Introduction

Comorbidities in chronic conditions such as Chronic Obstructive Pulmonary Diseases (COPD) are known to influence prognosis, health status and therapy options [1]. For large cohorts the assessment of comorbidities is challenging and different tools are used to obtain and evaluate information on coexisting diseases (e.g. Charlson-Comorbidity-Index [2], ATS-DLD-78 [3]).

Self-reporting of diagnoses is a common approach to collect data on health status [4]. Much work has been invested to assess the value of self-reported data in comparison to other sources of information like pharmacy records on medications, or medical records from general practitioners (GP), nurses etc. [5–8]. Information on comorbidities given by the patients has been described to be reliable [9–11] especially for chronic conditions such as diabetes or heart disease [12]. Despite this, depending on the disease and several influencing factors it was found that in some cases patients’ reports tend to underestimate relevant comorbidities [5–7] but overestimation could also be demonstrated [12]. Therefore it has been recommended to take into account all available sources of information in the assessment of comorbidities [7, 13].

To identify and predict frailty in the elderly, Coelho et al. not only collected health data from self-reports but also evaluated specific medications [14]. Medication was categorized into groups of indications (e.g. cardiovascular, metabolic), however the categories were rather broad and no detailed description was given how to categorize and match the medications. The need for a combined evaluation of comorbidities and medication is also evident in the German COPD cohort COSYCONET which comprises more than 2500 patients [15–17]. In order to use as much information as possible we developed a novel categorization in which the concordance between medication and patients’ reports on comorbidities is evaluated. This approach which is described in all necessary detail might also be useful for other large cohort studies on respiratory or other diseases probing the influence of comorbidities on outcome and prognosis.

Methods

COSYCONET is a multicenter cohort study focusing on disease progression over time in interaction with comorbidities [15]. All assessments were approved by the central (Marburg (Ethikkommission FB Medizin Marburg) and local ethical committees (see S1 Ethics Committees) and written informed consent was obtained from all patients. The study was in accordance with the declaration of Helsinki. The present analysis is based on data from the recruitment visit comprising COPD-patients of categories GOLD 0-IV [18] (n = 2653; for basic characteristics see Table 1). Data collection was performed from September 2010 to December 2013. All patients were required to be diagnosed with COPD or chronic bronchitis and an age of at least 40 years was required. Only few exclusion criteria were defined (e.g. lung surgery with great volume reduction, exacerbations within 4 weeks prior to baseline visit) since the study aimed at capturing a broad spectrum of patients with COPD. Comorbidities were assessed using a predefined list comprising a total of 51 (combined) diseases, as well as free text information. Moreover, patients were asked to bring all medication (original packages) to their study visit. Patients’ previous medical records were not evaluated due to issues of data quality and completeness. We solely used self-reports of physician-diagnosed conditions for morbidities, with medication brought to the visit being the supplementary source to be compared with diagnoses. Two approaches (I and II, see description below) were developed for a quality comparison. Approach II aimed at overcoming some of the limitations inherent to Approach I.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Basic characteristics of the patients stratified according to COPD GOLD categories 0–4.

https://doi.org/10.1371/journal.pone.0163408.t001

Approach I for comparison of self-reported diagnoses

The first step was to transform comorbidities and medications into codes which could be matched by computer analysis. “Non-insulin dependent diabetes mellitus” and “insulin-dependent diabetes mellitus” were combined into “diabetes mellitus”, and reflux, gastritis and gastric and duodenal ulcerations into “gastro-intestinal disorders” (GI). In addition to the individual analyses, hypertension, coronary heart disease and heart failure were merged into a disease entity “combined cardiovascular disorder” to take account of the fact that there was a lack of medication specific for the single diseases, but specific medication for their combination is available. “Mental disorders” comprised depression, anxiety, panic attacks and psychotic disorders.

ICD10-Codes for comorbidities.

To classify the reported diagnoses the system of International Classification of Diseases, version 10 (ICD10 [19]) was employed. ICD10-Codes comprise up to 5 digits (structure: X00.00). From the patients’ reports only limited information was available necessitating the definition of lumped categories. Thus, comorbidities were coded by ICD10-Codes but in most cases without applying subcategories. Except for sleep apnea, disturbed blood flow in the legs, and vein thrombosis we used only the first three digits.

In some instances a combination of three-digit codes was required since the single codes were not sufficient to describe the comorbidities. This applied e.g. for the integrative category of “coronary heart disease” which comprised I20: Angina pectoris, I24: Other acute ischemic heart diseases, and I25: Chronic ischemic heart disease. Table 2 shows all comorbidities mentioned by more than 400 patients or considered relevant in COPD. The full table of categories can be found in the S1 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. ICD10 coding of relevant and/or prevalent comorbidities as used in the analysis of Approach I.

https://doi.org/10.1371/journal.pone.0163408.t002

Transformation of ICD10-Codes of medication for the purpose of matching.

Indications for each reported drug were retrieved from the AiDKlinik software program [20] using the PZN (pharmacological registration number in Germany) if available, or alternatively names, dosages and companies. In the next step a special dictionary was created to translate and re-code from the detailed 5-digit ICD10-Codes of the medication into single or multiple 3-digit-codes. This enabled the matching with the comorbidity codes. Table 3 illustrates the re-coding algorithm for selected comorbidities (for the full dictionary see S2 Table). Each code occurred only once for each patient even if more than one drug with the same indication had been prescribed. The conversion from detailed ICD10-Codes into the comprehensive, merged codes and their matching with comorbidity ICD10-Codes was performed in the programming language R 3.2.2 [21].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. List of ICD10-Codes used for comparison of comorbidities and medication.

https://doi.org/10.1371/journal.pone.0163408.t003

Matching of ICD10-Codes between comorbidities and medication.

The comparison led to three different classes. Either the intake of a medication was concordant with the reported disease (class 1), or a disease was reported without a matching medication (class 2). The possibility that there was a medication without a corresponding disease in the patient’s report was not evaluated as medications were not necessarily specific for a disease. The natural limitation of this approach results from the fact that medications were classified according to their indication in a given comorbidity but without requiring that this comorbidity was the only indication for this medication.

Comparison and Combination of Approaches I and II

The patients which could not be categorized into concordance scores A or B using specific medications could only be analyzed using the weaker information from Approach I. Approach II provides a sufficient condition for the presence of the disease, therefore patients can be considered with high confidence to have the respective comorbidity. In contrast in Approach I there was only medication compatible with the disease or no medication for the disease at all. The scores A and B from Approach II were kept in the combined approach. Patients with class 1 from Approach I (intake of a (nonspecific) drug plus report of a corresponding disease) were labeled as concordance score C, and patients with class 2 (report only) as concordance score D. The combined approach thus enabled the definition of four concordance scores which were computed for the comorbidities selected for analysis (see Tables 2 and 4). The matching procedures were again performed by special R-codes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Overview on the information used for defining the concordance scores.

https://doi.org/10.1371/journal.pone.0163408.t004

Results

Table 5 shows the distribution of scores for the 9 analyzed comorbidities and the combined cardiovascular disorder. The distribution varied between diseases and in particular the contributions of the scores A and D were very different. For selected diseases the pattern is illustrated in Fig 1 (for all diseases see S1 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Examples of different distribution patterns of the concordance scores for the diseases asthma, diabetes, hyperuricemia and GI disorders.

The values are percentages relative to the total number of patients (n = 2653). The blue part (A) represents the concordance between reported disease and specific medication, the red part (C) illustrates self-reports confirmed by non-specific medication. Green parts show the proportion of patients only reporting a disease without any suitable medication (D). The violet part (B) on top presents patients without the report of a disease but identified as likely having the disease due to the intake of a specific medication. The sum of A, C and D represents the prevalence according to self-reports (see Table 5). The distribution patterns vary widely among the different diseases.

https://doi.org/10.1371/journal.pone.0163408.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Distribution of concordance scores for different comorbidities (percentages based on the total number of included patients (n = 2653)).

https://doi.org/10.1371/journal.pone.0163408.t005

The proportion of patients in whom the comorbidity was in accordance with specific medication was highest for diabetes mellitus. Only few patients were identified as taking diabetes-specific medication without reported diabetes (2.4% of 371 patients). Similarly, the proportion of patients with reported diabetes without specific or non-specific medication was low (16.4% of 371 patients).

In contrast to diabetes, reported asthma was associated with specific medication only in a minority of patients. This was a consequence of the fact that there are very few drugs on the market which are specific for asthma and not at the same time applicable in COPD. This reduces the potential to confirm asthma by medication in a cohort in which patients were required to have COPD. The percentage of patients reporting asthma without any medication indicating asthma was very low.

An intermediate pattern was observed for hyperuricemia. In a considerable proportion of patients the self-reported diagnosis was in accordance with the reported intake of specific medication, and the proportion of non-specific medication was low. On the other hand 47.7% of patients reported to have been diagnosed with hyperuricemia but took neither specific nor non-specific medication.

The analysis of coronary heart disease and hypertension showed a very high proportion of patients in whom the self-reported diagnosis was in accordance to non-specific medication. In contrast the percentage identified by specific medication was low. This reflected the fact that few drugs are available which are specifically targeted against one of these two diseases. This becomes apparent in the relatively low number in category A. Only a minority of patients with coronary heart disease or hypertension could be identified by specific drugs in the absence of a self-reported diagnosis. In comparison to the single diseases, the level of accordance between diagnosis and specific medication was much higher for the combined entity comprising hypertension, coronary heart disease and heart failure. This shows that the combined entity allowed for a greater number of specific medications (see S12 Table).

The patterns observed in the other diseases were similar to those of the diseases described above. Overall, about 51.5% of the reported comorbidities were confirmed by Approach I, without very large differences between diseases although the number of comorbidities analyzed was much higher than in Approach II. Despite the lower number of comorbidities analyzed, in Approach II the percentage confirmed by specific medication (concordance score A) showed even larger differences between diseases (Table 5 and S1 Fig).

The strategy comprising the combined Approaches I and II was equivalent to the flow chart shown in Fig 2. The structure of this chart also illustrates why it was possible to add the percentages of scores A-D as shown in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Diagram showing the logical structure of the combined categorization procedure (Approach I plus Approach II).

A-D indicates the concordance scores, Ø the absence of the disease under study. ATC-Codes refer to the patients‘ medication, and ICD10 matching to the comparison of medication with the revised ICD10-Codes of the disease (for details see text).

https://doi.org/10.1371/journal.pone.0163408.g002

Discussion

In COSYCONET, as in many other studies, patients were asked to report physician-diagnosed comorbidities (“Have you ever been diagnosed with…”). This type of assessment is required since it is practically impossible to verify a broad spectrum of comorbidities by clinical assessments at the study visits. Obviously, the information depends on the patients’ knowledge, memory and cooperation. It is also based on the sensitivity and specificity of the assessment of comorbidities during usual care by the patients’ physicians. Therefore it would be helpful to increase the reliability of patients’ reports with the help of other data available in the registry. For this purpose we developed a categorization system comprising four concordance scores for the reported diseases by using the information on medication. The major outcome was that about 50% of comorbidities were in accordance with medication when using the non-specific Approach I. For a selected number of important comorbidities Approach II showed that the degree of confirmation differed very much between diseases, particularly diabetes and asthma.

Many studies on multimorbidity have relied on questionnaires asking for self-reported doctor-diagnosed diseases [4, 22]. Studies on the relationship between self-reported and other medical data [5–7, 9–11] indicated sufficient reliability, especially for highly prevalent chronic conditions [6–9, 12], however the relationship differed between diseases. Using medical records as a gold standard [7, 9, 23] the sensitivities for different diseases turned out to be >0.7 indicating that self-reports can be a valid way to capture comorbidities despite the finding that for depression/anxiety and peptic ulcer disease sensitivities were lower [23]. These findings were explained by the fact that patients’ reports depend on disease severity, symptoms, the characteristics of the comorbidity and the need for ongoing monitoring [10]. Similar results were shown in a study that did not use a gold standard [12]; the concordance between self-reports and other sources of information was high for easily memorable, clearly defined conditions such as diabetes or heart disease, while it was lower for depression and lung disorders. Higher prevalences for self-reports compared to medical records might be due to patients’ recall bias or incomplete records [7, 24] but underreporting is also possible for diseases with low impact on daily symptoms [6].

In view of these findings the primary aim of our study was to evaluate to which extent information derived from medication can be used in different comorbidities; for this purpose we developed and described in detail a method utilizing these data.

Coelho et al. [14] already addressed the confirmation of self-reports by using medication, but no detailed information on the method was given. Computerized tools like the RxRisk Model [25] that have been used to relate ambulatory pharmacy data to chronic conditions cannot be used in all countries, since the pharmacy management differs and data may not be available in the proper format. In general, it seems that taking into account several sources of information for comorbidities leads to the best description of the patients’ comorbidity status [7, 13]. There is one study in patients with COPD which used medication (diabetes) or biomarkers (hypertension) to assess the prevalence of clinically relevant comorbidities, but no details were given on the procedure [26]. Biomarkers pose the problem that they might be affected by treatment. We therefore developed and describe such a procedure based on medication. As a further novel contribution concordance scores were defined providing different levels of confidence into the self-reported information on comorbidities.

In Approach I more than half of the reported diseases were found to be compatible with the intake of medication. This was encouraging when considering that some chronic conditions do not require regular medication and some may only have been present in the past. Nonetheless, this approach was limited in its confirmatory ability (concordance score C), since it did not account for the specificity of medication and all possible indications for a drug were admitted, e.g. in glucocorticoids more than 120.

For this reason Approach II was developed which allowed the addition of two further concordance scores (A and B). The basic feature was the use of disease-specific medications. If the patient’s report and the medication matched, the highest concordance score A was assigned. In analogy to the RxRisk model score B was defined as probable disease indicated by specific medication but without reporting of the disease by the patient. Although score B is to be considered lower than score A, the specificity of medication is strongly suggestive of the comorbidity. The lowest level of credibility (concordance score D) was given, when neither non-specific nor specific medication was identified although a patient reported a comorbidity. This is difficult to interpret since the comorbidity may have been cured or may be mild and does not require medication. In this case only more specific information, the comparison with the patients’ clinical state and/or biomarkers is capable of giving further clues.

The identification of specific drugs for GI disorders was difficult since many GI problems are treated not for chronic but acute conditions. Moreover acute problems of the GI tract are often caused by other conditions (e.g. diarrhea due to viral infection; vomiting due to use of cytostatic drugs). Therefore drugs for these types of symptoms were excluded (e.g. loperamide, metoclopramide). Another problem was that Over-the-counter (OTC) medication is available for the treatment of chronic GI problems (e.g. omeprazole). A prescribed medication supports the assumption that a physician diagnosed the self-reported comorbidity, whereas in case of OTC medications the self-reported disease must not be based on a physician’s diagnosis. Moreover, pantoprazole or omeprazole are often prescribed to prevent GI irritations caused by other drugs. This reduces the ability to infer a comorbidity from specific medication in this particular case.

The differentiation between asthma and other obstructive lung diseases via medication is a difficult issue. In COSYCONET this posed a particular problem due to the fact that all COSYCONET-patients were required to have a diagnosis of COPD. Many COPD medications are also prescribed in asthma, such as SABAs, LABAs or inhaled corticosteroids, and there are very few medications for asthma only. Therefore the opportunity for evaluation of asthma was low. Possibly a combined diagnosis “obstructive airway disorder” comprising asthma and COPD would result in a high degree of concordance.

Hypertension was the most often reported comorbidity in COSYCONET. Notably, many patients with hypertension are not categorized into the highest concordance score A since their hypertension therapeutics have many further indications. Coronary heart disease is also often treated by non-specific medications. We therefore combined coronary heart disease, hypertension and heart failure into a complex entity and used medication specific for one or more of these diseases to define concordance scores. This resulted in a high proportion of patients in concordance score A in comparison to the single diseases. In addition to the pharmaceutical argument given in the Methods section, we included the analysis of the combined group in order to illustrate that even with few, clinically heterogeneous diseases a remarkable gain in concordance score could be achieved, provided that there is medication specific for up to three of these diseases.

It will depend on the aim of the analyses regarding these comorbidities, whether a high degree of confirmation for a broader entity or a separation into single diseases is better. Due to this we cannot give a general recommendation how to use the scores. Self-reports are often considered sufficient in large epidemiological studies assessing prevalence estimates but score B may improve the estimates, for diseases on which most medication is specific.

In addition a detailed analysis of medication might be helpful for comparisons between patients merely reporting the disease versus those adequately treated. Information on specific medication might also be useful if this medication has pleiotropic effects extending beyond their specific indication [27, 28]. In contrast, patients with score D could be considered with caution in the analysis. This applies e.g. to the cardiovascular disorders which commonly require regular treatment. Overreporting as possibly reflected in score D might be based on patients’ misinterpretations of medical terminology [29]. We propose that by comparing different medication-based definitions of the disease it should be analyzed, whether exclusion of patients with score D creates more reliable groups. In addition, we expect that such comparisons could also be valuable in the analysis of quality of life data which might depend on the intake of medication in relation to disease severity. In our view, the ICD10-Code-based approach is useful for the quality check of data in studies relying on self-reports, since at least it can indicate compatibility between medication and report. These considerations illustrate that the different approaches and categories developed by us offer several options that can be chosen and adapted according to specific diseases and needs.

Supporting Information