Cause of death coding in Switzerland: evaluation based on a nationwide individual linkage of mortality and hospital in-patient records

Cause of death statistics are an important tool for monitoring the health of populations and for responding effectively to changing epidemiological circumstances [1, 2]. They are also a tool for quality control of the health care system. For example deaths due to causes that should not be fatal in the presence of effective medical care, known as amenable mortality, is an indicator of national levels of personal health-care access and quality [3]. Because of the long-standing supervision by WHO the statistics are, in principle, comparable over time and between countries [4]. Their reliability, however, is controversial [5‐7]. Substantial variation in certification practices between countries is a known problem [8, 9]. Even within the same country, “different coding practices, socio-cultural milieus, and individual socio-demographic characteristics such as age are most likely to influence the cause of death assigned on the death certificate” [10].

For these reasons it is important to evaluate the reliability of cause-of-death assignment and coding. Usually validation consists in obtaining the coding of the same set of death certificates from different raters [1, 7, 10, 11]. Unfortunately an agreement between the original and the reviewed data may us tell more about reproducibility and less aboutaccuracy of the information [12]. The forms look correct but do not provide an accurate description of the case [13]: Less frequently, death certificates were compared with their corresponding medical records (e.g., [14, 15]), a procedure rated as too slow and expensive for routine use [13] and therefore implemented only occasionally and studies spanning short periods [15‐17]. One of these studies, however, confirmed that medical incompatibility between underlying cause of death and main hospital discharge condition is a marker for greater risk of erroneous cause-of-death certification [16].

Switzerland introduced cause of death statistics in 1876. Certifying by a physician became mandatory a few years later [18]. In the assessments by WHO, the quality of Swiss data was rated as medium [19] or medium-high [5], due to a high proportion of ill-defined causes. This is in line with a recent evaluation by the Global Burden of Disease 2016 Causes of Death Collaborators, which concluded that in Switzerland since 2000 a larger proportion of “well certified deaths” than in neighboring Germany and France (but still not reaching the proportions in neighboring Austria and Italy) [20]. The only evaluation studies comparing cause of death statistics with other medical information date from the 1980s [21, 22]. More recent validation studies were restricted to death certificates and the implications of the adoption of ICD-10 in 1995, trying to identify correction factors with which to correct time series [23, 24].

In Swiss hospitals, cause of death reporting to the Federal Statistical Office is generally in the charge of the responsible attending physician, who also completes the medical files which in turn serve as the basis for the standardized registration of diseases and treatments in the hospital discharge statistics. It is therefore compelling to analyze the consistency of ICD mapping on death certificates and in hospital discharge statistics. Agreement of both, however, is still more a marker for reproducibility than validity of the information. Nevertheless, incompatibility of diagnoses and causes of death may point to quality problems [14].

In contrast to others we did not target an overall quality measure for hospital and cause of death statistics in Switzerland (e.g., the proportion of ill-defined causes) but aimed at assessing the reliability of as many as possible specified diagnoses / causes of death. Within this frame, we first aim at exploring to what extent the underlying cause of death in the cause of death statistics is in agreement with principal and additional diagnoses registered at hospital discharge in the Medical Statistics of Swiss Hospitals (MS). Second, we aim at exploring the opposite, i.e., the agreement of principal diagnosis at hospital discharge with cause of death information on the death certificate. Which causes of death can be traced in the MS especially well or badly and which principal hospital diagnoses can be traced in the cause of death statistics especially well or poorly? In addition, it may be of interest to know which diseases have the most similarity or variation of agreement rates between the two approaches. Finally, we aimed at evaluating combined vs. separate disease categories where the related ICD-codes may not always be easily distinguishable, and at looking for sociodemographic determinants that could influence the agreement between hospital and cause of death data.

For all analyses comparing hospital discharge diagnoses with cause of death information, one has, however, to be aware that the definitions of principal hospital diagnosis and underlying cause of death do not necessarily agree. Cause of death statistics have to deliver a unicausal result, also for multimorbid subjects for whom this concept is not appropriate. On the other hand, medical coding in hospitals first of all serves reimbursement and this may result in distortions [12]. The principal hospital diagnosis may also be a complication of the underlying cause of death. For these reasons, taking into account additional diagnoses as well as concomitant causes of death will be essential.

In Switzerland, on an individual level cause of death information in official cause of death statistics and diagnoses in the hospital discharge statistics are generally compatible, especially if additional diagnoses and contributory causes are taken into account. In 83% of analyzed deaths, the underlying cause of death could be traced in one of the diagnoses of the terminal hospital discharge and in 77% the principal discharge diagnosis was also reported as primary or contributory cause of death (principal diagnosis = primary cause: 57%). This is fairly in line with Johansson&Westerling who found agreement of principal hospital condition with underlying cause of death of 59% and with any information on the death certificate of 83% [13]. The latter number is somewhat lower than the 89% reported in two studies based on automated coding of diseases and causes of death [14, 17] reducing variation in the evaluation of complex situations but not necessarily resulting in higher validity of agreement [12].

The proportion of agreement varied widely depending on disease/cause of death, but was generally in line with known patterns from similar studies [13, 15, 22] as well as from intercoder agreement studies (e.g., [1, 11]), i.e., excellent in cancers and cerebrovascular disease and very good in respiratory and liver diseases. Discordance in less clearly defined chronic diseases may therefore be due to a large part to increased difficulty in reaching a consensus on diagnosis.

Of note, for individuals with long-lasting chronic diseases like multiple sclerosis, the immediate reason for a hospitalization is often due rather to sequelae which consequently appear as the principal hospital diagnosis. As studies from Sweden suggest, a substantial part of individual cases with discrepancies may be attributable to incomplete or inadequate transformation of the diagnostic information in the patient charts into disease assignment [31] or mistakes in death certification [16]. The fact that the logic and structure of the ICD differs from the clinical way of thinking [12] may also play a role, as well as the well-known difficulty to define a single, disease-specific, underlying cause of death in older people [6].

The proportion of agreement that was found between the underlying cause of death and hospital diagnoses was very similar to that found by Minder&Zingg in their 1979 sample of deaths in Switzerland [22]. The only notable differences were somewhat higher agreement rates in our study for malignant neoplasms of colon, rectum and larynx, while for leukaemia and diabetes mellitus in the older study the agreement was closer.

Substantial concordance with the Minder&Zingg study also emerged for the proportions of agreement between principal hospital diagnosis and combined primary and contributory cause of death information. The figures from our study show a closer agreement for several cancers (such as larynx, oropharynx and colon) and, most clearly, for chronic lower respiratory disease, however again less agreement in the case of diabetes mellitus. At least for cancers the proportions of agreement were also very similar as in a large Swedish study based on 1995 data [13]. This high degree of similarity of patterns also supports the notion that different evaluation of complex situations (see e.g. [32]) rather than insufficient reporting is the main reason for the variance between hospital discharge and cause of death registration.

In line with others [1, 6] disagreement increased with age. Our data support the notion that this may be related to an increasing number of reported diagnoses/causes [1].

Our data do not suggest substantial variation in the individual agreement between male and female decedents, which is in line with Alpérovitch et al. [6] who however could establish associations with determinants not available in our study (history of vascular diseases, presence of incapacities, MMSE <24, >5 medications).

The odds ratios resulting from multiple logistic regression (Table 3) however suggest substantial variation in reporting and maybe in diagnostic practice that can not entirely be explained by age and cause of death of the deceased. Comparatively lower agreement for the less educated and those speaking a non-European language could point to a lower standard of reporting (and maybe also examination) among less advantaged patients. These findings and the substantial regional variation indicate potential for improving procedures in cause of death as well as hospital discharge reporting.

What measures could improve the quality of the MS and the cause of death statistics in Switzerland? First, the responsible authority and data owner (Swiss Federal Statistical Office, SFSO) could do much more than at present. Currently, the SFSO publishes data reports on a yearly basis and researchers can request access to individual data, but there are no published reports about data quality. Second, the regular use of data automatically helps to improve data quality, because problems become visible and can be overcome. Especially studies combining information from different data sources may promote the evaluation of corresponding strengths and limits [33]. Third, a case review based on full clinical documentation should be carried out for the areas with lower scores in this study. For example, a sample of deaths due to diabetes could be studied in detail. The example of cancer shows that the work of cancer registries has an impact on medical documentation. An individual comparison of the cause of death statistics with the data of the cancer registries is required by law from 2020. Finally, the instruction in the scope and methods of the cause of death statistics during the basic and continuing education of physicians should be improved.

Depending on cause of death / hospital diagnosis, the patterns of agreement vary widely, but do not substantially differ from those found in other studies. This also and notably applies to interrater reliability studies as well as studies comparing individual data from different data sources, supporting the view that a substantial proportion of variation is due to different evaluation of complex situations by individual physicians.

The agreement rates in this study also did only slightly differ from those found almost 30 years ago, in spite of a necropsy study suggesting a significant decline of major discrepancies between 1972 and 1992 [35].

Agreement at the individual level remains limited and suggests a potential for improving data quality [14]. Even if this should not affect the reliability of population frequencies [10], the substantial regional variation (with lower agreement rates in French-speaking populations) hints at differential exploitation of available diagnosis information for cause of death statistics. Potentially worrying is the evidence for lower retrieval performance of the underlying cause of death in hospital diagnoses among socially disadvantaged groups like single, less educated, or culturally less integrated persons.

For all these reasons, studies of this kind should be regularly conducted as a quality monitoring of hospital diagnoses and causes of death.