Characterising long term Covid-19: a living systematic review

Protocol registration

This report was structured according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement guidelines.[15] The protocol was registered with PROSPERO (CRD42020211131) and published in a peer-reviewed journal.[16]

Search strategy

The following databases were searched: Medline and CINAHL (EBSCO), Global Health (Ovid), WHO Global Research Database on Covid-19, and LitCOVID from 1st January 2020 to 17th March 2021. Additionally, we searched Google Scholar on 17th March 2021, screening the first 500 titles. A ‘backwards’ snowball search was conducted of the references of systematic reviews. Full search terms are included in supplement 1. The search terms and inclusion criteria have, for this first version, been designed to cast a wide net, and will be modified in line with new evidence, research priorities, and clinical and policy needs.

Eligibility criteria

Peer-reviewed studies were considered eligible if they included at least 100 people with suspected, laboratory confirmed, and/or clinically diagnosed Covid-19. Without a clear or internationally agreed case definition, we included studies that reported symptoms or outcomes assessed at 12 or more weeks post Covid-19 onset.[5] There were no language restrictions. Reviews and opinion pieces were excluded. Studies were excluded if they included fewer than 100 participants or the follow-up was unclear or less than 12 weeks post-onset.

Screening

Screening was performed independently by two reviewers. Any disagreements were resolved via consensus or a third reviewer. Non-English articles were translated using Google Translate or reviewed by a reviewer with good knowledge of the language. The data were managed using the review software Rayyan.[17]

Data extraction

Data extraction was performed using Microsoft Excel. A data extraction template informed by a previous review [18] was reviewed, updated, and piloted before being finalised. Data extracted included study design, population characteristics, outcomes, prevalence, duration of symptoms, and risk factors. Data extraction was performed by one reviewer and checked by a second reviewer. Disagreements were resolved through consensus. To avoid duplication of data in future updates and ensure robustness, data extraction was not performed for non-peer-reviewed preprints.

Risk of bias assessment

The included studies were assessed for risk of bias using a modified version of the tool produced by Hoy et al. [19] (Supplement 2). This assessment checklist is a validated tool for assessing risk of bias in prevalence studies. The checklist has ten domains for assessing risk of bias, used to calculate a cumulative overall risk of bias for the whole study.

Data analysis

We undertook individual descriptive analysis for each study. We presented symptom proportions by different settings, as presented in the individual studies: hospitalised, non-hospitalised, or a mix of both populations if no subset data was available. Symptoms were broadly grouped into physiological clusters through discussion with clinicians. Proportion of symptoms and its 95% confidence intervals (CIs) were estimated using the exact method.[20] If there were two or more studies in each symptom, a meta-analysis was performed using a random intercept logistic regression model with Hartung-Knapp modification due to the heterogeneity and skewed sample sizes.[21,22] Heterogeneity between estimates was assessed using the I² statistic.[23] Additional subgroup analysis was conducted to explore the modification of the following factors on proportion of symptoms: hospitalisation, settings, continents, and follow-up timing. We also conducted meta-regression analysis on the percentage of females and ICU patients where there were more than 10 studies in the symptom. Sensitivity analyses were conducted to examine the impact of high risk of bias studies and statistical methods, Freeman-Tukey Double arcsine transformation using inverse variance meta-analysis, on the estimates. Funnel plots were plotted using proportion of the symptom against the precision and sample sizes [22] where there were more than 10 studies in the symptom to explore risk of publication bias. All analysis and data presentation were performed using metaprop and ggplot2 in R (version 4.0.5) via RStudio (version 1.3.1093). The data is presented using a combination of infographics, prepared by a design company, Design Science, and scientific tables to facilitate interpretation by different stakeholders, including non-specialists.

Patient and public involvement

The study team includes members who have been affected by long term Covid-19 sequalae (CH, MO, and JCS). CH, MO, and JCS are members of Long Covid Support, a patient support group with global reach, with approximately 40,000 members.

They actively contributed to the development of the study protocol, to inform the research questions and interpretation and presentation of the findings, to communicate the results to different audiences. The results of this LSR will be disseminated to Long Covid patient forums for discussion and feedback to inform research priorities and updates.

Characteristics of included studies

Most studies were set in Europe (62%, 24/39), followed by Asia (23%, 9/39), North America (8%, 3/39), and the Middle East 8% (3/39). There was no study set in a low-middle income (LMIC) country (Figure 1).[24] Most were cohort studies (82%, 32/39), followed by cross sectional studies (15%, 6/39), and a case control study (3%, 1/39).

• Download figure
• Open in new tab

These studies present data on 10,951 (range: 100-1,733) people in 12 countries, aged from 9 months to 93 years old, 48% (5206/10931) were female. Most studies included adults, while 10% (4/39) also included children.[25–28] Only 15% (6/39) of studies reported ethnicity of the participants [29–34], but without stratification. Table 1 presents the included study characteristics.

Most studies (67%, 26/39) were cohorts of hospitalised patients post-discharge, 10% (4/39) were set in the community (non-hospitalised population) whilst 23% (9/39) included both, with 78% (8520/10951) of individuals hospitalised during acute Covid-19. Twenty-two studies included people requiring ICU admission during the acute phase.[25,27–29,31,32,34–49]

The longest follow-up period in any study was a mean of 221.7 (SD: 10.9) days post-onset. Only 56% (22/39) of studies specified Covid-19 severity [25,27–29,31,32,34–49], 31% (12/39) treatment received during the acute phase [30,34,35,39,40,44,47,50–54], and 62% (24/39) described ventilation support requirements.[30–36,39,40,42–45,47,48,50,51,54–60] Pre-existing comorbidities were reported in the majority of studies (85%, 33/39), with hypertension and diabetes most commonly documented.[27,29–51,53–57,59,61–63]

Risk of bias

Twelve studies were assessed as high risk of bias, 22 moderate, and five low risk of bias. Most studies had a high risk of bias with regards to the generalisability of their results to the wider population with Covid-19. High risk-of-bias ratings were most common for external validity, with item 1 (representation of target population) and item 3 (random selection) having the most high risk of bias ratings (Supplement 2). Further, the recruitment process and response rates were often not well-described, and several studies applied different data collection methods. Although many studies applied validated measurement methods to assess participants, most were not designed to detect symptoms arising from Covid-19. Only four studies included a comparative control group.[29,30,37,38]

Symptoms & signs

People suffering from Long Covid report a wide range of new or persistent symptoms, in both the hospitalised and non-hospitalised populations. Symptoms were broadly organised into physiological ‘clusters’ for the purpose of presentation and interpretation of this review (Figure 2).

• Download figure
• Open in new tab

The focus of each study included in our analysis varied. Some authors focused solely on a specialty, such as dentistry, or a specific symptom, such as cognition, making comparative analysis difficult. Even amongst those studies which took a broad approach, the prevalence of symptoms was diverse. Similarly, the prevalence of the more commonly reported symptoms varied markedly.

Within these limitations, we performed a meta-analysis of the most commonly reported symptoms and signs of Long Covid. The most commonly described symptoms (with prevalence of 25% or greater) were weakness (41%, 95% CI 25.43 to 59.01), general malaise (33%, 95% CI 14.91 to 57.36), fatigue (31%, 95% CI 23.91 to 39.03), concentration impairment (26%, 95% CI 20.96 to 31.73) and breathlessness (25%, 95% CI 17.86 to 33.97). Across studies, 37% (95% CI 18.43 to 59.93) of people reported reduced quality of life. Although high I² values (>80%) were observed, they resulted from narrow dispersions in the estimates and well-separated estimates and confidence intervals between studies (Supplement 4). The differences between these symptoms, and the heterogeneity within them is likely to be, to some extent, due to other factors (e.g., study settings, populations, and different measurement tools used).

A diverse array of less prevalent symptoms, including systemic, musculoskeletal, neurological, and psychological symptoms such as sweating, chest pain, sore throat, anxiety, and headaches, amongst others, were also reported. The prevalence of these symptoms was lower, usually less than 20%. Figure 3 presents the range of symptoms.

• Download figure
• Open in new tab

Figure 4 displays this data by population. We also performed subgroup analysis based upon setting (hospitalised vs never hospitalised) and follow-up time. In several symptoms and signs, the heterogeneity of the results was found to be associated with level of hospitalisation, hospital settings, location of the studies, and follow-up timing using subgroup analysis (Supplement 5-8). Using meta-regression, the proportion of females in the studies was positively associated with headache and smell and taste disturbance (Supplement 9), while the proportion of ICU patients in the studies was positively associated with muscle pain (Supplement 10). No major difference was found in the sensitivity analyses (Supplement 11-12). Asymmetries found in the funnel plots suggest reporting biases and limited methodological quality in the included studies (Supplement 13).

• Download figure
• Open in new tab

Imaging & diagnostics

Multiple studies assessed lung sequelae and respiratory performance through outpatient visits follow-up (49%, 19/39).[25,31–37,40,42,43,45–48,50,54,55,60] Imaging results were reported in 33% (13/39) [25,31–33,37,40,42,46–48,50,55,60] of the cohort studies, with one including controls,[37] and one including children.[25] Authors used heterogenous measurement techniques, including artificial intelligence and point-of-care ultrasound.[37,48] Studies found abnormal CT results, including consolidation, reticulation, residual ground glass opacity, interstitial thickening, and fibrotic changes. Some of these studies presented comparisons between initial CT findings and those at follow-up, showing improvements overtime in pulmonary clinical measures and radiologic resolutions at serial visits. [31,33,40,42,48]

Pulmonary function tests were reported in 26% (10/39) of studies [31,32,35–37,42,43,45,47,50,55], including spirometry, diffusion capacity, lung volume, and exercise tests. These studies found evidence of altered pulmonary function, most frequently significant reduction of diffusing lung capacity for carbon monoxide (DLCO). Across these, the study with the longest follow up, at an average of 186 (175–199) days post-hospital discharge, following up adults (median age 57 (IQR 47– 65) years; 52% men) found impaired 6-min walk test in 24–56%, with higher impairment correlating with increased acute phase severity. Using high resolution CT and ultrasonography they detected corresponding lung diffusion impairment in 22% (severity scale 3), 29% (severity scale 4), and 56% (scale 5–6). Median CT scores were 3 (IQR 2–5) for severity scale 3, 4 (3–5) for scale 4, and 5 (4–6) for scale 5-6.[65]

Whereas one study assessing thrombotic complications in Covid-19 with a minimum of 90 days follow-up from critical care admission did not detect any post-discharge venous thromboembolism using CT or doppler, in a cohort of 129 patients.[60]

Another study assessed kidney function in people with Covid-19 –associated acute kidney injury (AKI) compared with people with non-Covid-19 associated AKI, and found that Covid-19–related AKI was associated with decreased kidney recovery during outpatient follow-up (median (IQR): 92.9 (52.5-127.7) days).[30]

Risk factors

Exploring the literature, we sought to produce a meta-analysis of risk factors for Long Covid. We found a considerable diversity of reported risk factors, including age, sex, comorbidities, ethnicity, and severity of the acute phase.

Several cohorts (64%, 25/39) assessed whether there was an association between the severity of initial Covid-19, including symptom load, level of hospital care, and need for mechanical ventilation, and the risk of persisting sequelae. An association between female gender and risk of prolonged recovery and post-acute SARS-CoV2 sequelae has been identified in longitudinal studies (20.5%, 8/39), as has the association between presence of pre-existing comorbidity,[34,49,51,57,62,66] age,[26,28,44,49,56,57] and minority ethnicity (e.g. Latinx and unspecified)[34,61].

The heterogeneity in study design and evidence base, limited number of studies including non-hospital cohorts, and lack of control groups limits the statistical conclusions or identification of a case definition at this time. We have summarised the reported significant associations to date (Supplement 14) and suggest that these associations be explored in prospective controlled trials to forward the evidence into long term Covid-19 outcomes in different at-risk populations.