Sensitivity of RT-PCR testing of upper respiratory tract samples for SARS-CoV-2 in hospitalised patients: a retrospective cohort study

Background: This study aimed to determine the sensitivity and specificity of reverse transcription PCR (RT-PCR) testing of upper respiratory tract samples from hospitalised patients with coronavirus disease 2019 (COVID-19), compared to the gold standard of a clinical diagnosis. Methods: All RT-PCR testing for severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) in NHS Lothian, Scotland, United Kingdom between the 7 th of February and 19 th April 2020 (inclusive) was reviewed, and hospitalised patients were identified. All upper respiratory tract RT-PCR tests were analysed for each patient to determine the sequence of negative and positive results. For those who were tested twice or more but never received a positive result, case records were reviewed, and a clinical diagnosis of COVID-19 allocated based on clinical features, discharge diagnosis, and radiology and haematology results. For those who had a negative RT-PCR test but a clinical diagnosis of COVID-19, respiratory samples were retested using a multiplex respiratory panel, a second SARS-CoV-2 RT-PCR assay, and a human RNase P control. Results: Compared to the gold standard of a clinical diagnosis of COVID-19, the sensitivity of a single upper respiratory tract RT-PCR for COVID-19 was 82.2% (95% confidence interval 79.0-85.1%). The sensitivity of two upper respiratory tract RT-PCR tests increased sensitivity to 90.6% (CI 88.0-92.7%). A further 2.2% and 0.9% of patients who received a clinical diagnosis of COVID-19 were positive on a third and fourth test; this may be an underestimate of the value of further testing as the majority of patients 93.0% (2999/3226) only had one or two RT-PCR tests. Conclusions: The sensitivity of a single RT-PCR test of upper respiratory tract samples in hospitalised patients is 82.2%. Sensitivity increases to 90.6% when patients are tested twice. A proportion of cases with clinically defined COVID-19 never test positive on RT-PCR despite repeat testing.


Introduction
The coronavirus disease 2019 (COVID- 19) pandemic in Europe has already caused significant morbidity and mortality, not least within the United Kingdom. As well as causing large numbers of community-acquired cases, severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) has also been shown to circulate effectively within hospitals 1 , necessitating the creation of COVID-19 specific areas. An estimate of the sensitivity of reverse transcription PCR (RT-PCR) testing for SARS-CoV-2 is therefore critical. Overestimation of sensitivity by clinical staff, and a lack of use of testing results in combination with clinical features of their presentation, may lead to patients with disease being incorrectly diagnosed, and placed in non-COVID-19 areas with the subsequent risk of infection to others; underestimation of the sensitivity by clinical staff may lead to patients who are SARS-CoV-2 negative being erroneously placed in COVID-19 areas.
A recent meta-analysis 2 estimates the sensitivity of reverse transcription polymerase chain reaction (RT-PCR) testing of upper respiratory tract samples as 89%, but this metaanalysis, and a subsequent one 3 highlight a number of limitations in the literature. These include small sample size (<100 patients with COVID-19) [4][5][6][7][8][9][10][11] , reliance on RT-PCR itself as the gold standard for diagnosis 12,13 , use of computed tomography (CT) scans rather than clinical criteria as a gold standard for the diagnosis of COVID-19 14,15 , and absence of comprehensive RT-PCR testing for all included patients 16 . Finally, only a single study to our knowledge has examined the cumulative sensitivity of repeat testing for SARS-CoV-2 14 .
Here we examine in a large, comprehensive dataset the sensitivity of RT-PCR testing of upper respiratory tract specimens for COVID-19, compared to the gold standard of clinical diagnosis.

Data source
All RT-PCR testing conducted for SARS-CoV-2 in NHS Lothian between the 7 th of February and 19 th April 2020 (inclusive) was reviewed. NHS Lothian covers a population of 907,580 people 17 and during the period of the study the Royal Infirmary of Edinburgh was the only regional centre conducting SARS-CoV-2 testing. Hospitalised patients were identified by cross-matching patient identification numbers against the NHS Lothian TrakCare Patient Clinical Management System database. In this study we comply with the principles of the STARD 18 reporting guidelines for diagnostic accuracy studies.

Data collection
For the analysis of all patients (community and hospitalised), RT-PCR tests for SARS-CoV-2 conducted by the Royal Infirmary of Edinburgh virology laboratory (the only laboratory in the region conducting testing at this point) were identified.
For hospitalised patients, a record of samples sent for RT-PCR testing in the study period were identified, and only unambiguous positive or negative results, as authorised by laboratory staff, selected. Testing patterns were allocated for each patient, determining the sequence of RT-PCR tests and whether each test had yielded a negative or positive result (Table 1).

Amendments from Version 1
-A comment from both our reviewers was that as a large number of patients only received a single negative test, there was the possibility that our study over-estimates the sensitivity of RT-PCR testing for SARS-CoV-2, for hospitalised patients in this earliest stage of the pandemic. We believe that the risk of this is low, as clinical guidelines at the time recommended that an RT-PCR test should be repeated if there was clinical suspicion of COVID-19 after an initial negative test. However, this is an important limitation, which we have highlighted in our revised manuscript.
-As this is a diagnostic accuracy study we have referenced the relevant guidelines (STARD).
-We have made clear that bulk of our analysis pertains to hospitalised patients, although we also present some data for all those tested, in the community and in hospital. We have also clarified that blinding was not performed in judging whether patients were judged to have clinical evidence of COVID-19 infection, which is another limitation of our study.
-A small number of ambiguous results were not included in our analysis; we have made this clear in the Results section and given the number of these.
-We have removed the abbreviation "URT" (upper respiratory tract) in order to improve the readibility of the manuscript.
-No additional authors have been added to this paper, and no additional data or code is presented.

Case definitions
Hospitalised patients with a single negative test result were classified as a true negative, as clinical guidelines in place at the time specified that if there was clinical suspicion of COVID-19, an RT-PCR test should be repeated if the first test was negative. For those who initially tested negative on one or more occasions and then positive, case records were reviewed to determine whether this represented two discrete presentations or the same presentation. If they were classified as two distinct presentations, the negative followed by positive test was treated as a single positive test.
For those tested twice or more but who never received a positive result from RT-PCR testing, case records were reviewed, and a clinical diagnosis of COVID-19 was allocated based on a discharge diagnosis from the clinical team (or death certificate documentation) and clinical review. Clinical features, radiology reports and haematology results were reviewed. A positive clinical diagnosis was based on European Centres for Disease Control (ECDC) and World Health Organisation (WHO) criteria 19 . Based on previously published studies 20,21 , cases were judged to be more likely to represent COVID if a chest X-ray showed patchy bilateral infiltrative changes, or a CT scan showed ground glass changes and if there was lymphopaenia in the presence of a normal neutrophil count 22 . Case records were reviewed by two clinicians (EW and TCW); if a consensus decision could not be reached, the case records were reviewed by a third clinician (DW) to arrive at a final clinical diagnosis. For patients classified as a possible false negative, their initial respiratory sample was retested using a multiplex respiratory panel, a second SARS-CoV-2 RT-PCR assay on the SeeGene platform as detailed below, and a human RNase P RT-PCR.
For patients who tested positive on their initial test, case records were reviewed to ensure they met the clinical criteria for COVID-19, as described above. As for those who tested negative on two or more occasions, a positive RT-PCR case was not part of the diagnostic criteria for COVID-19 infection. If they did not meet these clinical criteria, the samples IDs were matched against samples which had undergone whole genome sequencing (WGS) as part of the COVID-19 Genomics UK sequencing consortium 23 . If WGS had been completed successfully for a sample, this was assumed to represent a true positive. For those that had not, RT-PCR re-testing was conducted using the SeeGene platform as detailed below. For patients who tested positive for a new respiratory pathogen, the case records were reviewed to ascertain whether the diagnosis was best explained by SARS-CoV-2 infection or the subsequently identified respiratory pathogen. Convalescent serology samples (>14 days after onset of symptoms), if available, were analysed using the Abbott SARS-CoV-2 IgG assay on the Abbott Architect platform 29 .

Statistical analyses
The sensitivity was calculated as the proportion of true positives detected on initial testing and re-testing of suspected false negatives, divided by the number of true positives added to convincing false negatives, as estimated on the basis of further respiratory testing and serology testing. The specificity was calculated by dividing true negatives by the number of true negatives added to those judged to be false positives, on the basis of repeat RT-PCR retesting. The positive predictive value was determined by dividing the number of true positive by the sum of the true positives and false positives. The negative predictive value was calculated by dividing the number of true negatives by the sum of the true negatives and false negatives. Confidence intervals for these estimates were calculated using a two-sided exact binomial test with a confidence level of 0.95, implemented in R 28 .

Ethics statement
As part of the study protocol, specimens and associated clinical data were collected and anonymized before additional molecular/serological testing in accordance with local ethical approval (South East Scotland Scottish Academic Health Sciences Collaboration Human Annotated BioResource reference no. 10/S1402/33). As the study formed part of a service evaluation, with no publication of patient identifiable information, the need for informed consent was waived by the local Caldicott Guardian.

Results
A total of 10,601 RT-PCR tests with unambiguous results for SARS-CoV-2 for 8311 patients were conducted on upper respiratory tract specimens by the Royal Infirmary of Edinburgh laboratories between the 7 th of February and the 19 th of April 2020 on patients for whom data was available. "In addition to these, there were 37 tests with ambiguous or indeterminate results; of these 25 were in hospitalised patients. These results were not included in our analysis." These tests included community testing for patients who were never admitted to hospital, and testing for patients outside NHS Lothian for Boards that did not perform their own SARS-CoV-2 testing. From this testing, 1667 patients received a positive result for SARS-CoV-2 testing ( Table 2). The overall sensitivity of an initial upper respiratory tract RT-PCR test for the whole cohort (using a gold standard of an eventual molecular diagnosis of SARS-CoV-2 on upper respiratory tract RT-PCR) is 91.8%, rising to 98.4% after 2 tests.

Testing for other respiratory pathogens
Of the total cohort, 3226 patients were hospitalised in NHS Lothian. The data analysis for these patients is summarised in the flowchart in Figure 1. In total, 73 patients received a clinical diagnosis of COVID-19 but did not receive a positive RT-PCR result at the time. The RNA extract used for the initial SARS-CoV-2 RT-PCR was retested for common respiratory pathogens using the Luminex assay. Out of the 73, nine (12.3%) tested positive for a respiratory pathogen. On clinical review, all nine cases were judged to be better explained by this new diagnosis rather than COVID-19 (Table 3).

Retesting with the Seegene assay
Retesting of the remaining 64 samples from suspected false negative cases with the Seegene assay for SARS-CoV-2 showed 27 (42.2%) of these were positive. Of the 37 remaining samples that neither tested positive for a respiratory pathogen nor for SARS-CoV-2 on repeat testing, all showed a positive result for human RNase P. Comparing Ct values for human RNase P for the samples that tested positive compared to those that tested negative (for SARS-CoV-2 on the Seegene assay) showed no difference using a Welch two sample t-test (p=0.49, Figure 3. The flowchart in Figure 2 summarises the results for the patients with a clinical assessment of COVID-19 but negative initial RT-PCR testing. For an initial test, the sensitivity of RT-PCR for SARS-CoV-2 infection was 82.2% (95% confidence interval 79.0-85.1%) with a specificity of 100% (CI 99.9-100%). The positive predictive value of an initial test was 100%; the negative predictive value of an initial test was 95.7% (Table 4).

Convalescent serology
Out of the cohort of 64 patients who received a clinical diagnosis of COVID-19 with initial negative testing, and negative testing for other viruses, convalescent serology (>14 days after onset of symptoms) was available for seven patients. Of these, four were positive (57.1%).

Summary of principal findings
Here we show, using a comprehensively examined dataset, that the sensitivity of RT-PCR testing of upper respiratory tract specimens for the diagnosis of COVID-19 is 82.2% on initial testing, and 90.6% after two consecutive tests. Subsequent tests showed a small increase in diagnostic yield (2.2% for three tests and a further 0.9% for four tests), although this may represent an underestimate, as a number of patients given a diagnosis of COVID-19 based on clinical criteria were only tested twice.

Findings of the present study in light of what has been published before
A previous meta-analysis gives a pooled sensitivity for RT-PCR of 89% (CI 81-94%) for the diagnosis of COVID-19 2 ; our results sit at the lower range of this estimate, but with overlapping confidence intervals. As highlighted in the introduction, the included studies suffer from a number of limitations including reliance on RT-PCR itself as the diagnostic gold standard, which would lead to an increase in the estimated sensitivity. We are not aware of any studies which have used a clinical diagnosis of COVID-19 against which to assess the sensitivity of RT-PCR. Here we show that the sensitivity of an initial test is lower than reported in this meta-analysis, but that the chance of a false negative result (17.8%) is lower than the 29% estimated in a subsequent meta-analysis 3 using a subset of studies included in 2. These widely varying estimates highlight the importance of more data to inform our understanding of the strengths and weaknesses of RT-PCR testing.

Strengths and limitations
The strengths of the study include the large dataset of both COVID-19 positive and negative patients, and extensive further testing to rule out false negative RT-PCR results and alternative diagnoses in those patients given a clinical diagnosis of COVID-19. We also studied whether inadequate sampling might be a possible explanation for false negatives. However in a cohort of 37 possible false negatives all samples had detectable RT-PCR for human RNase P, with no difference between this group and those that tested positive for SARS-CoV-2, showing that this was not a factor in determining the sensitivity of RT-PCR in this population.
A limitation of the study is that the WHO/ECDC case definition of COVID-19 is likely to be highly sensitive but have low specificity. This means that a number of the cases we identified as potential false negatives could in fact represent other case presentations (a false positive in terms of the clinical diagnosis), and thus underestimate the sensitivity of the assay. This interpretation is supported by the findings from serology, where four out of seven patients who met the clinical case criteria and had a convalescent serology sample had a positive serological test. Conversely, we did not examine the case records of the 1837 patients who tested negative on a single occasion, some of whom are likely to have received a clinical diagnosis of COVID-19, or may have had atypical COVID-19 disease. An increased number of false negatives would lead to a decreased sensitivity for the assay and therefore the sensitivity of this assay may be less than what we report.
A final limitation is that this is a retrospective, not prospective study, and that cases were not blinded to RT-PCR COVID-19 status at the point of assessment of whether they met the case criteria.

Meaning of the study and understanding possible mechanisms
The result from our study suggest that there may be a small proportion of patients with SARS-CoV-2 infection who meet the clinical case definition but never test positive on RT-PCR testing. It is possible that, in patients with severe disease, infection is entirely in the LRT, or that by time of presentation in the disease course the virus may only be present at very low levels in the upper respiratory tract 30 ; this is supported by our findings in the ICU cohort, where 6.0% of patients never tested positive on upper respiratory tract RT-PCR.    Implications for practice or policy, and suggestions for future research Reliance on RT-PCR testing may result in patients with COVID-19 being inappropriately labelled with alternative diagnoses. These possibly infectious patients will subsequently pose a risk to healthcare workers and other patients. A more detailed picture of the sensitivity of RT-PCR testing will be aided by comprehensive serological testing of hospitalised patients with suspected infection.

Data availability
Underlying data As part of a service evaluation project, this study received local Caldicott Guardian approval but no permission was granted for publication of any patient identifiable information. Therefore, the raw data underlying the analyses has not been made publicly available. Anonymised data will be provided to researchers at accredited institutions who wish to conduct their own analysis or run meta-analyses after consultation with the local Institutional Review Board. Requests for access to the data should be made to Kate Templeton (kate.templeton@ nhslothian.scot.nhs.uk).

Open Peer Review
inappropriately labelled with alternative diagnoses.") will probably still be true. Abstract, Results: "The sensitivity of two upper respiratory tract RT-PCR tests increased sensitivity to 90.6%". Please rephrase. Suggestion: "Using two upper respiratory tract RT-PCR tests increased sensitivity to 90.6%".

1.
Introduction: "Overestimation of sensitivity may lead to patients with disease being incorrectly diagnosed". I think it is not necessarily the overestimation leading to misdiagnosis, but rather the sensitivity (and actually the fact that a test results in any false negative) in itself.
the abstract) involves both sensitivity and specificity. Please be consistent in this. Also, if the aim of the study was to investigate specificity, why were potential false positives not further followed up? So maybe elucidating the specificity was not the aim of this study after all.
Limitations: the retrospective design and the lack of blinding should also be mentioned as limitations.