 2 weeks of COVID-19 diagnosis. Rheumatic disease patients were matched based on age (± 5 years), sex, hospital site at which the chest radiograph was performed (to account for technical differences), and the date of positive COVID-19 test result (± 14 days, to control for calendar time since testing criteria, hospital capacity, and treatment strategies changed over time). The date of the earliest positive COVID-19 test was used as the index date. All patients included in the study had at least one chest radiograph within ± 2 weeks of COVID-19 diagnosis. Chest radiographs were assessed using a previously validated convolutional Siamese neural network algorithm to assess COVID-19 lung disease severity, which had been further tuned and tested for both anterior-posterior and posterior-anterior frontal chest views. 8, 9 In this model, the inputs are pixel-level image data from frontal chest radiographs, and the output is a quantitative score of COVID-19 lung severity--the PXS score (https://github.com/QTIM-Lab/PXS-score). The PXS score quantifies the extent as well as the density of any pulmonary opacities and is not specific for COVID-19-related pulmonary findings. PXS scores range from approximately 0 to 24, with higher PXS scores indicative of more severe radiographic lung disease (Figure 1 ). We defined a PXS score of ≤ 9 as indicating non-severe disease and a score > 9 as indicating severe disease based on previously determined thresholds. 10 Prior data have shown 5fold higher odds of subsequent intubation or death within 3 days of hospital admission in the general population with PXS score > 9 compared to all patients n with COVID-19. 10 For those patients with multiple radiographs, we used the maximum PXS score within ± 14 days of COVID-19 diagnosis for each patient for the primary analysis; we also performed secondary analyses using initial PXS scores. For those who were mechanically ventilated, we used the maximum PXS score prior to the time of mechanical ventilation in the analysis assessing the risk of mechanical ventilation with high versus low maximum PXS score. For the patients with rheumatic disease, clinical variables of interest were extracted from the electronic health record (EHR) by manual record review. These included the rheumatic disease diagnosis and duration, immunomodulatory medications at the time of COVID-19 diagnosis (including specific dose of glucocorticoids when applicable), and disease activity level (based on global assessment by the treating provider as documented in the EHR). For rheumatic disease patients and comparators, additional variables were extracted from the COVID-19 Data Mart, an EHR-based data enclave established by MGB that includes all patients diagnosed with COVID-19. Variables extracted from the COVID-19 Data Mart included location (outpatient, emergency department, or inpatient) of the maximally scored chest radiograph, site of COVID-19 test, demographics (age, sex, and self-identified race/ethnicity), smoking status, and medical comorbidities. 11 Baseline characteristics including comorbidities and body mass index (BMI) were assessed in the 1 year prior to the index date, and the Charlson Comorbidity Index (CCI) was calculated from diagnosis codes in the EHR using any data available prior to the index date. 12 COVID-19 clinical outcomes, specifically mechanical ventilation and death, were also extracted from the COVID-19 Data Mart. Categorical variables were presented as number (%) and continuous variables were presented as mean ± standard deviation (SD) or median ± interquartile range (IQR), as appropriate. Continuous variables were compared using a two-sample t-test for continuous normally distributed variables or Wilcoxon test for continuous non-normally distributed variables. Categorical variables were compared using Chi-square tests. For each group (rheumatic disease patients and comparators), we determined the PXS score at the 10 th , 20 th , 30 th , 40 th , 50 th , 60 th , 70 th , 80 th , and 90 th percentiles. We then compared the PXS scores between matched rheumatic disease patients and comparators using unadjusted and adjusted quantile regression at each percentile. We initially used unadjusted analyses since rheumatic disease patients and comparators were matched on a number of relevant potential confounders (e.g., age, sex, time of COVID-19 infection), and other covariates (e.g., comorbidity scores) are likely mediators of the association between the exposure (rheumatic disease versus no rheumatic disease) and the outcome (PXS score). We also performed adjusted quantile regression, adjusting for age, Charlson Comorbidity Index (dichotomized by <2 or ≥2), and interstitial lung disease in the 12 months prior to COVID-19 diagnosis. We used Poisson regression to compare the number of chest radiographs per subject. For both the time from COVID-19 diagnosis to initial or maximum PXS score and median initial PXS score, we used a Wilcoxon test for unadjusted analyses and quantile regression for adjusted analyses. We used a Chi-square test to compare the site where the maximally scored chest radiograph was performed in unadjusted analyses and multinomial regression in adjusted analyses. In unadjusted and multivariable adjusted quantile regression, we added an "id" variable to indicate matched sets in the MODEL statement to account for the correlation among subjects within the matched set. In the analysis comparing the outcomes of mechanical ventilation and death between those with maximum PXS score > 9 (indicating severe lung disease) to those with a maximum PXS score ≤9 within ± 14 days of COVID-19 diagnosis, the index date was considered to be the time of either the initial or maximum PXS score. Person-months of follow-up were determined for each subject from the index date to the first of any of the following events: occurrence of the outcome of interest, death, or date of the last encounter in MGB at the time of data analysis. We calculated incidence rates as events/1,000 person-months. Cox proportional hazard regression models were used to estimate HRs and 95% CIs for the outcomes of mechanical ventilation and death, comparing patients with rheumatic diseases with maximum PXS score > 9 to those with a maximum PXS score ≤9. We also performed a secondary analysis among rheumatic disease patients evaluating those with an initial PXS score > 9 to those with an initial PXS score ≤9 as well as a secondary analysis among general population comparators evaluating those with either an initial or maximum PXS score > 9 to those with an initial or maximum PXS score ≤9, respectively. The level of significance was set as a two-tailed p<0.05, and statistical analyses were completed using SAS statistical software (version 9.4; SAS Institute, Inc.). Out of 624 patients with rheumatic disease and confirmed COVID-19 who were seen at In 70 rheumatic disease patients, there were a total of 206 chest radiographs analyzed within two weeks of COVID-19 diagnosis date compared with 463 comparators who had 936 chest radiographs in the same time frame ( Table 3 Among patients with rheumatic disease (n=70), those with a maximum PXS score >9 (n=14) compared to those with PXS scores ≤9 (n=56) had higher rates of mechanical ventilation (578.9 vs. 2.7 events/1,000 months; HR 24.1 [95% CI: 6.7, 86.9], p<0.001) ( Table 4) . Those with PXS scores >9 had numerically higher rates of death compared to those with PXS scores ≤9, although this was not statistically significant (7.5 vs. 0.9 events/1,000 months; HR 8.2 [95% CI: 0.7, 90.4], p=0.09). These differences in outcomes between those with higher and lower PXS scores also persisted when we excluded patients with ILD from the rheumatic disease patient and comparator groups. Similar results were seen when using initial instead of maximum PXS score. Among patients with rheumatic diseases, those with an initial PXS score >9 compared to those with PXS Table 4 ). Using a previously validated deep learning-derived scoring algorithm, we found that patients with rheumatic disease and COVID-19 had more severe radiographic lung disease than matched comparators with COVID-19. Our results provide validation of the use of both the initial and maximum PXS score among rheumatic disease patients and corroborate previous studies suggesting increased risk for mechanical ventilation. As rheumatic disease patients may be more susceptible to severe lung involvement from COVID-19 than the general population, 1, 2 additional studies are needed to determine how PXS scores may be helpful in clinical research studies studying risk factors for poor COVID-19 outcomes in rheumatic disease and other patients. Our results emphasize the importance of COVID-19 risk mitigation strategies, such as vaccination with an additional dose and post-exposure prophylaxis with monoclonal antibodies, in patients with rheumatic diseases due to their increased susceptibility to severe pulmonary involvement from COVID-19. patients with COVID-19. In both training and validation data sets, PXS scores correlated well with radiologist-determined radiographic lung disease severity (Pearson correlation coefficient of 0.86), and the initial PXS score was predictive of subsequent intubation or death (receiver operating characteristic area under the curve = 0.80). 8 The PXS algorithm has been further tuned using inpatient and outpatient COVID-19 data sets in the U.S. and Brazil. 9 After an intervention in which radiologists were exposed to machine learning-derived PXS scores of different chest radiographs, the interrater agreement among subsequent diagnostic radiologists improved (Fleiss kappa coefficient 0.40 to 0.66), indicating that PXS scores could be integrated into the radiology workflow to standardize interpretation of COVID-19 chest radiographs without requiring additional training of radiologists. 10 findings (e.g., intracranial hemorrhage, infarction, or leukoencephalopathy) and found that they had significantly higher mean PXS scores than comparators (9.2 vs. 5.0, p<0.001). 13 Gibson et al found that right ventricular strain is common in intubated patients with COVID-19 but is not associated with radiographic lung disease. 14 We found that initial PXS scores were similar between rheumatic disease patients and general population comparators but that maximum PXS scores were higher in rheumatic disease patients; this underscores the similar nature of both groups at baseline as well as the natural history of acute COVID-19 which often does not reach peak severity until a week or more into the clinical course. 15 Several factors may explain why rheumatic disease patients are at higher risk for worsened lung disease over the course of acute COVID-19. First, patients with rheumatic diseases are commonly treated with immunosuppressive medications that might increase the severity of infection. Second, rheumatic diseases are often associated with chronic systemic inflammation which might predispose these patients to more severe hyperinflammatory responses to COVID-19. Indeed, our prior work demonstrated that, compared with general population patients, those with rheumatic diseases and COVID-19 had a more severe inflammatory state. 16 Third, some rheumatic diseases are associated with inflammatory lung disease (e.g., interstitial lung disease) and the trigger underlying this manifestation is unknown. Infectious etiologies, like COVID-19, may be responsible for precipitating inflammatory lung disease in predisposed rheumatic disease patients. These hypotheses deserve additional study. Notably, the differences in maximum PXS scores between groups persisted after adjusting for comorbidities, suggesting that either the underlying rheumatic disease or possibly immunosuppressive medications may contribute to these differences. Additionally, as a PXS score difference of >1 between two x-rays can be distinguished by a radiologist, these differences are likely to be clinically meaningful on a population level. Our study has some strengths and limitations. We performed a systematic interpretation of chest radiographs at two academic hospitals, and patients were matched by site to account for differences in chest radiograph technique and other site differences such as triage and preference of ordering radiographs. Details regarding rheumatic disease diagnosis were available for all patients by manual chart review. Most patients received their care within the MGB system, thus reducing missing data, and data were obtained systematically from a centralized data warehouse. Our study also has certain limitations. Chest radiographs were not available for all patients in our initial rheumatic disease cohort; thus, our sample may be biased toward those with more severe disease or in whom the clinician was concerned about more severe disease. However, we would not expect access to chest radiographs to be misclassified between rheumatic disease patients and comparators. It is possible that clinicians may have been more likely to order radiographs for rheumatic disease patients (while the median number of chest radiographs is 1 in each group, the interquartile range is 1-3 in rheumatic disease patients compared to 1-2 in comparators), which would have biased results towards the null and would not explain a higher proportion of severe scores in the rheumatic disease patients. We did not have pre-COVID-19 radiographs available on all patients so could not evaluate changes from baseline. While we had ILD diagnosis available, it is possible that some rheumatic disease patients had subclinical lung damage from ILD prior to COVID-19 that contributed to the findings though our results remained unchanged in sensitivity analyses excluding patients with known baseline ILD. Additionally, our study was conducted at two tertiary care hospitals in the same geographic region, which may limit generalizability. The small sample size reduced our ability to assess the impact of potential confounders or mediators on the relationship between rheumatic disease and PXS scores and to determine the significance of differences in deaths between those with severe and non-severe scores. Finally, our sample size limited our ability to perform subgroup analyses of patients with different rheumatic diseases or on different immunosuppressive medications. Nevertheless, our findings indicate that patients with rheumatic diseases have more severe pulmonary radiologic manifestations of COVID-19 and that higher PXS scores identify those most likely to experience subsequent respiratory decompensation. In conclusion, we used a machine learning algorithm that assessed the severity of chest xray findings and found that patients with rheumatic disease were more likely to have more severe COVID-19 related lung disease. Our study confirms that rheumatic disease patients are more susceptible to lung involvement from COVID-19 and corroborates previous findings suggesting increased risk for mechanical ventilation and mortality. The higher risk of severe pulmonary involvement from COVID-19 provides additional support for COVID-19 vaccination and risk mitigation among rheumatic disease patients. Clinical characteristics and outcomes of patients with coronavirus disease 2019 (COVID-19) and rheumatic disease: a comparative cohort study from a US 'hot spot Coronavirus disease 2019 outcomes among patients with rheumatic diseases 6 months into the pandemic COVID-19 Outcomes in Patients With Systemic Autoimmune Rheumatic Diseases Compared to the General Population: A US Multicenter Impact of the COVID-19 pandemic on morbidity and mortality in patients with inflammatory joint diseases and in the general population: a nationwide Swedish cohort study Risk and prognosis of SARS-CoV-2 infection and vaccination against SARS-CoV-2 in rheumatic and musculoskeletal diseases: a systematic literature review to inform EULAR recommendations SARS-CoV-2 Infection and COVID-19 Outcomes in Rheumatic Disease: A Systematic Literature Review And Meta-Analysis Clinical and Chest Radiography Features Determine Patient Outcomes in Young and Middle-aged Adults with COVID-19 Pulmonary Disease Severity on Chest Radiographs using Convolutional Siamese Neural Networks Improvement and Multi-Population Generalizability of a Deep Learning-Based Chest Radiograph Severity Score for COVID-19 Multi-Radiologist User Study for Artificial Intelligence-Guided Grading of COVID-19 Lung Disease Severity on Chest Radiographs New COVID-19 Tools for Researchers A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation Severity of Chest Imaging is Correlated with Risk of Acute Neuroimaging Findings among Patients with COVID-19 Right Ventricular Strain Is Common in Intubated COVID-19 Patients and Does Not Reflect Severity of Respiratory Illness Timing of Invasive Mechanical Ventilation and Mortality among Patients with Severe COVID-19-associated Acute Respiratory Distress Syndrome Laboratory trends, hyperinflammation, and clinical outcomes for patients with a systemic rheumatic disease admitted to hospital for COVID-19: a retrospective, comparative cohort study. The Lancet Rheumatology DMARD, disease-modifying anti-rheumatic drug IL, interleukin; CTLA-4, cytotoxic T-lymphocyte-associated protein 4 * Includes other inflammatory arthritis (n=8), psoriatic arthritis (n=2), non-systemic juvenile idiopathic arthritis (n=1), and ankylosing spondylitis (n=1) † Includes other vasculitides (n=3), giant cell arteritis (n=2), granulomatosis with polyangiitis (n=1), and Behçet's disease Includes antiphospholipid syndrome (n=2), mixed connective tissue disease (n=2), IgG4-related disease (n=1), Sjögren's syndrome (n=1), systemic sclerosis (n=1) § Includes leflunomide (n=1), cyclophosphamide (n=1), sulfasalazine (n=1), tacrolimus (n=1) Contributorship statement: NJP, KMD, MDL, TYH, MDI, JK, JAS, and ZSW designed the study, were responsible for the acquisition, analysis, and interpretation of data, and drafted and revised the article. CC, LP, LM, KMMV and AZ were involved in data acquisition and revision of the manuscript. XF was involved in data analysis and interpretation and revision of the manuscript. YZ was involved in data interpretation and revision of the manuscript. All authors approved the final version of the article.