Abstract:Objectives
To assess the reporting and methodological quality of COVID-19 systematic reviews, and to analyze trends and gaps in the quality, clinical topics, author countries, and populations of the reviews using an evidence mapping approach.
Study Design and Setting
A structured search for systematic reviews concerning COVID-19 was performed using PubMed, Embase, Cochrane Library, Campbell Library, Web of Science, CBM, WanFang Data, CNKI, and CQVIP from inception until… Show more
“…Out of 17 available reviews published before September 1, 2020, 5 (29%) were found to be of low and the remaining 12 (71%) of critically low quality. This is also in line with Li et al [31] who assessed 63 systematic reviews (25%) to have low and 150 (62%) to have critically low quality. The authors also evaluated reporting using PRISMA [77], and the median score was 14 (10-18).…”
Section: Discussionsupporting
confidence: 89%
“…In some recent studies and clinical trials, AI has been demonstrated to match or even exceed the performance of expert radiologists, which could potentially offer expedited and less expensive diagnostics [25–30]. A recent study and meta-analysis by Li et al [31] with 31,587 identified and 82 included studies shows deep learning is capable of slightly outperforming health care professionals in detecting diseases from medical images with a pooled sensitivity of 87% (vs 86% of health care professionals) and pooled specificity of 93% (vs 91% respectively). Overlapping confidence intervals suggest that there is no statistically significant difference in performance between AI and human.…”
Objective: In this umbrella systematic review, we screen existing reviews on using artificial intelligence(AI) techniques to diagnose COVID-19 in patients of any age and sex (both hospitalised and ambulatory) using medical images and assess their methodological quality.
Methods: We searched seven databases (MEDLINE, EMBASE, Web of Science, Scopus, dblp, CochraneLibrary, IEEE Xplore) and two preprint services (arXiv, OSF Preprints) up to September 1, 2020. Eligible studies were identified as reviews or surveys where any metric of classification of detection of COVID-19 using AI was provided. Two independent reviewers did all steps of identification of records (titles and abstracts screening, full texts assessment, essential data extraction, and quality assessment). Any discrepancies were resolved by discussion. We qualitatively analyse methodological credibility ofthe reviews using AMSTAR 2 and evaluate reporting using PRISMA-DTA tools, leaving quantitative analysis for further publications.
Results: We included 22 reviews out of 725 records covering 165 primary studies. This review covers 416,254 participants in total, including 50,022 diagnosed with COVID-19. The methodological quality of all eligible studies was rated as critically low. 91% of papers had significant flaws in reporting quality. More than half of the reviews did not comment on the results of previously published reviews at all. Almost three fourth of the studies included less than 10% of available studies.
Discussion: In this umbrella review, we focus on the descriptive summary of included papers. Much wasting time and resources could be avoided if referring to previous reviews and following methodological guidelines. Due to the low credibility of evidence and flawed reporting, any recommendation about automated COVID-19 clinical diagnosis from medical images using AI at this point cannot be provided.
Funding: PO was supported by NIH grant AI116794 (the funding body had no role in the design, in any stage of the review, or in writing the manuscript); PJ and DS did not receive any funding.
Registration:The protocol of this review was registered on the OSF platform at osf.io/kxrmh
“…Out of 17 available reviews published before September 1, 2020, 5 (29%) were found to be of low and the remaining 12 (71%) of critically low quality. This is also in line with Li et al [31] who assessed 63 systematic reviews (25%) to have low and 150 (62%) to have critically low quality. The authors also evaluated reporting using PRISMA [77], and the median score was 14 (10-18).…”
Section: Discussionsupporting
confidence: 89%
“…In some recent studies and clinical trials, AI has been demonstrated to match or even exceed the performance of expert radiologists, which could potentially offer expedited and less expensive diagnostics [25–30]. A recent study and meta-analysis by Li et al [31] with 31,587 identified and 82 included studies shows deep learning is capable of slightly outperforming health care professionals in detecting diseases from medical images with a pooled sensitivity of 87% (vs 86% of health care professionals) and pooled specificity of 93% (vs 91% respectively). Overlapping confidence intervals suggest that there is no statistically significant difference in performance between AI and human.…”
Objective: In this umbrella systematic review, we screen existing reviews on using artificial intelligence(AI) techniques to diagnose COVID-19 in patients of any age and sex (both hospitalised and ambulatory) using medical images and assess their methodological quality.
Methods: We searched seven databases (MEDLINE, EMBASE, Web of Science, Scopus, dblp, CochraneLibrary, IEEE Xplore) and two preprint services (arXiv, OSF Preprints) up to September 1, 2020. Eligible studies were identified as reviews or surveys where any metric of classification of detection of COVID-19 using AI was provided. Two independent reviewers did all steps of identification of records (titles and abstracts screening, full texts assessment, essential data extraction, and quality assessment). Any discrepancies were resolved by discussion. We qualitatively analyse methodological credibility ofthe reviews using AMSTAR 2 and evaluate reporting using PRISMA-DTA tools, leaving quantitative analysis for further publications.
Results: We included 22 reviews out of 725 records covering 165 primary studies. This review covers 416,254 participants in total, including 50,022 diagnosed with COVID-19. The methodological quality of all eligible studies was rated as critically low. 91% of papers had significant flaws in reporting quality. More than half of the reviews did not comment on the results of previously published reviews at all. Almost three fourth of the studies included less than 10% of available studies.
Discussion: In this umbrella review, we focus on the descriptive summary of included papers. Much wasting time and resources could be avoided if referring to previous reviews and following methodological guidelines. Due to the low credibility of evidence and flawed reporting, any recommendation about automated COVID-19 clinical diagnosis from medical images using AI at this point cannot be provided.
Funding: PO was supported by NIH grant AI116794 (the funding body had no role in the design, in any stage of the review, or in writing the manuscript); PJ and DS did not receive any funding.
Registration:The protocol of this review was registered on the OSF platform at osf.io/kxrmh
“…The quality of the published work was not assessed in our analysis, given the broad scope and huge diversity of the included papers. Nevertheless, many surveys of the quality of COVID-19 publications already exist [15,[25][26][27][28][29][30][31][32][33][34][35][36][37]. Although existing surveys of the quality of COVID-19 research do not cover all subfields of investigation and quality is often difficult to measure precisely, the consistent finding of the high prevalence of low-quality studies across very different types of study designs suggests that a large portion ( perhaps even the large majority) of the immense and rapidly growing COVID-19 literature may be of low quality.…”
We examined the extent to which the scientific workforce in different fields was engaged in publishing COVID-19-related papers. According to Scopus (data cut, 1 August 2021), 210 183 COVID-19-related publications included 720 801 unique authors, of which 360 005 authors had published at least five full papers in their career and 23 520 authors were at the top 2% of their scientific subfield based on a career-long composite citation indicator. The growth of COVID-19 authors was far more rapid and massive compared with cohorts of authors historically publishing on H1N1, Zika, Ebola, HIV/AIDS and tuberculosis. All 174 scientific subfields had some specialists who had published on COVID-19. In 109 of the 174 subfields of science, at least one in 10 active, influential (top 2% composite citation indicator) authors in the subfield had authored something on COVID-19. Fifty-three hyper-prolific authors had already at least 60 (and up to 227) COVID-19 publications each. Among the 300 authors with the highest composite citation indicator for their COVID-19 publications, most common countries were USA (
n
= 67), China (
n
= 52), UK (
n
= 32) and Italy (
n
= 18). The rapid and massive involvement of the scientific workforce in COVID-19-related work is unprecedented and creates opportunities and challenges. There is evidence for hyper-prolific productivity.
“…Nevertheless, many surveys of the quality of COVID-19 publications already exist. [25][26][27][28][29][30][31][32][33][34][35][36][37][38] Although existing surveys of the quality of COVID-19 research do not cover all subfields of investigation and quality is often difficult to measure precisely, the consistent finding of high prevalence of low quality studies across very different types of study designs suggests that a large portion (perhaps even the large majority) of the immense and rapidly growing COVID-19 literature may be of low quality. Moreover, massive productivity has been described in the pre-COVID era, as affecting researchers across many fields 39 and may be a particular feature for COVID-19 research.…”
Importance: COVID-19 is a major global crisis and the scientific community has been mobilized to deal with this crisis.
Objective: To estimate the extent to which the scientific workforce in different fields has been engaged publishing papers relative to the COVID-19 pandemic.
Design, setting, and participants: We evaluated Scopus (data cut, December 1, 2020) for all indexed published papers and preprints relevant to COVID-19. We mapped this COVID-19 literature in terms of its authors across 174 subfields of science according to the Science Metrix classification. We also evaluated the extent to which the most influential scientists across science (based on a composite citation indicator) had published COVID-19-related research. Finally, we assessed the features of authors who published the highest number of COVID-19 publications and of those with the highest impact in the COVID-19 field based on the composite citation indicator limited to COVID-19 publications.
Main outcomes and measures: Publishing scientists (authors) and their published papers and citation impact.
Results: 84,180 indexed publications were relevant to COVID-19 including 322,279 unique authors. The highest rates of COVID-19 publications were seen for authors classified in Public Health and in Clinical Medicine, where 11.3% (6,388/56,516) and 11.1% (92,570/833,060) of authors, respectively, had published on COVID-19. Almost all (173/174) subfields (except for Automobile Design & Engineering) had some authors publishing on COVID-19. Among active scientists at the top 2% of citation impact, 15,803 (13.3%) had published on COVID-19 in their publications in the first 11 months of 2020. The rates were the highest in the fields of Clinical Medicine (27.7%) and Public Health (26.8%). In 83 of the 174 subfields of science, at least one in ten active, influential authors in that field had authored something on COVID-19. 65 authors had already at least 30 (and up to 133) COVID-19 publications each. Among the 300 authors with the highest composite citation indicator for COVID-19 publications, 26 were journalists or editors publishing news stories or editorials in prestigious journals; most common countries for the remaining were China (n=77), USA (n=66), UK (n=27), and Italy (n=20).
Conclusions and relevance: The scientific literature and publishing scientists have been rapidly and massively infected by COVID-19 creating opportunities and challenges. There is evidence for hyper-prolific productivity.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.