The search was conducted until April 20, 2023. Two independent authors searched the PubMed, Scopus, and Web of Science databases for studies that provided information about cough as a predictor of any conceivable disease. The Cochrane Library was searched for additional data, supplemented by a search for gray literature in the web-based database of the German Journal of Family Medicine. To detect all relevant areas, a complex search strategy was developed, using not only the term “Cough,” but also terms presumed to be associated with cough. These were “Pulmonary Disease, Chronic Obstructive,” “Asthma,” “Pneumonia,” “Respiratory Tract Infections,” “Carcinoma, Bronchogenic,” “Gastroesophageal Reflux,” “Heart Failure,” and “Angiotensin-Converting Enzyme Inhibitors.” To identify studies reporting on cough as a predictor, we added the terms “predict*,” “Odds Ratio,” and “Likelihood Functions.” The terms “Family Practice” and “General Practice” were included to ensure relevance to primary health care. To exclude studies focusing on therapeutic issues, the search was combined with the term NOT “Therapeutics.” As far as possible, Medical Subject Headings terms were used. The start date for the search set was 1995 because this is the first year from which studies were available in sufficient volume. The complete search strategy is available in Multimedia Appendix 2. A total of 2 identified experts in the field of cough were contacted in April 2022 and asked about unpublished or ongoing studies.

After the exclusion of duplicates, titles and abstracts of the search results were screened independently by 2 reviewers to exclude clearly irrelevant articles. Then a full-text analysis was performed to identify all relevant articles. Discrepancies during the screening process were discussed at all stages at regular consensus meetings. A third experienced reviewer was consulted to resolve disagreements. The entire screening process was visualized using the PRISMA flowchart, as presented in Figure 1 (adaptation from the study by Page et al [25]).

We included all studies that reported data with diagnostic value related to cough. More precisely, only studies reporting on cough-related ORs, hazard ratios (HRs), positive predictive values (PPVs), negative predictive values (NPVs), and positive and negative likelihood ratios (LRs; LR+/LR-) were considered. Another inclusion criterion was for the study to have taken place in a primary care setting. We did not consider any therapeutic issues or studies reporting only the prevalence of cough. Further exclusion criteria were languages other than English or German and articles other than original studies (ie, reviews, letters, and case reports).

The extraction of valid data was conducted by 2 reviewers (NE and JC), if necessary, under the supervision of an experienced third reviewer (JS). Data extraction included authors and year, study design, study population, country, and main results concerning ORs, (HRs), PPVs, NPVs, and LR+/LR-. The results were presented in tabular form and sorted thematically by diseases.

The search from 1995 up until April 20, 2023, yielded 231 records from PubMed, 869 from Web of Science, and 831 from Scopus. In total, 7 articles were identified from the search of the Cochrane Library, and the search for gray literature in the German Journal of Family Medicine yielded 8 additional results. Of these 1946 results, 488 records were removed because of duplication (Figure 1). Title and abstract screening of the remaining 1458 records led to the exclusion of another 1299 articles. The remaining 159 articles were examined using a full-text inductive thematic analysis. The reasons for exclusion of 124 articles after full-text screening are presented in the PRISMA flowchart (Figure 1). Finally, a total of 35 articles were identified to meet the inclusion criteria. The 2 experts contacted were not aware of any additional unpublished or ongoing studies.

The included articles (for summarized results related to cough, refer to Multimedia Appendix 3 [26-60]) covered the following diagnoses: COPD, asthma, obstructive airway disease (OAD), influenza or influenza-like illness, respiratory tract infection (RTI), bronchial carcinoma (BC), community-acquired pneumonia, COVID-19, dyspepsia, gastroesophageal reflux disease (GERD), differential diagnosis of pulmonary embolism, heart failure, and adverse effects of angiotensin-converting enzyme (ACE) inhibitors. Of these studies, 10 were prospective studies (Vandevoorde et al [27], Freeman et al [30], Price et al [34], Eysink et al [38], Kable et al [39], Navarro-Marí et al [44], Thursky et al [45], Senn et al [46], Govaert et al [26], and Hollenz et al [59]), 8 were prospective cohort or diagnostic studies (Hamers et al [29], Schneider et al [37], Pescatore et al [40], Buffels et al [42], Schneider et al [53], Hippisley-Cox et al [54], Erkens et al [57], and Wallander et al [58]), 5 were cross-sectional studies (van Schayck et al [33], Melbye et al [36], Schneider et al [43], Kool et al [49], Hopstaken et al [52]), 4 were case-control studies (Haroon et al [35], Hamilton et al [55], Iyen-Omofoman et al [56], and Visser et al [60]), 2 were multicenter studies (Vrijhoef et al [32] and van Elden et al [47]), and 1 each was a descriptive (Thiadens et al [41]), retrospective (Nakanishi et al [51]), cohort (Geijer et al [28]), prospective case series (Bloom et al [50]), prospective, systematic sampling study (Sočan et al [48]) and comparative study (van Schayck et al [31]). Of these, 1 study included a secondary analysis of the data (Erkens et al [57]). Another study was listed as a randomized controlled trial, whereas the results listed here are considered prospective [26].

The number of patients varied widely; the largest study had >3 million patients, 5 included more than 5000 patients, another 5 had more than 1000 patients, 22 had between 100 and 1000 patients, and 3 had ≤100 patients. In total, 12 studies were conducted in the Netherlands; 7 in the United Kingdom; 4 in Germany; 2 each in Belgium and Australia; 1 each in Brazil, Japan, Norway, Slovenia, Spain, and Switzerland; and 1 combined in the United Kingdom and the United States. An overview of all the included studies is presented in Multimedia Appendix 3 [26-60].

In this scoping review, we included 35 studies reporting on cough as a predictor of different diagnoses in general practice. Most articles were found on obstructive airway diseases, followed by respiratory tract infections. Within these topics, symptom complexes yielded higher ORs than cough as a single symptom. A total of 3 articles were found on BC. Other entities were identified only in 1 article, respectively. Overall, data for diagnosis based on the symptom of cough are scarce. Comparisons of the studies have limitations because of the varying study sizes and designs.

The ORs for cough as a predictor of COPD are heterogeneous [27,28,30,31,33,36,37]. Overall, the study results suggest that cough is a relevant predictor of COPD, although only to a small extent as a single symptom. The combination of symptoms such as “cough and dyspnea” and “chronic cough and phlegm,” in turn, led to higher ORs through further specification [29,31]. Nevertheless, other authors report age and spirometry predicting COPD rather than the detection of clinical symptoms [61]. A relevance of the symptom cough for the diagnosis of chronic obstructive airway diseases other than COPD (ie, asthma and OAD) cannot be inferred from the inconsistent results of the rather small studies presented here [37,39-43]. Others report a stronger association of clinical symptoms and asthma [61]. However, they highlight wheezing and thoracic tightness, not cough, as the strongest predictors.

The results of cough as a predictor of infectious diseases are more consistent, although only a few studies are involved in each case. A significant correlation of influenza infection with “cough” was shown by all included studies on this topic, albeit with varying strengths of the associations [26,44-48]. The association was again enhanced by the combination of symptoms (here “cough and fever” and “fever, coughing, and acute onset”) [26,44,46]. This is in line with the situation in practice, where usually more than 1 symptom is considered for the diagnosis. Nevertheless, the predictive value might vary with years and the region where the infection occurs, as well as between children and adults [45,48] This could be related to different virus variants, for example. Only 1 included study reported on the COVID-19 infection [53]. Although it did predict the COVID-19-infection, a described clinical decision rule did not include cough. Others also stated a low predictive value of cough, fever, and headache for the diagnosis of the COVID-19-infection [62].

The included studies do provide evidence of associations of hemoptysis and serious disease (especially BC) [54-57], as well as a stronger association of cough and GERD [59] compared with cough and dyspepsia [58]. Nevertheless, the small number of results requires confirmation by further studies. Among ACE inhibitors, only enalapril was shown to be statistically significantly associated with incident cough [60]. However, this refers to only 1 study and therefore does not allow a reliable statement.

In summary, although cough is one of the most frequent reasons for seeing a GP, it is surprising how few studies on cough as a predictor of various diseases were identified, which could serve as a database in a diagnostic algorithm.

However, studies and reviews tend to start from the diagnosis (eg, pneumonia) [63,64] rather than from the symptoms or the reason for consultation. Thus, their conclusions are drawn backward from the diagnosis to the symptoms, as it were. Therefore, these studies represent an advanced diagnostic process with a level of uncertainty that is already reduced (“entropy”). To reach this stage in primary care, GPs actively reduce the uncertainties of an initially complex situation, meaning that the theoretically possible diagnoses have already been limited [65]. Accordingly, it is reported that analysis of the reason for encounters leads to lower PPVs than when using data documented by the GPs [66]. In contrast, although PPVs of reasons for encounter tend to be low overall, differences can still be noted. In a study in which the reasons for seeking primary medical care were analyzed in association with cancer diagnoses, the PPV of hemoptysis for cancer diagnoses was 2.7%, whereas cough as a reason for encounter yielded a PPV of 0.1% [66]. The chance of having pneumonia is reported to be higher in patients who presented with cough and fever than in those who reported only cough as the reason for encounter (16.4% vs 5.6%) [67]. A broader study base that considers this approach would be desirable as a starting point for probability calculations in primary care—not least in terms of its potential application in the context of AI. Furthermore, in the patient-centered setting of general practice with a wide array of possible diagnoses, the cognitive strategies used by clinicians often contain unusual phenomena (ie, symptoms, pattern failure, and a sense of alarm) [68]. A lot of tacit information flows preconsciously into a diagnosis. This is not reflected in the studies analyzed in this paper. It seems questionable how this database could be used to translate the reality of primary care into a reliable machine-based diagnostic algorithm. Thus, as already assumed for internal medicine, general medicine may be an area where the implementation of AI-based algorithms is more complex than in disciplines with frequent use of diagnostic imaging [11]. Future research needs to always include the sector-adapted prevalence of a disease when looking at PPVs.

Our search strategy aimed to identify the broadest possible range of relevant articles. Nevertheless, the search focused on studies written in English or German. Thus, we might have missed studies written in other languages. The wide range of numbers and ages among the participants limits the comparability of the study results. The largest study used a dedicated database for research in the United Kingdom based on physicians’ routine documentation. Thus, even in this large study, a bias because of differences in quality and missing documentation must be considered [54]. In addition, the focus on primary care might have led to an underrepresentation of severe acute illnesses, as studies reporting these conditions may be more likely to be conducted in a hospital setting. Nevertheless, some studies reported different results in different settings [37,69]. Therefore, the results of studies in a hospital setting need to be transferred with caution to general practice and vice versa.

Although the symptom of cough is one of the most frequent reasons for consultation in general practice, data for probabilities of the underlying causes are scarce. The complex processes that occur in general practice on the path from naming the reason for the consultation to finding the diagnosis are still underrepresented in studies that include cough as a predictor of various diseases. This should be considered when planning the development of machine-based diagnostic algorithms for general practice. Until solid evidence is available in this field, it seems questionable whether AI solutions can be programmed to adequately reflect reality.