Prevalence, aetiologies and prognosis of the symptom cough in primary care: a systematic review and meta-analysis

We conducted a systematic review including all studies evaluating the symptom “cough” as a reason for encounter in primary care. The methods were based on the PRISMA statement [9] and on recommendations for symptom-evaluating studies by Donner-Banzhoff et al. 2001 [8]. The study methods including eligibility criteria and analysis were pre-specified in a protocol. Our research group applied the same methods for the symptoms tiredness, abdominal pain, headache, chest pain, dyspnoea, dizziness, and back pain [10‐14].

We performed a systematic search in MEDLINE (2012) and EMBASE (2015), updated 2019 resp. 2020, addressing publications in English, German, and French. A snowball search included the reference lists of all articles and reviews. The search syntax combined the terms “cough” AND “general practice” in various notations OR their MESH terms in title or abstract. Alternatively, we considered papers on “cough” published in journals representing primary care research OR papers in which the term “primary care” appeared in different notations in the affiliation of at least the main author. The entire search syntax can be found in Additional File 1.

We screened titles and abstracts and the eligible full text articles with respect to the criteria given in Table 1. Eligible studies focusing solely on children were excluded from data analysis and will be published elsewhere.

All steps of the selection process (except its update in 2019/2020) were performed and documented by two reviewers (MB, DB/SS) working independently. In case of disagreement, the full text evaluation was revised, inclusion criteria were discussed, and, if necessary, an expert (AB) was consulted.

We extracted bibliographic data (author, publication year, title, journal), country, inclusion criteria, definition of cough, characteristics of physicians and practices, study design, sample size and study duration. For outcomes we extracted data concerning prevalence/incidence, underlying aetiologies and the prognosis of cough. Seven study authors were contacted to complement published data. In case of multiple publications, we extracted data from all eligible reports.

Due to lack of standardized guidelines for assessing risk of bias in symptom-evaluating studies, we followed the criteria published by Donner-Banzhoff et al. [8], which entail four domains with pre-specified key questions related to the potential of bias. Domain A and B refer to all studies dealing with the selection of patients and physicians (description of symptom, inclusion criteria, recruitment, multicentricity), data collection, and patient flow (study design, dropouts). Domain C refers to the aetiological outcomes (the definition of aetiological categories, diagnostic workup). Domain D assesses the quality of the prognostic data (definition of the outcome, inclusion of a comparison group, prognostic workup). Again, two reviewers (MB, KH), working independently, assessed the risk of bias.

We calculated proportions (with a confidence interval of 95%) on prevalence/incidence data and the underlying aetiologies. If sensible, a meta-analysis was performed. To visualize probability estimates and between-study variation of our data, we used forest plots. To ensure comparability, we grouped studies according to the estimates’ denominators, the duration of cough (both pre-specified) and regional characteristics (post hoc).

For meta-analysis we used the random effects model (assuming a distribution of effects across studies) to weigh estimates of studies in proportion to their significance [15].

Outcomes vary due to differences in study design and bias (methodological heterogeneity) as well as in study population, inclusion criteria, healthcare system and diagnostic workup (clinical heterogeneity) [15]. To quantify heterogeneity, we used χ², p-value, and I². A high χ² and a low p-value correlate with a heterogeneity beyond chance; I² describes the portion of variability that is not due to chance [15].

There were only a few heterogeneous studies providing evidence of prognosis for cough. Therefore these results were analyzed descriptively.

For statistical analysis we used the software R (R Foundation for statistical Computing, Vienna, Austria, version 3.4.4) and RStudio V (RStudio, Inc., version 1.1.442).

We identified 2,985 references in MEDLINE, 2,719 additional references in EMBASE, and 19 by snowball searching. Screening of titles/abstracts and full texts identified 73 eligible references, of which 60 publications (31 studies) reported data on adults or on patients of all age groups. Of these, 22 provided data on prevalence of cough in primary care, 12 on aetiology and 4 on prognosis. Further details are presented in Fig. 1.

Most studies were conducted in Western countries: In Europe (n = 12), in North America (n = 6), in both Europe and North America (n = 2), and in Australia (n = 1). Five studies collected data in Asia, four in Africa, and one in South America, Africa and Asia. Time of publication varied between 1969 and 2018. Studies included 32 to 158,863 patients, 121 to 337,348 consultations, and 385 to 284,348 reasons for encounters. Forty-two per cent to 75% of study populations were women; the overall age ranged from 0 to 103 years (the mean age was 24 to 50 years). One study recruited only patients 65 years and above. Except for one, the study population was recruited prospectively. Further details on study characteristics are presented in Table 2.

Depending on the selection of patients and GPs (Domain A) most studies had a low risk of substantial variation and of risk of bias. Referring to data collection and patient flow (Domain B) the risk of bias was found to be low in most studies (n = 20), and none had a high risk of bias. In diagnostic workup (Domain C) most showed a high risk of bias (n = 7). The risk of bias in the prognostic workup (Domain D) was low in one study, unclear in another, and had different assessments in two studies, depending on the prognostic category. Only seven studies had an overall low risk of bias. A summary is presented in Table 3; detailed methodological description and risk of bias can be found in Additional File 2.

Twenty-two studies presented outcomes on the prevalence of cough; nine of these show a low risk of bias. Figure 2 presents the prevalences and incidences of cough in Western primary care. Incidental consultations showed about three times as many estimates in comparison with prevalences. Outliers were characterized by study populations recruited in a single primary care practice with one or two GPs [38, 50] or by excluding consultations for cough of < 2 and > 15 weeks duration [54]. Comparably low prevalences were seen in a study population of patients aged ≥ 65 years [51] and in studies including not only consultations for symptoms, but also for prescriptions, follow-up visits, tests, procedures and administrative visits to the denominator [49, 50].

Studies with data collection in African, Asian and South American primary care settings show higher estimates of prevalence (13.8% for reasons for encounter and 10.3% for patients), while they show lower estimates of incidence (6.3% for consultations) (see Additional File 3). The presented estimates show a high heterogeneity across studies, indicated by high values of I² and χ².

Twelve studies assessed data on the aetiology of cough in primary care. Data referred to different durations of cough and a wide spectrum of differential diagnoses. Mostly, the given aetiologies were the working or presumptive diagnoses by the treating GPs, which correlate with a high risk of bias in the diagnostic workup process. No study had a low risk of bias in all categories. As there were differing denominators (reasons for encounter, (incidental) consultations, episodes of care, patients), no meta-analysis was performed and data is presented in forest plots (Fig. 3, Fig. 4). Data on acute cough and cough of all durations were collected in North America and Europe. The most frequent underlying conditions in acute cough were respiratory tract infections (ranging from 73–91.9%) and in cough of all durations, bronchitis/bronchiolitis (25.4–50.2%). Potentially serious diseases like pneumonia, chronic obstructive pulmonary disease (COPD), heart failure or suspected malignancy were rare. Findings on subacute/chronic cough derived from a study conducted in Zimbabwe (with an HIV prevalence of 83%) [45] and Malaysia [48], showing high prevalences of tuberculosis (6.0–43.0%) and pneumonia (2.8–16.0%) (see Additional File 4). The results of these studies are not applicable to the context of Western countries. The high quality study by Munyati et al. [45] is based on a sample with 83% HIV positive patients; the work by Nantha et al. [48] lacks sufficient information to estimate the risk of bias. In the foremost aetiological categories, we found substantial heterogeneity across studies, indicated by high values of I² and χ².

Four studies assessed prognostic outcomes, one with an overall low risk of bias. Studies included patients with acute cough of up to one [35] or four weeks [20, 30‐34, 57]. The follow-up duration was 28 days in all studies, assessed by a symptom diary or telephone interview.

The median duration of cough after first consultation was reported to be eight (IQR 6–14.5) days [30], with the median time to feeling recovered 9 [57] to 11 days. [34] The mean total illness duration was 20.4 days (standard deviation 10) in patients who felt recovered after four weeks [31]. A first improvement of cough was seen the third day after consultation in 52% of patients [35]. A major improvement or complete recovery was seen in 65.7% of patients after seven days and in 81.4% after 14 days [35]. 10.8% of patients felt completely recovered after seven days [35], 40.2% [35] to 67% [32] of patients after 14 days, and 79% [31] after 4 weeks. A prolonged illness (moderate or severe symptoms more than 3 weeks after consultation), was described in 7.9% of patients [32]. At day 28 after the first consultation, 21.3% of patients still didn’t feel recovered [31]. The re-consultation rate ranged from 21.1% [20] to 35% [30, 32]. Most patients re-consulted the GP during working hours (27.6%), 1.4% out of hours, 2.8% consulted a nurse, 2.7% a specialist, 0.5% a hospital emergency department and 17.2% visited a pharmacist [30]. Between 0% [30] and 1.3% [57] of patients were hospitalized for 3–3.5 days [57] because of cough. No patient died of cough during follow-up [32, 33].

Our study identified 31 studies evaluating the symptom cough in primary care. Data quality was heterogeneous with only seven studies having an overall low risk of bias. The prevalence of cough in Western primary care was 3.8–4.2%; the incidence was 12.5%. African, Asian and South American healthcare settings showed higher prevalences (10.3–13.8%) and lower incidences (6.3–6.5%). Respiratory tract infection (73–91.9%) was the most frequent aetiology in patients with acute cough; bronchitis/bronchiolitis was the most frequent aetiology (25.4–50.2%) in patients with cough of any duration. Other frequent underlying conditions in both were influenza (6–15.2%), asthma (3.2–15.0%), and laryngitis/tracheitis (3.6–9.0%). Serious diseases like pneumonia (4.0–4.2%), COPD (0.5–3.3%), heart failure (0.3%) and suspected malignancy (0.2–1.8%) were rare. Findings on subacute or chronic cough were based on two studies conducted in Zimbabwe and in Malaysia, showing high prevalences of infectious diseases (tuberculosis and pneumonia). For acute cough patients, the median time to feel recovered was 9 to 11 days. Complete recovery was reported by 40.2- 67% of patients after two weeks (79% after four weeks). 21.1- 35% of patients re-consulted, 0–1.3% were hospitalized and none died.

To our knowledge, there are no other reviews estimating the prevalence or incidence of cough in primary care. However, evidence is needed to set focus in priorities for research, resources, policy making, guideline development and training of primary care professionals [60]. In comparison with our data, the prevalence of cough in population-based surveys is higher (9% to 33%) than in primary care [1], most likely due to its self-limiting course. A population-based telephone survey in Italy showed that 23% of subjects would use domestic remedies, 21% would ask their pharmacist and only 33% would consult their doctor [61]. However, when it comes to consultation, for the majority of people (69.6%-73.7%) the GP is the first address [61, 62].

In Western countries, differences between prevalence and incidence estimates were quite high, with prevalences of about 4% and incidence at 12.5%. This is different in African, Asian and South American primary care settings (10.3–13.8% prevalence and 6.3–6.5% incidence). This might possibly be attributed to the high share of chronic diseases in Western countries, in relation to which cough is less relevant than when compared to a population with a high share of acute diseases. Moreover, study outcomes depend on cultural variance between countries (e.g. different healthcare systems, the patient’s own health traditions, and different thresholds for consulting a doctor) [14]. In developing countries with a higher rate of uninsured people and fewer health care providers (especially in rural areas) there are fewer consultations for self-limiting acute respiratory tract infections. Furthermore, environmental factors associated with poverty (cooking on an open fire and a higher burden of HIV-infections, accompanied by higher rates of tuberculosis) increase the prevalence of chronic cough.

International guidelines suggest classifying cough according to its duration, as either acute (< 3 weeks), subacute (3–8 weeks), or chronic cough (> 8 weeks) [5, 6, 63, 64], or as acute and chronic cough [7, 65‐67]. In fact, the most common definition for chronic cough is ≥ 3 months duration [68]. A categorisation seems necessary as acute cough is mostly caused by a respiratory tract infection, usually vanishing within two weeks [1]. In contrast, chronic cough is associated with a greater risk of serious diseases that require efficient treatment or referral [6]. This is confirmed by our results: we found respiratory tract infections to be the most common underlying conditions of acute cough, followed by exacerbations of asthma and influenza. This is in accordance with primary care guidelines recommending that laboratory tests, sputum evaluation, chest x-rays, and antibiotic treatment all be foregone when respiratory tract infection is clinically likely and no warning signs of serious disease are present [69].

Our results concerning aetiologies of chronic cough are based mainly on two studies from Malaysia [48] and Zimbabwe [45], with a cough > 2/ ≥ 3 weeks. Other than a study from Poland, assessing the prevalence of pertussis [54], we didn’t find any evidence for chronic cough in Western primary care and none concerning subacute cough. Our data do not confirm the big three causes of chronic cough (Chronic upper airway cough syndrome, asthma, and gastroesophageal reflux disease (GERD), nor any other differential diagnosis. The respective recommendations on subacute or chronic cough are based on secondary or tertiary care studies [6, 70]. In fact, given the different case mix, it is likely that the distribution of causes is different in primary care.

Accurate prediction of the course of cough could decrease antibiotic overprescribing [71, 72]. Half of antibiotic prescriptions for acute respiratory conditions in US ambulatory care visits seem to be unnecessary [73]. About 53% of acute cough patients in Europe receive antibiotics [34] – despite the high prevalence of underlying self-limiting viral infection [6, 74]. We found no death, a low rate of hospital admissions, an improvement in half of patients after three days and complete recovery in 79% of patients after one month. A benign course of acute cough was also found by Bruyndonckx et al. [71]. A systematic review assessing primary, secondary, and tertiary care found a weighted mean duration of any cough of 17.8 days (range 15.3 to 28.6 days) and 13.9 days for productive cough (range 13.3 to 17.4 days) [75]. In our study the mean total illness duration was 20.4 days (standard deviation 10). As for acute cough, symptom control without diagnosis ('wait and see approach') seems more sensible than investing in unnecessary diagnostic resources [76]. To reassure patients with low risk, and to confine patients with a high risk of complication, primary care prediction tools like RISSC85 [71] are helpful.

We didn’t identify any studies presenting evidence on prognostic outcomes concerning subacute or chronic cough in primary care; this should be addressed in future research.

Guidelines define cough of more than eight weeks as chronic [6, 63, 64]. In fact, the longest follow-up in prognostic studies was 28 days. Outcome assessment varied vastly across prognostic studies; accordingly, standardization seems mandatory. None of the included prognostic studies contained an untreated or alternative control group, leading to a high risk of bias.

Our work comes at a time when the epidemiology of cough has shifted due to the Covid 19 pandemic. Struyf et al. [77] performed a systematic review over the accuracy of Covid-19 symptoms in primary care and in hospital outpatient settings. They identified 44 studies, including three from primary care settings. In a sample including 21% patients suffering from Covid19, they found 65% of patients presenting with cough, of whom 142 would have Covid-19. The search strategy (searching for Covid-19 studies) was different from our study design and symptoms were actively asked for, so frequencies are overestimated. But even if the study had fit our requirements, these data would be outliers. During a pandemic, the prevalence of diseases and symptoms shifts. In addition, the utilization behaviour, the diagnostics and the frequency of aetiologies as well as the morbidity change. Interventions related to Covid-19 like facial masks are displacing diseases such as influenza and, at the same time, pneumonia is increasing as a cause of cough due to viral illness. Studies conducted during the pandemic are not comparable to the everyday situation of a family practice, which we would like to depict in our review. We must point out that the results of our study apply only to the periods leading up to the pandemic. After that, it will be important to examine whether behavioural changes (such as refraining from shaking hands) as a result of the pandemic will change the observed epidemiological data in our study.

Apart from this temporal classification, we must consider the typical weaknesses of a systematic review. Conclusions of any systematic review can only be as valid as the available literature and the accuracy of the included studies’ protocols [75]. Important aspects are (1) limitations to the internal validity of the included studies (e.g. imprecise inclusion criteria or incomplete recruitment); (2) criteria affecting the external validity of the included studies (e.g. characteristics of the setting, or recruitment practice); (3) methodological aspects of our review affecting the internal validity of our review (e.g. accuracy in literature search, screening process or data analysis); (4) aspects influencing the review’s external validity [10, 13].

Accordingly, we performed strict quality assessment and implemented clear inclusion criteria. Our research was comprehensive and thorough, with almost all abstracts and full texts screened by two reviewers. To minimize selection bias, we excluded all studies that explicitly included or excluded certain groups of cough patients and we contacted study authors to acquire missing information. Still, in some cases uncertainty remained regarding eligibility criteria, definition of outcomes or denominators of given data. This may have introduced error into our data synthesis.

We didn’t control the risk of bias across studies and the publication bias, as the number of studies concerning the respective outcome was too low. However, it is rather unlikely that prevalences of cough or underlying conditions are not published.

Limitations to our review are the substantial methodological and clinical heterogeneity across included studies. As Higgins et al. postulated “every amount of heterogeneity is acceptable, providing both that the predefined eligibility criteria for the meta-analysis are sound and that the data are correct” [78]. We built subgroups referring to denominators, duration of cough and cultural variances in healthcare systems. In aetiological outcomes, the formation of categories was difficult and overlapping of categories is likely. Given (sub-)categories differed widely. Denominators weren’t always specified, which may have influenced data synthesis.

The attribution of countries to the subgroups Western resp. African/Asian/South American countries corresponds with the United Nations classification system of developed and developing countries [79]. We didn’t use the latter terms, because people’s health demands depend not only on the economic situation of a country, but also on health systems, people’s health convictions and utilization of health care.

The assessment of the methodological quality and the risk of bias should be based on standardized checklists. Yet, there are no published criteria referring to studies evaluating symptoms [13]. Therefore our research group has developed a tool for assessing methodological quality and risk of bias, based on work done by Donner-Banzhoff et al. and on the Standards for the Reporting of Diagnostic Accuracy (STARD) on diagnostic accuracy studies [8, 80]. Applying our tool, we found an overall low risk of bias in only ten studies with prevalence outcomes and in one study with prognostic outcomes, while there was no such study presenting aetiological results. The latter is caused mainly by the fact that the majority of aetiological studies evaluated clinical diagnoses without a standardised diagnostic approach or follow-up. Despite these limitations, most studies in subgroups had similar results, and we think our results are currently the best approach wehave to guide the GP in his everyday decisions.

Statistical limitations can be quantified. Content-related aspects can only be discussed and made transparent. We discussed seasonal effects and differences between countries. We ourselves see no reason to exclude older studies as long as they meet the inclusion criteria, and as long as their sample shows an appropriate external validity. This would be different if we knew of any event that calls into question the epidemiological situation at the time, but as far as we know there is nothing we have to consider. If we were already 10 years further along, we would probably exclude the studies of today because of the special situation under pandemic conditions.