Phenotypes of cough in children: a latent class analysis

2.1 Study design and population

The Swiss Paediatric Airway Cohort (SPAC) is a prospective clinical study integrated into the routine care of 8 paediatric respiratory outpatient clinics and 2 paediatric practices across Switzerland. SPAC design and methods are described in detail elsewhere.^[29] SPAC enrols children aged 0–16 years referred by general practitioners or paediatricians for the evaluation of common respiratory problems, such as cough, wheeze, and exercise-induced symptoms. At time of referral, parents or guardians complete a baseline questionnaire about respiratory symptoms, socio-demographic information, environmental exposures, personal and family history, and comorbidities. One year after completing the baseline questionnaire and yearly thereafter, parents receive a follow-up questionnaire asking about symptoms and treatments during the past 12 months. From hospital records, we retrieve information about results of diagnostic tests and paediatric pulmonologist diagnoses. The Bern Cantonal Ethics Committee (Kantonale Ethikkomission Bern 2016-02176) approved the study; parents and patients older than 14 years provided their written consent.

2.2 Inclusion criteria

We only included children aged 5–16 years at the time of visit because physiological traits, such as lung function and allergies, are routinely measured for this age group. We included all children with completed baseline parental questionnaires by December 31, 2020 who indicated suffering from any recurrent cough. We used 4 validated and commonly used questions about cough: “Does your child have a cough even without having a cold?” (cough without a cold);^[30] “In the last 12 months has your child had a dry cough at night, apart from a cough associated with a cold or a chest infection?” (dry night cough);^[31] “Do you think your child coughs more than other children?” (cough more than others);^[32] and “Has your child had a cough in the past 12 months that lasted more than 4 weeks in a row?” (chronic cough). We excluded children who reported wheeze in the past 12 months (current wheeze), but not those who reported wheeze earlier in life (> 12 months ago) or with physician-diagnosed asthma earlier in life.

2.3 Definition of indicator variables

As indicator variables for the latent class model, we included the following 9 variables, all reported during the last 12 months from the baseline questionnaire: frequency of colds (≥ 7/year), pneumonia, otitis media, regular snoring, rhino-conjunctivitis, dry night cough, type of cough (dry versus wet), duration of cough (> 2 months), and number of allergic cough triggers, including pets, pollen, and house dust, ranging from 0 to 3 (Table S1). We selected these cough-related symptoms and characteristics based on a consensus among authors as they were considered to be potential indicators of underlying aetiologies that can be easily asked in all clinical settings, including primary care.

2.4 Definition of variables to characterise and validate identified phenotypes

We characterised and validated the identified phenotypes by comparing how these differed according to a) parent-reported information, b) paediatric pulmonologist diagnoses, c) results of measurable traits, and d) prognosis of respiratory symptoms 1 year later. First, we compared the phenotypes with respect to exposure, family history, and comorbidity information reported by parents in the baseline questionnaire. Second, we compared the phenotypes with diagnoses given by paediatric pulmonologists who were unaware of the LCA results. For simplicity, we grouped diagnoses into broad categories— asthma, chronic cough, allergies, and others. A child could have several diagnoses. Appendix S1 provides details of parent-reported information and physician diagnoses. Third, we validated the identified phenotypes against measurable traits. We used body mass index (BMI), spirometry, bronchodilator response (BDR), fractional exhaled nitric oxide (FeNO), and allergy tests [skin prick test (SPT) and allergen specific IgE (sIgE)]. We converted BMI into z-scores using reference values from the World Health Organisation.^[33] For spirometry, we transformed forced vital capacity (FVC), forced expiratory volume in the first second (FEV1), FEV1/FVC, and the forced expiratory flow between the 25% and 75% of the FVC (FEF 25–75) into z-scores using the Global Lung Initiative reference values.^[34] We defined a positive BDR test as an increase in FEV1 ≥ 12% (from baseline FEV1) after the use of a short acting beta-agonist. FeNO measurements differed across clinics, with cut-offs varying from 10 parts per billion (ppb) when using an offline method to 20–25 ppb for online methods.^[35] Thus, we first distinguished low versus high FeNO based on specified cut- offs for each centre. We also assessed FeNO (measured online) as a continuous variable and expressed in ppb. A positive allergy test was defined as either a positive SPT (mean wheal size of ≥ 3mm) or a positive sIgE with values ≥ 0.35 kU/L. Since SPAC is an observational study, physiological measurements are performed based on local protocols and indications and not for all children. However, availability of measurements did not differ across phenotypes (Table S2).

Fourth, we compared prognosis 1 year later between phenotypes by looking at reported cough without a cold, dry night cough, cough more than others, cough more than 4 weeks, and wheeze during the past 12 months. Parents of 360 (68%) children completed the first year follow-up questionnaire and response rates did not differ by phenotype (Table S2).

2.5 Statistical analysis

To identify different phenotypes of cough, we fitted latent class models to the dataset including the selected cough-related variables. To select number of latent classes (interpreted here as phenotypes), we repeatedly fitted the model with the number of classes increasing stepwise from 2 to 10 and used Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) to select the optimal model.^[36] These criteria favour models that are parsimonious, i.e. have few model parameters (in our context this means fewer classes), while still achieving good model fit. For the final selection, we also took into account interpretability and clinical relevance. We fitted models using the generalised structural equation modelling function (gsem) in STATA. The dichotomised symptom variables were assumed to be binomially distributed within classes and the number of allergic triggers as Poisson distributed.

For validation of phenotypes, we first assigned each individual to the phenotypes for which it had the highest posterior membership probability.^[14] We then compared the resulting groups with respect to parent-reported characteristics, paediatric pulmonologists diagnoses, measured traits and prognosis of respiratory symptoms after 1 year. We tested differences between phenotypes using chi-squared and Fisher exact tests for categorical variables. For non-normally distributed continuous variables we used Kruskal –Wallis tests.

We recoded missing values for questions about symptoms as “no” since we assumed absent or mild symptoms when parents have not answered “yes” (Table S1). We performed all analyses in STATA (Version 15.1, StataCorp, TX, USA).

3.1 Characteristics of study population

Our analysis included 531 children with median age 11 years (interquartile range 8–13) (Figure S1, Table 1). Parents of 446 (84%) children reported cough without a cold, 289 (54%) cough more than others, 276 (52%) dry night cough, and 215 (40%) cough more than 4 weeks.

3.2 Description of phenotypes

Results of BIC and AIC were discordant: the lowest BIC indicated a model with 2 phenotypes and the lowest AIC a model with 4 phenotypes (Table S3). Since identifying cough phenotypes using LCA is exploratory, we present the model with 4 phenotypes which we think is more clinically-relevant, assuming there are more than 2 underlying causes for recurrent cough among children^[13]. We summarise class size, main characteristics, and summary labels of the 4 phenotypes in the textbox, Figure 1, and Table S4.

• Download figure
• Open in new tab

Figure 1:

Estimated class sizes (in brackets)* and estimated probabilities of cough-related symptoms and characteristics of the four-class model identified by latent class analysis based on the lowest Akaike Information Criterion (N=531)

*The width of the lines represents the size of the classes, i.e. the proportion of children in the different classes. The estimated probabilities which can take values from 0 to 1 and represent the probability of answering ‘yes’ for the dichotomous indicators variables (frequent colds ≥ 7, pneumonia, otitis media, night cough, rhino-conjunctivitis, snoring, mostly dry cough, cough > 2 months) for each of the 4 classes, while the estimated mean which can take values from 0 to 3 represents the average number of allergic cough triggers for each of the 4 classes.

For completeness, we also present the model with 2 phenotypes in Table S5; it identified a larger phenotype (73% of children) with allergic features (rhino-conjunctivitis, allergic triggers), as the four-class model, and a smaller, less specific phenotype.

3.3 Validation of identified phenotypes

Parent-reported characteristics, paediatric pulmonologist diagnoses, measurable traits, and respiratory prognosis varied between the different phenotypes. Children with phenotype 2 were more often male (72%) and younger (median age: 7 years) compared with other phenotypes (Table 2). Family history of asthma and hay fever and atopy-related comorbidities, such as previous wheeze, physician-diagnosed asthma, hay fever, and eczema, were reported for all phenotypes but were most common among children with phenotype 4 (allergic cough).

Paediatric pulmonologist diagnoses differed across the 4 phenotypes (Table 3). Asthma diagnosis was most frequent in phenotype 4 (48%), followed by phenotype 1 (26%), phenotype 3 (20%), and phenotype 2 (6%). Phenotype 3 contained the highest proportion of children with a specific cough diagnosis (33%) by a paediatric pulmonologist, whilst in phenotype 4 this number was smallest (8%). Children with phenotype 4 more often received a diagnosis of rhinitis/rhino-conjunctivitis and allergic sensitisation/atopy.

Regarding measurable traits, BMI did not differ across the 4 phenotypes (p=0.161) (Table 4). Median z-scores of spirometry parameters (FCV, FEV1/FVC, and FEF25-75) were lowest for phenotype 2 and highest for phenotype 3 (Table 4, Figures 2A and S2). Fewer than half the children had a BDR test done and among those, the proportion with a positive BDR test did not differ across phenotypes (p=0.406). FeNO values differed across groups (p<0.001) being highest among children with phenotype 4 [17.9 ppb, 95% confidence interval (CI) 15.6–19.5] and lowest for children with phenotype 2 (7.0 ppb, 95% CI 5.0–17.7) (Table 4, Figure 2B). Similarly, positive allergy tests differed across phenotypes (p< 0.001) and were more common among children with phenotype 4 (70%, 95% CI 64–76), and relatively rare for children with phenotype 1 (31%, 95% CI 23–41).

• Download figure
• Open in new tab

Figure 2.

(A): Distribution of spirometry parameters FEV1/FVC z-score (N=436) and (B) Distribution of FeNO values based on the 4 identified cough phenotypes (N=450)

FEV1: Forced Expiratory Volume in one second; FVC: Forced Vital Capacity; FeNO: Fractional Exhaled Nitric Oxide; ppb: parts per billion; include FeNO values measured using the online method

Class 1: Unspecific dry cough

Class 2: Non-allergic infectious and night cough with snoring and otitis

Class 2: Allergic dry night cough with snoring

Class 4: Allergic cough

Regarding prognosis after 1 year, overall, cough was less often reported at follow-up (Figure 3) compared with baseline (Table 1). Children with phenotype 3 more often reported cough without a cold (66%), cough more than others (52%), and cough more than 4 weeks (24%) compared with children in other phenotypes (Figure 3). Night cough was more prevalent for children with phenotype 2 (57%) followed by children with phenotype 3 (41%). Wheeze was reported among 78 (22%) children and the proportion differed across the 4 phenotypes (p=0.005). Children with phenotype 4 wheezed almost 3 times more after 1 year compared with children in phenotype 1.

• Download figure
• Open in new tab

Figure 3:

Prognostic outcomes after 1 year for the 4 cough phenotypes

Class 1: Unspecific dry cough

Class 2: Non-allergic infectious and night cough with snoring and otitis

Class 3: Allergic dry night cough with snoring

Class 4: Allergic cough

4.1 Strengths and limitations

Ours is the first study identifying cough phenotypes in a clinical cohort of children by applying a data-driven unsupervised learning approach to a wide range of cough symptoms and characteristics. We used broad inclusion criteria—answering positively to any question about cough—for our study population, which minimised the risk of missing some phenotypes. In a previous population-based study, we found only a partial overlap between positive answers to different cough questions, which likely reflects the existence of subgroups of children with differences in underlying pathophysiology.^[37] For phenotype definitions, we used only cough-related symptoms and characteristics easily collected in primary care settings to make our results applicable to all healthcare settings. Lastly, we used different ways to validate the 4 phenotypes, including parent-reported information, paediatric pulmonologist diagnoses, measurable traits, and 1-year prognosis to help ascertain clinical relevance. Our study also has some limitations. First, despite our use of a data-driven approach, selecting variables to include in the model always involves some degree of subjectivity. Second, our baseline questionnaire lacked information about signs and symptoms of some possible aetiologies of cough, such as gastroesophageal reflux and psychogenic and habit/tic cough, which thus could not serve as indicator variables to build the latent class model. Third, our study population is mainly representative of German- and French-speaking families in Switzerland and includes few recent immigrants. Finally, we did not externally validate our findings among an independent clinical population of children— an interest for future studies to determine if we identify similar phenotypes. However, we are unaware of a similar study for validation.

4.2 Comparison with other studies

Our study is the first of its type to use LCA for identifying cough phenotypes among a clinical population of children. We cannot fully compare our findings with population-based studies also using LCA or similar clustering techniques to distinguish cough phenotypes.^{[15, 26}-^{28, 38]} We provide a summary of study designs and populations, inclusion criteria, indicator variables, identified phenotypes/trajectories, and validation processes from existing studies in Table S6. Spycher et al. and Weinmayr et al. included children who wheeze and used mostly wheeze-related symptoms with only few questions about cough as indicator variables in the LCA. ^{[15, 27]} Spycher et al. also included physiological measurements as indicator variables in the LCA.^[15] In contrast, we excluded children who wheezed, and as indicator variables, we used a rich set of only cough-related questions allowing us to identify more cough phenotypes and to describe them in more detail.

Divaret-Chauveau et al. and Rancière et al. investigated cough without cold and dry night cough trajectories, respectively.^{[28, 38]} Both studies used only cough questions collected at different time points to define cough trajectories/patterns, but they included children who wheeze; and the children were also younger (Table S6). In an analysis stratified by presence or absence of wheeze, Divaret-Chauveau et al. found physician-diagnosed asthma reported among a small proportion of children in all cough trajectories (2 to 12%) even in the absence of wheeze.^[28] Rancière et al. also reported physician-diagnosed asthma ever in all cough patterns (7 to 26%) but this included children who wheeze.^[38] In our study, parent-reported, physician-diagnosed asthma ever was also reported in all identified cough phenotypes, yet in much higher proportions (22 to 55%). This might be explained by the difference in study populations, since children visiting respiratory outpatient clinics likely present with more severe respiratory problems than children from the general population.

4.3 Interpretation and implication of findings

Our study distinguished 4 phenotypes of cough among children. Based only on parent- reported characteristics, distinctions between different phenotypes were not clear-cut, yet reflect the realities and difficulties of clinical practice when trying to establish underlying aetiologies of cough based on history alone with further examinations usually required. Nevertheless, interesting and clinically-relevant patterns, such as cough driven by allergies and upper airway pathology, correlating with several parent-reported characteristics, paediatric pulmonologist diagnoses, measurable traits and prognosis, emerged from the latent class model.

Interestingly, most children (70%) had phenotypes related to allergy (Phenotypes 3 and 4). In terms of clinical implications, children who cough with allergic symptoms possibly need allergy and asthma work-ups even in absence of reported wheeze. Divaret-Chauveau et al. also concluded as such; they found an association between different cough trajectories and asthma/allergic rhinitis.^[28] Similarly, Rancière et al. found an association between children with rising cough patterns—late and persistent coughers—and allergic comorbidity, including asthma, and atopy, independent of the presence of wheeze.^[38]

Additionally, the relatively high proportion of children reporting wheeze at follow-up (overall: 22%, range 10 to 29% for all 4 phenotypes) supports the need for further asthma work-ups. Our statistical model identified specific groups of children who more often develop wheeze and possibly asthma in the future. Spycher et al. also found that risk of reporting wheeze at follow-up was higher among children of the persistent cough phenotype compared with the transient cough phenotype (25% vs 10%).^[15] Rancière et al. found prevalence of wheeze at follow-up lowest among preschool children with infrequent night cough (4%) compared with transient (8%) and rising pattern (consisting of late and persistent cough, 16%).^[38] Findings from these studies and ours highlight the importance of identifying subgroups of children with cough since their prognoses differ.

Another interesting finding of our study is the identification of 2 cough phenotypes potentially driven by upper airway problems—1 allergic (phenotype 3) and 1 non-allergic (phenotype 2). The presence of night cough and snoring, long duration of cough, and rhino- conjunctivitis in phenotype 3 suggests house dust mite allergy. Children with phenotype 2 also had a high probability of night-time cough and snoring, yet in contrast to children with phenotype 3, without an allergic component, with more frequent infections (colds, pneumonia, otitis), and rarely reported dry cough. Phenotype 2 possibly represents children with recurrent upper airway infections and adenoid hypertrophy or postnasal rhinorrhoea— a small group (4% of children) perhaps restricted by our study population of children aged 5 years and older when respiratory infections become less common.^[10] This group also had the lowest estimated median lung function parameters but the number of children in this group was small and estimates were subject to much uncertainty.

The second largest group (phenotype 1: unspecific dry cough) possibly represents children with the so-called non-specific isolated cough—a “dry cough in the absence of identifiable respiratory or known aetiology.”^[39] This group of children seems to have a more transient cough, as compared with the other phenotypes, their prognosis after 1 year was better with less cough and less wheeze.