Multidimensional analyses to assess the relations between treatment choices by physicians and patients’ characteristics: the example of COPD

The following data were recorded: socio-demographic characteristics, risk factors (active and passive smoking, occupational exposures), history of COPD (year of first symptoms and year of diagnosis), comorbidities, number and severity of exacerbations, most recent lung function data, symptoms (cough, sputum production, MRC dyspnea grade) using questions derived from the European Communicty Respiratory Health Survey [24], usual maintenance pharmacological and non-pharmacological treatments, yearly number of visits to GPs and chest physicians, investigations performed during the previous year. Exacerbations were defined as an increase in cough and/or sputum production and/or dyspnea during at least 48 hours and requiring a change in treatment (i.e., at least an increase in the dose of bronchodilators). Their severity was classified as mild (self-managed change in treatment), moderate (requiring a visit to a physician), severe (hospitalisation in a medical ward) or very severe (hospitalisation in an intensive care unit).

The planned sample size was set to 2400 patients to obtain a 2% precision in the prevalence estimate of a patient characteristic that is present in about 50% of the population. A two-step process was followed, first identifying typologies of patients and treatments, then exploring their relationships.

Two analyses were used to find typologies of patients on the one hand and typologies of treatments on the other: the multiple correspondence analysis (MCA) and the ascending hierarchical complete linkage clustering analysis (AHCLCA)[25, 26]. MCA allows mapping patients’ characteristics or treatments in a reduced space according to their deviation from independence (association or mutual exclusion). Clustering is based on the distance between variables, which is equal to 1 minus the correlation. This method forces the inclusion of all variables in a given group. As a consequence, small groups are constituted of positively related variables, while larger groups may pool together variables that may be unrelated to each other or even negatively related. Results of MCA and AHCLCA were combined to find a typology.

Once groups of patient characteristics and groups of treatments were defined, multiple logistic regression was used to assess (i) associations between each clinical subtype and the 6 treatment subgroups and (ii) associations between each treatment subgroup and the 6 clinical subtypes, to identify the clinical subtype to which a given treatment type is preferentially prescribed or in which it is underprescribed. These two approaches should provide convergent but not exactly similar results because the odds ratios measuring associations between clinical subtypes and treatment subgroups are not adjusted on the same set of covariates (adjustment on other treatment subgroups in one case and on other clinical subtypes in the other). Canonical analysis of correspondence (CAC)[25, 26] and more precisely principal component analysis (PCA) with respect to instrumental variables [27] was then used to map treatment variables and patient characteristics in a reduced space and identify (without any clustering) which patient characteristics are associated to which treatments. Two canonical redundancy analyses were performed: the first one with prescribed COPD medication including oxygen therapy as response variables (16 variables), the second with other (i.e., non-pharmacological) COPD treatments and patient follow-up variables (8 variables: number of antibiotic courses, homeopathy, flu vaccine, pneumococcal vaccine, chest physiotherapy, frequency of visits to respiratory physician, frequency of visits to general practitioners and frequency of lung function tests).

More details on these methods can be found in the electronic supplementary material.

About 3000 respiratory physicians (i.e., all French respiratory physicians) were contacted; 515 accepted to participate and were compared to a random sample of 245 non-participating physicians: no significant difference was found in terms of type of clinical practice (general hospital, university hospital, private practice). The 515 participating physicians were harmoniously distributed on the whole French mainland territory. They recruited 2494 patients, among whom recent lung function data was unavailable in only 10 (protocol deviations). The general characteristics and treatments of the population are summarized in Table 1.

Detailed results of MCA and cluster analysis of clinical subtypes are provided in the electronic supplementary material. Altogether, MCA allowed defining 9 groups within the first four interpretable axes, while 7 clusters were identified from clustering. Putting all these results together lead to define 6 clinical subtypes (Table 2). It has to be noted that these subtypes are not exclusive: for instance, a given patient may belong to subtype 4 (women) and subtype 1 (mild severity).

Detailed results of MCA and cluster analysis of treatment characteristics are provided in the electronic supplementary material. Briefly, again four axes of the correspondence analysis were interpretable and MCA allowed defining 5 groups while clustering found 6 clusters. The combination of ordination and clustering lead to distinguish 6 treatment subgroups (Table 3). Oral corticosteroids belong to treatment subgroup 1 (nebulised treatments) because they can be associated with (i.e., they are not negatively related to) nebulised drugs. MCA and clustering show that treatments subgroups 1 and 2 (fixed ICS + LABA combinations) are usually mutually exclusive (opposite position on axis 1, and linkage at a negative correlation). Like clinical subtypes, treatment subgroups are not exclusive and two types of treatments may therefore be prescribed to the same patient.

Odds ratio (OR) estimated by the 12 multivariate logistic regressions are presented in Table 4. Each row indicates whether the clinical subtype is overrepresented (OR > 1) or under-represented (OR < 1) in the prescriptions of each treatment type. Each column indicates if the treatment type is preferably prescribed (OR > 1) or preferably avoided (OR < 1) in patients of a given clinical subtype. Odds ratios within the same cell of the table are slightly different because they are not adjusted on the same set of covariates (see the methods section). Overall, the results show that associations are more quantitative than qualitative in that all treatment types are significantly less prescribed to less severe patients (i.e., patients subtype 1) and more prescribed to the most severe (patients subtype 3). More specific associations were as follows: fixed combinations (treatment subgroup 2) are significantly more prescribed to elderly patients (patients subtype 6) while nebulized treatments (treatment subgroup 1) are significantly less prescribed to these patients. Vaccines (treatment subgroup 6) are more prescribed in patients with symptoms of chronic bronchitis (subtype 5) and respiratory support (treatment subgroup 4) is significantly related to clinical subtype 2 (overweight smokers with comorbidities).

Only 6.23% of variations in the use of pharmacological treatments are explained by patient characteristics as described by 6 families of variables: Table 5 shows the components of each of these families and the percentage of treatment variations that it explained. Altogether, the described associations are vague and reflect trends or tendencies rather than tight relations. Significant explanatory variables for components of care other than medications are shown in Table 6. Altogether, 14.29% of variations of these components of care are explained by patient characteristics.

Using factor analyses was justified primarily by (i) the large number of patients characteristics and possible therapies as well as (ii) the known wide overlap between GOLD stages of airflow obstruction for most clinical variables used to describe patients (such as dyspnea or exacerbation frequency), which makes it difficult to identify subtypes using conventional analysis. Factor analysis is basically a descriptive approach that summarizes associations without any a priori hypothesis. Indeed, it confirmed that many factors other than FEV₁ add to the description of the population: risk factors, symptoms (chronic bronchitis, dyspnea), comorbidities, demographics (age, gender) and anthropometric characteristics (weight) are also important contributors to characterization. The main limitation of these methods is their high dependency on active variables that participate to the construction of axes. Addition or subtraction of some variables (characteristics) may change the typology. Moreover, the choice of characteristics constituting one group (subtype) is not fully objective (preferential collection of characteristics that are believed to be relevant). Thus, the choice of recorded variables may be questioned: some variables that may predict treatment choices may have been omitted when building the case-report form; this could participate to explain why clinical characteristics as a whole were poorly associated to treatment subgroups. However, it must be outlined that the selection of collected data included all patients and disease characteristics that are routinely accessible to any practitioner: therefore, if some additional information was lost (e.g., degree of emphysema on CT-scan, lung volumes or diffusion capacity, characteristics of airway inflammation as measured in induced sputum), it would not be relevant in routine clinical practice.

The large sample size of the present study provides a good power for detecting associations (80% chance of detecting a true correlation of 0.057 and 70% chance of detecting a true correlation of 0.05) and therefore allows considering that, if some real relations are not detected, they would be very small and of no clinical relevance. However, it might also be that treatment choices rely on factors completely different from those recorded in our study. We believe that this last explanation is rather unlikely since virtually all routinely collected clinical and lung function characteristics of COPD were assessed.

Another limitation to be addressed relates to the statistical methods used for factorial analyses: a group of redundant characteristics (information) may be well represented because of their association (redundancy) while an independent clinically relevant information could be lost in the blank noise and not be well represented by the interpretable axes. The risk of such an event is minimized by the use of 3 different statistical approaches that search for associations through different procedures. For example, cluster analysis does not lose any variable in the blank noise.

We did not include the new GOLD classification (A-B-C-D) [1] into the analyses, Indeed, our purpose was to study the relationship between treatment decisions and patients’ characteristics with no a priori hypothesis rather than to test the relations between treatments and already defined patients subgroups such as those proposed in the new GOLD document. Interestingly, the variables that participate most to the definition of the patients subgroups identified by cluster analyses actually do not fit with those used in the GOLD A-B-C-D classification (see Table 2). Therefore, this classification is quite unlikely to by closely related to treatment choices. Finally, the new GOLD classification was not available when the study was performed, and could therefore not be used per se for treatment choices by physicians.

While care for COPD is quite heterogeneous in Europe, the present study was performed in only one European country where local guidelines (taking into account the GOLD documents) exist and have been widely disseminated, Therefore, it might have been hypothesized that practice would be more homogeneous than in a multinational survey. Even though, it remained difficult to determine how physicians make their choices.

There appeared to be separate gradients in the use of ICS + LABA, short acting bronchodilators, nebulised treatments and oral corticosteroids, and respiratory support. Such data suggest that several independent (at least in part) factors influence treatment choices. The very low variation of pharmacological treatments explained by clinical subtype is in line with data from other studies, which showed that the treatment of COPD is notably heterogeneous, either for stable disease or exacerbations. This can be due to several factors. Firstly, some physicians may have their own “specific” and “patient-independent” therapeutic habits. This could have prevented us from detecting an overall strategy since the investigator effect could not be taken into account due to the too low number of patients per physician. Secondly, the investigator’s therapeutic strategy may depend mostly on what was already prescribed to the patient in the past. This might not be detected since this was a cross-sectional study, providing only a snapshot of the situation at a given time and preventing from capturing the evolution of the treatment strategy over time. In addition, we did not record treatment changes that were decided during the visit, which would have allowed us to detect how new prescriptions were influenced by previous treatments. Thirdly and most importantly, physicians may consider that currently available treatments are not markedly different in terms of efficacy and safety profile. This would be in accordance with the impossibility to identify well-defined subgroups of responders or non-responders in recent major treatment trials [16‐20]. This can be explained in two ways: either recorded patients characteristics are not the correct predictors of response, or response is not adequately defined (e.g., by FEV₁, which is used in most studies). Indeed, available guidelines are quite vague about clinical criteria that may allow to choose, e.g., between available classes of inhaled bronchodilators. This relates to the lack of convincing evidence on the superiority of any treatment in well-defined subgroups, even in large randomised controlled trials [16, 17]. Considering these observations, there is a clear need for further research on phenotypes and predictors of response to treatments.

Even though, efforts should also be directed at improving adherence of physicians to guidelines and understanding the determinants of their therapeutic choices. Dissemination is obviously required but also clearly insufficient to ensure wide guidelines implementation and appropriate treatment use. Barriers have to be identified and may relate to guidelines themselves (when unclear or too difficult to implement) but also to physicians (inertia,lack of awareness, familiarity, agreement, self-efficacy, outcome expectancy, ease of prescription, anticipated drawbacks), patients (underlying the importance of taking their preferences into account) or healthcare systems (“external barriers”). These aspects have been extensively reviewed elsewhere, especially in a 2005 Canadian Wokshop focused on asthma and COPD [28].

However, it must also be underlined that explaining 6.23% of treatment variations by patients characteristics is not amazingly low compared to results obtained in other therapeutic areas: in most studies of the same kind with other medications available on the market (e.g., anti-depressants), less than 10% of treatment choices can be attributed to specific patients characteristics. Thus, even if the percentage of explained variation is not large, relationships were found and the two main criteria for the choice of treatments according to current guidelines, FEV₁ and exacerbations, were confirmed as predictors of treatment choices in MCA and canonical analyses. Moreover, it was possible to go further and find far more associations with other factors including dyspnea and comorbidities. Finally, it must be kept in mind that individual treatment responses were not recorded in the study. Therefore, our findings do not refute the hypothesis of specific phenotypes based on response to treatments. Actually, we cannot test this hypothesis due to the cross-sectional nature of the study. While we can conclude that the choice of a given treatment does not markedly relate to patients’ characteristics at a given time-point, we cannot exclude that it relates to improvements reported by the patient or recorded by the physician, following the first prescription of this treatment.

The % of variation of non pharmacological treatments that is explained by clinical characteristics of patients represents more than twice the amount of variation of medications for COPD that was explained by these characteristics. Components of care other than medications are thus more dependent upon patient characteristics or/and more stereotyped across physicians than medication strategy, which may suggest greater confidence in the specific (or differentiated) effects of non-pharmacological treatments. Moreover, clinical variables explaining non-pharmacological treatments are not strictly the same as for medications, which reflects the use of different patient characteristics for adjusting different components of care.