Identifying patients with chronic conditions using pharmacy data in Switzerland: an updated mapping approach to the classification of medications

The evaluation of the population health status, patients’ health care needs and its associated costs is a priority issue in health policy, decision-making and resource allocation debates. In general, data of national disease registries and prevalence surveys, including clinical diagnoses according to the International Classification of Diseases (ICD-10-codes) have regularly been used to estimate the health status of a population. However, this type of data pool is not available in all health care systems. In Switzerland, for example, epidemiological data, providing information on the prevalence of chronic diseases and comorbidities in general population, are not widely available.

Administrative databases such as drug prescription data have thus been frequently used to identify persons with chronic conditions, quasi as an indirect method to estimate prevalence. Pharmacy based claims data provide a consistently available information source, which is reliable, covers a large population and might be extremely useful for assessment of morbidity [1–6]. Pharmacy-based diagnosis were used in risk adjustment models [7–9], illness severity measurement [10, 11], prevalence estimates [12–15] and epidemiological studies for comorbidity adjustments [16, 17]. However, in these studies the clustering of the Anatomical Therapeutic Chemical (ATC)-codes has not been applied consistently, and even in few studies the used ATCs are not documented. Moreover, the well-known ATC-algorithm of Lamers/van Vliet [18–20], the Pharmacy-based Cost Group (PCG) model, was often used in various modified and unspecified defined versions. The PCG model distinguishes 22 chronic conditions and was primarily developed to predict cost of diseases for risk adjustment. However, this model has some limitations. Medication classifications based on data, which have been recorded about 10 years ago. New drugs, which had not been developed and thus were not commercially available in the past years, were not included in the model. Furthermore, previous studies claimed the possibility of an exact differentiation between specific diseases via ATC code [12, 19–21]. For example beta-blockers and diuretics were assigned to the category “hypertension” [19, 22]. However, beta-blockers were also prescribed in patients with other cardiovascular diseases. Another example, diuretics were included in the category “cardiovascular diseases” although diuretics were also frequently used in patients with renal diseases [12, 21]. In several medication classes, an ambiguous assignment of medication to chronic conditions is challenging and even, in certain cases, infeasible.

To overcome the limitations of previous mapping approaches, and to suggest a standardised and transparent use of the mapped medication classes to chronic conditions, we aimed to develop an updated mapping algorithm with a special focus on the unambiguous assignment of prescription drugs to chronic diseases. We provide an updated and rather conservative mapping approach to the classification of medications. Our classification is on the one side detailed as possible and on the other side we summarise categories to the superior category of disease when needed. Furthermore, we give an overview of the proportions of chronically ill patients in Switzerland using pharmacy data.

This study developed an updated mapping approach to identify chronically ill patients using pharmacy data. Databases on prescribed drugs are an useful source of information on morbidity, if reliable medical diagnoses are missing.

In our classification model, we identified a total of 22 chronic conditions based on the 2011 WHO-ATC-classification. We found that the largest proportions of persons suffered from pain, rheumatologic conditions and cardiovascular diseases, including hypertension. These findings are in line with previous studies in Italy estimating prevalence of chronic diseases showing highest rates in cardiovascular diseases and rheumatologic conditions [12, 13]. On the other hand, our results are different from the Dutch estimates based on the PCG-model showing a low prevalence rate in rheumatologic conditions [19]. One explanation is the fact that a different and partly out-dated medication classification for the chronic conditions was used. New drugs, which have been developed during the last decade, were not included in the PCG-model performed in the nineties. We assume that the Italian results may be more exact, since a more recent medication classification was used in comparison to the Dutch study. Several of our estimated proportions of patients with chronic conditions are comparable to prevalence estimates in other European countries. For example, the prevalence of diabetes mellitus, as one of the leading health problems in the industrialised countries, is similar to the prevalence in Germany and the Netherlands [15, 36, 37]. Additionally, the prevalence rates of asthma are comparable to those reported in other European countries such as France, Norway and Germany [38]. Not surprisingly and in line with several studies, we could show that chronic conditions affected all ages, but with a significant increase in older age groups [39–42]. Furthermore, we found that cardiovascular diseases are mainly attributable to classical risk factors such as diabetes and hyperlipidemia. A large body of epidemiological and pathological data also demonstrated that diabetes and dyslipidemia are independent risk factors for cardiovascular diseases in both men and women. Diabetes and hyperlipidemia are common comorbidities of cardiovascular diseases and contributed significantly to its overall mortality [43–47].

A further aim of our study was to compare our estimated proportions of chronically ill patients with the reliable prevalence estimates from national studies. Therefore, we compared our estimated proportions of patients with cardiovascular diseases, diabetes mellitus, hyperlipidemia, cancer and respiratory illness (asthma, COPD) to those proportions from available data sources in Switzerland. Data on further chronic conditions could not be presented, since epidemiological data in Switzerland are very scarce. We found a high level of agreement between the estimates using pharmacy data and those reported by epidemiological studies. Proportions of cancer, cardiovascular diseases, diabetes and respiratory illness matched quite well [25, 28, 31, 34]. For hyperlipidemia, study results showed lower prevalence estimates than those revealed in our pharmacy based study [27, 33]. One explanation for the discrepancy between our findings and those from Pèchere-Bertschi et al. [27] and Firmann et al. [33] is the questionable representativeness for the general Swiss population. For example, Firmann et al. [33] collected data from a sample of 6′084 subjects living in one Swiss city, Lausanne, with a relatively high percentage of non-Swiss compared to other cities. Moreover, this study was conducted six years ago and recent trends in prevalence are not taken into account.

There are some limitations addressing the accuracy of our estimates. About 3% of the used claims invoices were paid directly by the patients and not by the health insurer. Drug data may be underrepresented since claims data are not completely recorded. Furthermore, the number of prescribed drugs was underreported because they were obtained exclusively in outpatient settings. Additionally, drug-based diagnoses are naturally regarded as a proxy for a medical diagnosis. Since we excluded ATC-codes which could not be matched to certain diseases, proportions may be underestimated. Moreover, we are not able to distinguish specific chronic diseases such as hypertension from other cardiovascular diseases. For legal and ethical reasons as well as regarding data security, data do not include medical diagnoses and clinical procedures, and thus we are unable to give more information on chronic conditions as well as to provide prevalence estimates based on medical diagnoses (e.g. ICD-10) for comparisons in the same data source. Furthermore, in Switzerland reliable epidemiological data are scare. Appropriate additional, independent samples for comparing the prevalence estimates are not widely available. However, drug-based diagnoses could not completely substitute, but add clinical data. Mapping medication classes to chronic conditions is a frequently recommended and used in previous studies [10, 12, 21]. Moreover, we assume that no patient is prescribed a drug for chronic conditions such as insulin without a serious indication of diabetes. However, this study has several strengths. Our study is based on a very comprehensive administrative claims database for a large population across Switzerland with an almost complete recording of prescribed drug data. Administrative data are a consistently available source of information on morbidity, health care use and costs of care. They are reliable, longitudinal, practice-based, large in size and widely accepted in epidemiological, health services and outcomes research to assess prevalence. Moreover, studies developing risk models emphasised that prescription data derived from administrative claims database have several advantages over diagnostic data regarding reliability, timeliness and completeness of data [7, 11, 48]. In the Dutch health care system, the use of information on chronic conditions obtained from claims for prescribed drugs are well established as a helpful instrument for improving risk equalisation between the sickness funds as well as in capitation modelling [19, 20, 49]. In Switzerland, there is an on-going debate about the optimal risk equalisation scheme [50–53]. The present Swiss risk equalisation scheme is a demographics-based risk equalisation model which based on age, sex and an at least three successive night-stays in a hospital or nursing home (in the previous year) as equalising elements. In 2011, the Swiss parliament has decided to refine the risk equalisation by diagnoses based on pharmacy data, since no appropriate data on diagnosed diseases exist. Consequently, it is particularly important to identify those ATC-codes which are exclusively used for the treatment of population’s common diseases. Pharmacy-based, estimated prevalence data might be a useful addition for risk adjustment modelling in health care delivery systems with demographics-based risk equalisation schemes.

Using pharmacy data, we developed an updated and improved approach for a feasible and efficient measure of patients’ chronic disease status. Pharmacy drug data may be a valuable source for measuring population’s burden of disease, when clinical data are missing. This approach may contribute to health policy debates about health services sources and risk adjustment modelling.