The Diagnostic Performance of Various Clinical Specimens for the Detection of COVID-19: A Meta-Analysis of RT-PCR Studies

Background: The diagnostic performance of numerous clinical specimens to diagnose COVID-19 through RT-PCR techniques is very important, and the test result outcome is still unclear. This review aimed to analyze the diagnostic performance of clinical samples for COVID-19 detection by RT-PCR through a systematic literature review process. Methodology: A compressive literature search was performed in PubMed/Medline, Scopus, Embase, and Cochrane Library from inception to November 2022. A snowball search on Google, Google Scholar, Research Gate, and MedRxiv, as well as bibliographic research, was performed to identify any other relevant articles. Observational studies that assessed the clinical usefulness of the RT-PCR technique in different human samples for the detection or screening of COVID-19 among patients or patient samples were considered for this review. The primary outcomes considered were sensitivity and specificity, while parameters such as positive predictive value (PPV), negative predictive value (NPV), and kappa coefficient were considered secondary outcomes. Results: A total of 85 studies out of 10,213 non-duplicate records were included for the systematic review, of which 69 articles were considered for the meta-analysis. The meta-analysis indicated better pooled sensitivity with the nasopharyngeal swab (NPS) than saliva (91.06% vs. 76.70%) and was comparable with the combined NPS/oropharyngeal swab (OPS; 92%). Nevertheless, specificity was observed to be better with saliva (98.27%) than the combined NPS/OPS (98.08%) and NPS (95.57%). The other parameters were comparable among different samples. The respiratory samples and throat samples showed a promising result relative to other specimens. The sensitivity and specificity of samples such as nasopharyngeal swabs, saliva, combined nasopharyngeal/oropharyngeal, respiratory, sputum, broncho aspirate, throat swab, gargle, serum, and the mixed sample were found to be 91.06%, 76.70%, 92.00%, 99.44%, 86%, 96%, 94.4%, 95.3%, 73.63%, and above 98; and 95.57%, 98.27%, 98.08%, 100%, 37%, 100%, 100%, 97.6%, and above 97, respectively. Conclusions: NPS was observed to have relatively better sensitivity, but not specificity when compared with other clinical specimens. Head-to-head comparisons between the different samples and the time of sample collection are warranted to strengthen this evidence.


Introduction
The recent global pandemic was caused by a respiratory tract infection in the Wuhan province of China in December 2019. The causative organism was recognized as a severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The coronavirus disease 2019  has spread across the world and contributed to many deaths in a huge proportion of the population. Fast and accurate detection of viruses and/or diseases is essential to

Criteria for Considering the Studies for This Review
The observational studies assessed the clinical usefulness of the RT-PCR technique in various human samples for the detection or screening of COVID-19 among the patients or patient samples that were considered for this review. Only the studies with full-text availability in the English language were considered. Studies comparing the numerous samples were also considered for this review. The primary outcomes considered were sensitivity and specificity, while parameters such as positive predictive value (PPV), negative predictive value (NPV), and kappa coefficient were considered secondary outcomes. We have considered all types of RT-PCR techniques in our review, per the author's discretion. Any studies that used RT-PCR as a reference to assess the performance of other screening techniques were excluded. Studies such as reviews, descriptive studies, non-clinical studies, non-COVID-19 participants, commentary, guidelines, and qualitative analyses were excluded.

Search Methods for Identification of Studies
PubMed/Medline, Scopus, Embase, and Cochrane Library) were accessed through a comprehensive search strategy using all the possible keywords and entry terms from inception to November 2022. We also performed a snowball search on Google, Google Scholar Research Gate, and MedRxiv to identify any relevant articles. The reference lists of potential articles were also screened to identify additional potentially relevant citations. A detailed search strategy in various databases is provided in Supplementary File S1.

Study Selection
All the identified records through a database literature search were retrieved in an Excel sheet and screened against the pre-defined criteria. The studies were screened by reading the title and abstracts in the initial stage, followed by the full text. Only the studies passing these two stages were considered for final inclusion in the review. Two independent reviewers were involved in the study selection to limit bias, and discrepancies were resolved through consensus or discussion with another member of the research team.

Data Extraction
The data were extracted to a well-defined data extraction form by two independent reviewers. The author's first name and year of publication were used to identify the studies. The study detailed information such as year, country, study design, and study settings; the participants' information including the total number of samples/participants, age and gender of cohort, and clinical presentation or characteristics; type of specimen; and the characteristics of RT-PCR techniques were captured from the studies. The outcomes were collected from the studies or calculated from the available data in terms of percentage with a 95% confidence interval. The highest values of primary and secondary outcomes were captured in the case of multiple RT-PCR kits used in the same study. Two independent reviewers were involved in the data extraction, and disagreements were resolved through discussion or consultation with another reviewer.

Evidence Synthesis and Meta-Analysis
All the evidence extracted through the systematic process was summarized narratively and presented in tabular form. The studies that have sufficient homogenous data or if there is a sufficient number of studies to perform meta-analysis were only considered for meta-analysis. Review Manager 5.4 was used to conduct the meta-analysis [14]. The available data were converted into percentage and standard error and presented as pooled outcomes with a 95% confidence interval. We used the random effect model, as there was substantial heterogeneity (I2 > 50%; p < 0.10) in all analyses.

Publication Bias and Sensitivity Analysis
The visual inspection of the funnel plot for the sensitivity of RT-PCR in COVID-19 diagnosis was used to check publication bias using RevMan 5.4, which was further assessed for statistical significance with Egger's and Begg's test using comprehensive meta-analysis (trial version). A probability of less than 0.05 was considered to be statistically significant [15,16]. The sensitivity analysis was performed to check the robustness of the findings by removing the study with the lowest weight in the analysis [17].

Study Selection Process
A total of 32,006 records were identified from literature sources and 10,213 records were screened by title and abstract following duplicate removal. A total of 8265 records (animal studies and case reports: 403; non-diagnostic and treatments: 303; guidelines and protocols: 42; non-English: 72; not RT-PCR: 6925; pediatric: 315; qualitative research and reviews: 205) were excluded at this stage, and the remaining 1948 full texts were considered for their eligibility. Following the exclusion of 1863 articles with numerous reasons (animal studies and case reports: 78; duplicate: 1; not outcome of interest: 1248; non-English: 43; not RT-PCR: 398; review: 95), 85 studies  were considered for this systematic review. Hence, a total of 69 articles with homogenous data were used for the meta-analysis. A detailed description of the study selection process is depicted in Figure 1.

Study Characteristics
The studies were published between the years 2020 and 2022 from different parts of the world with a major contribution from the USA, the UK, and India. The studies were observational making them retrospective, prospective, and cross-sectional in nature. The studies were from hospital settings or sample collection centers. The human samples were analyzed across the included studies. The majority of the studies included adult participants with an average age of 18 to 65 years. The participants were asymptomatic or symptomatic, severe or non-severe, and positive or negative at the time of sample collection. A detailed description of the studies and participant characteristics are provided in Table 1.

Characteristics of RT-PCR Techniques
Many human samples such as the nasopharyngeal swab, oropharyngeal swab, respiratory tract specimens (bronchoalveolar lavage and broncho aspirates), throat, nasal, saliva, sputum, fecal, gargle, or mixed were used for the detection of COVID-19 using various RT-PCR techniques. The samples were stored at a cool temperature ranging from 40 • C to −800 • C. Many in-house and modified RT-PCR techniques were used by the studies by numerous companies. Detailed information on the RT-PCR techniques is depicted in Supplementary File S2.

Kappa Coefficient
A meta-analysis of 5 studies indicated a pooled kappa coefficient of 0.61 (95%CI: 0.44 to 0.79; I2: 98%) in the saliva samples using different RT-PCR techniques. (Figure 3D). The diagnostic parameters of RT-PCR in the saliva samples are provided in Figure 3.

Kappa Coefficient
A meta-analysis of five studies indicated a pooled kappa coefficient of 0.82 (95%CI: 0.67 to 0.98; I2: 98%) in the combined nasopharyngeal/oropharyngeal samples using various RT-PCR techniques ( Figure 4D). The diagnostic parameters of RT-PCR in the nasopharyngeal and oropharyngeal samples are provided in Figure 4. The meta-analysis findings on various samples are provided in Table 2.

Respiratory Samples
Only two studies [26,69] reported the sensitivity and specificity of RT-PCR in respiratory samples. Wu et al., [26] indicated a sensitivity and specificity of 100% and Nakura Y et al. [69] reported a sensitivity and specificity of 99.44% and 100%, respectively. The studies by Wu S et al. [26] and Pekoz A et al. [41] recorded a PPV of 100% and 73.7; and an NPV of 100% and 100%, respectively. Additionally, Price T K et al. [83] recorded an NPV of 98% among their samples. Wu et al. [26] recorded a kappa coefficient of 1 among 52 samples analyzed. The details are provided in Table 3.

Sputum Samples
The study by Torres A et al. [33] indicated a sensitivity, specificity, PPV, NPV, and kappa coefficient of 86%, 37%, 38%, 85%, and 0.73, respectively. The study by Villota S D et al. [22] reported a sensitivity and specificity of 86% and 37%, respectively. The details are provided in Table 3.

Broncho Aspirate Samples
Only a single study by Pace V D et al. [23] used broncho aspirate samples for the detection of COVID-19 using RT-PCR. The sensitivity, specificity, and kappa coefficient were 96%, 100%, and 0.94, respectively. The details are provided in Table 3.

Gargle Samples
The study by Dumaresq J et al. [43] reported a sensitivity and kappa coefficient of 95.3% and 0.94, respectively, in gargle samples for the detection of COVID-19 using the RT-PCR technique. The details are provided in Table 3.

Publication Bias
A visual inspection of the funnel plot reveals an obvious asymmetry, which represents the chances of publication bias. This was confirmed statistically by Egger's test (p = 0.00003) but not Begg's test (p = 0.0982). The funnel plot is provided in Supplementary File S3.

Sensitivity Analysis
The sensitivity analysis was performed by altering the analysis model from the random effect model to the fixed effect model on NPS sensitivity analysis (Figure 2A). This made a small change in the overall effect measure, which is 91.06% (95%CI: 88.91 to 93.21) in the random effect model and 94.53% (95%CI: 94.53 to 94.54) in the fixed effect model. The sensitivity analysis result is provided in Supplementary File S4.

Discussion
COVID-19 can manifest in a variety of forms ranging from simple flu-like illness to death [103]. Various samples are used for the diagnosis of COVID-19 using many techniques, including RT-PCR. The diagnostic performance of various sampling approaches needs to be investigated to gain a better picture of all these aspects [104].
Our review provided evidence that pharyngeal samples (combined nasopharyngeal/oropharyngeal) have an equivalent sensitivity to nasopharyngeal samples, whereas saliva samples have a lesser sensitivity compared to the two other types of samples. A previous systematic review reported a comparable diagnostic performance with pooled nasal and throat swabs in comparison with nasopharyngeal swabs, which is considered to be the gold standard technique. Moreover, the self-collection of samples has influenced diagnostic accuracy [104].
As indicated in our review, respiratory samples, combined nasopharyngeal/oropharyngeal samples, broncho aspirate samples, throat swab samples, gargle samples, and mixed samples had better sensitivity than other samples, like serum and saliva, compared to nasopharyngeal swabs. Similarly, the study by Becker et al. recorded that saliva had approximately 30% lesser sensitivity than NPS, and it was 50% less sensitive in those cohorts of samples taken less than 21 days from the first symptom occurrence [105]. Similar findings were observed in the previous meta-analysis by Lee et al. [106].
Combined NPS/OPS, saliva, respiratory, broncho aspirate, throat, and mixed samples had better specificity than NPS. The current review indicates lower specificity with NPS than other specimens except for the sputum sample, which had reduced specificity compared to NPS. A community study by Torres et al. reported that saliva had 99.1% relative specificity to NPS [107]. Better diagnostic accuracy and specificity with saliva samples have been reported by many other studies [108][109][110]. Moreover, another study by Sasikala et al. reported that there was no difference between the diagnostic performance of saliva samples collected by healthcare workers and the patients themselves [111].
This review suggests that throat and respiratory samples had a similar positive predictive value (PPV) compared to NPS, while all other specimens had a lower PPV than NPS. Wang H et al. also found that NPS had better performance than other samples and recommended it as the best specimen for detecting COVID-19 through RT-PCR techniques [112]. The findings from this study can be used to develop protocols and guidelines for diagnosing COVID-19 and similar infections. Although NPS is considered the gold standard for diagnosing COVID-19, other samples have also been found to be equally helpful. Head-tohead analysis between different sampling strategies and specimens needs to be studied to develop the best alternative, cost-effective, and accurate diagnostic techniques.
This review had some limitations. First, there was a significant level of heterogeneity in all the meta-analyses performed, so caution should be taken when interpreting the findings. Second, English language restriction might have contributed to the exclusion of studies. However, comprehensive literature searches in all the available databases helped to collate the maximum possible information. Third, the variation in the RT-PCR techniques used and their processes might have contributed to the findings. Fourth, there was a lack of information with respect to sampling techniques and time of sampling. Hence, further research studies should focus on this. Future meta-analyses that emphasize subgroup analysis based on COVID-19 status, severity, and other important parameters should be planned.

Conclusions
The current meta-analysis suggests that NPS has a better or similar sensitivity than other samples, especially the specimens collected from any parts of the respiratory system, while the relative specificity of NPS was lower compared to other samples. Caution should be taken while interpreting the results due to the high heterogeneity in the analyses.