COVID-MATCH65—A prospectively derived clinical decision rule for severe acute respiratory syndrome coronavirus 2

Objectives We report on the key clinical predictors of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection and present a clinical decision rule that can risk stratify patients for COVID-19. Design, participants and setting A prospective cohort of patients assessed for COVID-19 at a screening clinic in Melbourne, Australia. The primary outcome was a positive COVID-19 test from nasopharyngeal swab. A backwards stepwise logistic regression was used to derive a model of clinical variables predictive of a positive COVID-19 test. Internal validation of the final model was performed using bootstrapped samples and the model scoring derived from the coefficients, with modelling performed for increasing prevalence. Results Of 4226 patients with suspected COVID-19 who were assessed, 2976 patients underwent SARS-CoV-2 testing (n = 108 SARS-CoV-2 positive) and were used to determine factors associated with a positive COVID-19 test. The 7 features associated with a positive COVID-19 test on multivariable analysis were: COVID-19 patient exposure or international travel, Myalgia/malaise, Anosmia or ageusia, Temperature, Coryza/sore throat, Hypoxia–oxygen saturation < 97%, 65 years or older—summarized in the mnemonic COVID-MATCH65. Internal validation showed an AUC of 0.836. A cut-off of ≥ 1.5 points was associated with a 92.6% sensitivity and 99.5% negative predictive value (NPV) for COVID-19. Conclusions From the largest prospective outpatient cohort of suspected COVID-19 we define the clinical factors predictive of a positive SARS-CoV-2 test. The subsequent clinical decision rule, COVID-MATCH65, has a high sensitivity and NPV for SARS-CoV-2 and can be employed in the pandemic, adjusted for disease prevalence, to aid COVID-19 risk-assessment and vital testing resource allocation.


Background
The COVID-19 pandemic caused by severe acute respiratory syndrome coronavirus 2 (SARS-  was first reported in China and has now infected over 9 million people globally (1). A range of clinical symptoms and syndromes have been reported in confirmed . However, there have been limited prospective reports of the clinical and epidemiological predictors of . We report on the clinical and epidemiological predictors of COVID-19 from a uniquely derived prospective database and present a point-of-care COVID-19 clinical decision tool.

Methods
A COVID-19 rapid assessment screening clinic was established at Austin Health on 11 March 2020 with prospective electronic medical record (EMR; eMethods) data of patients presenting to the clinic systematically collected by medical staff from 11 March to 22 April 2020. Patients were predominantly adults -children over 6 months were seen at clinician discretion. Modifications to the EMR were made during the study period to align with the Victorian Department of Health and Human Services (DHHS) testing criteria (6) (eMethods). Only those patients that met the DHHS criteria for SARS-CoV-2 testing had nasopharyngeal swab collected for SARS-Cov-2 nucleic acid detection by polymerase chain reaction (PCR). Patients with swabs that had SARS-CoV-2 nucleic acid detected were termed "COVID-19 test positive"; those with swabs where SARS-CoV-2 nucleic acid was not detected were termed "COVID-19 test negative". This study was approved by the Austin Health Human Research and Ethics Committee.

Derivation and Internal Validation Cohort
Clinical data from the data collection tool (baseline demographics, clinical symptoms, clinical observations) and COVID-19 testing results were extracted from Austin Health EMR platform (Cerner®) by the Data Analytics Research and Evaluation (DARE) Centre (Austin Health/University of Melbourne).

Statistical analysis
All rights reserved. No reuse allowed without permission.
was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020. . https: //doi.org/10.1101//doi.org/10. /2020 All results are presented according to TRIPOD guidelines(7). Categorical variables are presented as frequency (percentage) and continuous variables as median (interquartile range [IQR]). Fisher's exact test or rank sum test were used to compare characteristics between tested and not tested patients. To determine the predictors of a positive COVID-19 test, a multivariable logistic regression with backward stepwise procedure was used, eliminating variables with p>0.10 and re-inclusion of variables with p<0.05. Bootstrapping was used for internal validation. Further details on variable selection, model development and performance, internal validation and score derivation are outlined in eMethods.

Study population and setting
During the study period 4359 assessments were performed in 4226 patients (eTable 1). For those with multiple presentations (n=118) only their first testing date was used (for patients that were not tested, their first assessment was taken). Median (IQR) number of daily assessments was 96 (71, 134) with an average of 51% of patients being tested each day (eFigure1).

COVID-19 testing
Testing was performed on 2976 patients (70%). The characteristics of those with suspected COVID-19, stratified by testing performed status, is outlined in eTable 2. The most frequently reported symptoms in both groups were any fever (reported or documented), cough, sore throat and coryza as outlined in eTable 2.

COVID-19 test positivity
Of the 2976 patients that were tested, 41 were excluded from the analysis due to pending results (n=38) or indeterminate results (n=3). The prevalence of a positive COVID-19 test in the final cohort was 3.7% (108/2935). Characteristics of those patients with a positive COVID-19 test are shown in Table 1.
All rights reserved. No reuse allowed without permission.
was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020. . https://doi.org/10. 1101/2020 Demographic, epidemiological and clinical factors associated with a positive COVID-19 test The characteristics associated with a positive COVID-19 test in univariate and multivariable analysis are shown in Table 2. The seven features associated with a COVID-19 test on multivariable analysis were summarized in the mnemonic COVID-MATCH65 (Figure 1). The model showed good discrimination (AUC = 0.843, Hosmer-Lemeshow chi 2 =4.96, p=0.762) and calibration (calibration slope = 1.00, Brier score = 0.03, product-moment correlation between observed and predicted probability = 0.35). Internal validation showed minimal mean optimism of 0.007 with internally validated AUC of 0.836 (eFigure 2 & 3). The resulting score ranges from -1 to 6.5 points with score ≤ 1 representing low risk of a positive test (<1%) and scores above 4 having beyond 20% probability of a positive test (Figure 1).

Discussion
Whilst the clinical features of COVID-19 have been well reported, robust prospective from patients presenting for COVID-19 assessment that are both SARS-CoV-2 positive and negative on testing remains absence. Therefore, to date an accurate assessment of the clinical predictors associated with a All rights reserved. No reuse allowed without permission.
was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020. . https://doi.org/10.1101/2020.06.30.20143818 doi: medRxiv preprint positive SARS-CoV-2 test has been ill defined. Whilst fever has been the predominant presenting feature of confirmed COVID-19 cases from published inpatient populations(4), it was in fact observed less frequently (36.5%) in our outpatient cohort, potentially the result of earlier presentation (5 days[median] from symptom onset). Bajema et al.(5) reported fever in 68% in a retrospective cohort study (n=210) from the USA with similar incidence rate of COVID-19 positive tests to our cohort (5% USA vs. 4.7% AUS). Whilst in the earliest reports from confirmed cases in China the figures were 83-98%(2, 3). Whilst coryza and sore throat were frequently reported, the presence of either was in fact a negative predictor of COVID-19 infection. Anosmia or ageusia as seen in other emerging studies was a strong predictor of a positive COVID-19 test (8). Whilst contact and/or international travel was a predictor of COVID-19 infection in our model, as seen in US model from Challenger et al. (9), it may be less relevant in outbreak settings and during periods of travel bans, however these criteria alone are not required for a patient to be at high risk of COVID-19.
Our model has some limitations, including the single centre prospective data source, jurisdictional was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020.

Acknowledgements -Nil
All rights reserved. No reuse allowed without permission.
was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020.  was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020. . https: //doi.org/10.1101//doi.org/10. /2020  was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.
was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint (which this version posted July 2, 2020. . https://doi.org/10.1101/2020.06.30.20143818 doi: medRxiv preprint  was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.