The Danish National Patient Registry: a review of content, data quality, and research potential.

The Danish Health and Medicines Authority has established guidelines for releasing data from the DNPR. Implementing the European Union Data Protection Directive (Directive 95/46/EC) on the protection of individuals with regard to the processing of personal data and on the free movement of such data, the Danish Act on Processing of Personal Data provides the legal basis for private and public institutions to obtain individually identifiable health data for research purposes. 217 This Act protects against abuse of such data and thus balances the privacy rights of individuals and the society’s need for quality research. In order to access data from the DNPR, researchers have to apply to Research Service (Danish, Forskerservice). 26 , 218 Use of any health data also requires project-specific permission from the Danish Data Protection Agency, 217 and, in many cases, additional permission from the Danish Health and Medicines Authority to link data from various registries. 26 The Danish Data Protection Agency specifies safety precautions for data processing and also sets cancellation deadlines, ensuring that data traceable to individuals will not be stored longer than required to complete a project. As well, it is necessary to obtain permission from the Danish Health and Medicines Authority and the chief physician from relevant hospital departments to retrieve medical record files for validation of DNPR data. 219

Potential uses of the DNPR, according to study design, are presented in Table 7 . Patient cohorts of interest may be identified, along with their medical history and outcomes. Thus, the DNPR may provide data on diseases, 170 , 171 treatments, 172 and diagnostic examinations as exposures. Seasonal variation as an exposure has also been examined. 173 Furthermore, the DNPR allows for identification of disease occurrence in the general population (risk studies), 174 where the exposure information could originate from other data sources involving primary or secondary data collection, eg, military conscription cohorts 175 or population-based health surveys such as the Danish Health Examination Survey, 176 the “How Are You?” study, 177 the Danish Diet, Cancer and Health study, 178 the Soon Parents cohort, 179 the Glostrup Population Studies, 180 and the Copenhagen City Heart study. 181 Extraordinary long-term follow-up (>35 years) for lifestyle-associated diseases is feasible. 175 Using techniques similar to that in risk studies, the DNPR can be used to study outcomes in well-defined patient groups (eg, diagnostic examinations, 182 recurrence, 112 and complications 183 ) and prognostic factors. 170 These patient groups may be identified from the DNPR itself, other registries, or surveys. Most recently, the DNPR has also been used to gather long-term follow-up data for randomized controlled trials using clinically driven outcome detection. 184 The automated event-detection feature of the DNPR allows for large, low-cost randomized trials that reflect daily clinical practice, cover a broad range of patients and end points, and include lifelong follow-up. 183 , 185 – 187 As with cohort studies, DNPR data may be used to identify exposures and cases/outcomes in case– control studies 112 , 188 , 189 and ecological studies 182 ( Table 7 ). The administrative data related to each patient contact allow for studies of health care utilization and how health care planning may affect patient outcomes. As an example, admission rates for the most common medical conditions in Denmark have been found to be higher during the regular office hours than during the weekend hours. 190 However, admissions during the weekend hours have been associated with higher mortality rates (weekend nighttime hours > weekend daytime hours > weekday out-of-hours > weekday office hours). 190 The availability of patient-identifiable data in the DNPR makes it technically easy to link to other Danish data sources using the CPR number. 20 Because Denmark’s registries are numerous and far reaching even by the high standards of the Nordic countries, 22 , 191 additional information on, eg, cancer staging, 34 laboratory test results, 167 general practice utilization, 192 socioeconomic data, 193 – 196 prescription use, 197 all-cause mortality, 20 and cause-specific mortality 198 can easily be obtained to supplement the DNPR. Figure 3 shows the time line for the DNPR relative to selected administrative and clinical registries in Denmark, illustrating the potential for record linkage by calendar year. As shown, nationwide data can be obtained on, eg, all twins in Denmark since 1870 (the Danish Twin Registry), 199 specific causes of death since 1943 (the Danish Register of Causes of Death), 198 detailed cancer diagnoses since 1943 (the Danish Cancer Registry), 34 migration and vital status since 1968 (the Danish Civil Registration System), 20 personal income since 1970 (the Income Statistics Registry), 195 labor market statistics and health services since 1980 (the Integrated Database for Labour Market Research 193 and Danish National Health Service Register), 192 education since 1981 (the Population’s Education Register), 194 prescribed medications since 1995 (the Danish National Prescription Registry), 168 and patient tissue samples and blood transfusions since 1997 (the Danish National Pathology Registry and Blood Transfusion Databases). 166 The Danish clinical registries constitute the infrastructure of the National Clinical Quality Databases and the Danish Multidisciplinary Cancer Groups. 200 The clinical registries contain information about individual patients used for quality improvement, research, and surveillance purposes. 200 Linkage to one or more of the current 69 clinical registries thus provides detailed information on a range of procedures (eg, hip arthroplasty and hysterectomy) and diseases (eg, heart failure, stroke, diabetes, and various malignant diseases; Figure 3 ). 200 , 201 Finally, individual-level linkage to data from randomized controlled trials, population surveys, and epidemiologic field studies is possible as previously described.

As the role of routine computerized health data in epidemiological research is growing, 1 there is a need to examine their strengths and limitations. 2 , 3 Typical shortcomings of such data include limited linkage possibilities, incomplete temporal or geographic coverage, restriction to selected patient groups, and lack of systematic follow-up. 4 – 7 Among the examples, the Dutch nationwide hospital registry has been in operation since 1963, but personal records are anonymized, and therefore not linkable to other data sources. 4 Also, the United Kingdom’s Clinical Practice Research Datalink has recorded detailed information on both diagnoses and prescriptions in primary care since 1987 but covers only part of the population and lacks information on patients who leave participating practices. 8 In the United States, the collection of routine health data is restricted to specific age groups (eg, Medicare beneficiaries), 6 income groups (eg, Medicaid beneficiaries), 6 professions (eg, the Veterans Affairs), 7 or members of private insurance plans (eg, Kaiser Permanente), 9 often without the possibility of linkage or long-term follow-up. In the Nordic countries, government-funded universal health care, combined with the tradition of record-keeping and individual-level linkage, has led to establishment of extensive networks of interlinkable longitudinal population-based registries covering entire nations. 10 , 11 Patient registries with complete nationwide coverage and individual-level linkage potential have existed in Finland since 1969, 12 in Sweden since 1987, 13 in Iceland since 1999, 14 and in Norway since 2008. 15 , 16 The Danish National Patient Registry (DNPR) is one such population-based administrative registry, which has collected data from all Danish hospitals since 1977 with complete nationwide coverage since 1978. 17 – 19 An epidemiologist setting out to use the DNPR must be familiar with the strengths and limitations of its data. Many studies have validated algorithms for identifying health events in the DNPR, but the reports are fragmented and no overview exists. Herein, we review the content and data quality of the DNPR and its potential as a research tool in epidemiology.

Denmark had 5,580,516 inhabitants in 2012, excluding inhabitants of Greenland and the Faroe Islands. 20 Although these areas are part of the Kingdom of Denmark, they are not covered by the DNPR. Since 2007, the Danish healthcare system has had three administrative levels: 10 , 21 1) the state, responsible for legislation, national guidelines, surveillance, and health financing through the Ministry of Health; 2) the regions (n=5), responsible for delivery of primary and hospital-based care; and 3) the municipalities (n=98), responsible for a broad range of welfare services, including school health, child dental treatment, home care, primary disease prevention, and rehabilitation. The Danish National Health Service provides tax-supported health care for the entire Danish population. 10 , 21 Redistributionist taxation finances ~85% of overall health care costs, including access to general practitioners (GPs), hospitals, outpatient specialty clinics, and partial reimbursement of prescribed medications. 21 Of note, outpatient specialty clinics include contacts from hospital-based (ambulatory) specialty clinics but not from private practice specialists or GPs. Patients’ out-of-pocket expenditures cover the remaining costs of medication and dental care. 21 Except in emergencies, GPs (including on-call GPs) provide referrals to hospitals and specialists. 21 Approximately 4,100 GPs and 4,600 dentists, as well as physiotherapists, chiropractors, and home nurses, work in the primary health care sector. 21 The Danish Civil Registration System is a key tool for epidemiological research in Denmark. 20 , 22 This nationwide registry of administrative information was established on April 2, 1968. 20 It assigns a unique ten-digit Civil Personal Register (CPR) number to all persons residing in Denmark, allowing for technically easy, cost-effective, and exact individual-level record linkage of all Danish registries. 20 The Danish Civil Registration System, which tracks and continuously updates information on migrations and vital status, permits long-term follow-up with accurate censoring at emigration or death. 20

The DNPR is a valuable tool for epidemiological research, providing longitudinal registration of diagnoses, treatments, and examinations, with complete nationwide coverage since 1978. Denmark’s constellation of universal health care, routine and long-standing registration of life and health events, and the possibility of exact individual-level linkage impart virtually unlimited research possibilities onto the DNPR. At the same time, varying completeness and validity of the individual variables underscore the need for validation of its clinical data before using the registry for research.

The classifications used in the DNPR are provided in the Health Care Classification System (Danish, Sundheds-væsenets Klassifikations System [SKS]). 43 The SKS is a collection of international, Nordic, and Danish classifications. 43 SKS codes contain up to ten alphanumeric characters, the first being a letter representing a primary group, following a monohierarchical classification system. 43 Thus, diagnoses are registered under “D”, surgery under “K”, other treatments under “B”, anesthesia under “N”, and examinations under “U” or “ZZ” ( Table 2 ). 36 To facilitate the search for SKS codes, the Danish National Health and Medicines Authority maintains a user-friendly SKS browser ( Figure 2 ), 44 searchable by code, by free text, or by browsing. Searching for acute myocardial infarction codes can be done by entering “DI21” or by typing the Danish or Latin term in a full phrase (akut myokardieinfarkt) or a partial phrase (eg, infarctus myo). 44 Manual browsing requires clicking the main group “Classification of diseases” (group D), then “Diseases of the cardiovascular system” (I), then “Ischemic heart disease” (I20–I25), and finally “Acute myocardial infarction” (I21). The SKS browser does not include historical codes, 44 but these are available online elsewhere. 45 Over time, the DNPR has adopted different classification systems for diagnoses, surgeries, and accidents ( Figure 1 ), whereas the classification systems for radiological procedures and in-hospital medications have remained unchanged since their introduction into the DNPR. 26 , 36 Diagnoses were classified according to the ICD-8 until the end of 1993 and the ICD-10 thereafter. The three-digit ICD-8 codes were used in a modified Danish version (with two supplementary digits), which explains in part why ICD-9 coding was never introduced in Denmark. Coding granularity improved in 1994 through introduction of the five-digit ICD-10 codes. Although the DNPR follows the current international standards for disease classification, the ICD-10 version used in Denmark often does not allow for identification of certain clinical details, such as disease severity. Supplementary codes (eg, the so-called “TUL” codes) sometimes allow for anatomical precision, eg, to identify location of a thrombosis or surgery site in right/left or upper/lower extremity, but these codes are used inconsistently. Sometimes, ABC extensions are added to specific diagnostic codes, eg, atrial fibrillation (I489B) and flutter (I489A), making the Danish version of the ICD-10 more detailed than the international ICD-10 but less detailed than the clinical modification of the ICD-10 (ICD-10-CM), which is not used in Denmark. 46 Surgeries were coded according to the three consecutive editions of the Danish Classification of Surgical Procedures and Therapies, from 1977 to 1995. 47 Since 1996, surgical procedures have been coded according to the Danish version of the Nordic Medico-Statistical Committee Classification of Surgical Procedures. 48 Accidents have been coded using the Danish Classification of Accidents. A detailed registration was introduced in 1987. The latest version of the classification, the Nordic Classification of External Causes of Injury, also included suicide attempts and violence. 26 It was adopted in 2008 and used until a new Danish Classification of External Causes of Injury was incorporated in the SKS, in 2014. 36 , 37 Although closely related to the Nordic classification in structure, the new Danish classification facilitates a simpler registration of external causes of injury. Radiological procedures (without results) are coded according to the Danish Classification of Radiological Procedures (UX codes). This classification system follows the general principles used for registration of treatments in the SKS. 36 In-hospital medication use (without dispensed dose or route of administration) is registered using different modules consistent with the Anatomical Therapeutic Chemical (ATC) classification system. Data on in-hospital medical treatment are not commonly used in research, except for drugs exclusively administered at hospitals, eg, fibrinolysis or cancer/immune-modulating treatments such as antibody, radiation, cytostatic, and biological therapies ( Table 2 ). These drugs are primarily registered with a SKS treatment code, but their ATC codes can also be used as supplemental codes (eg, fibrinolysis is covered by SKS code BOHA1 and ATC code B01AD).

Methodological considerations related to the internal validity of cohort studies conducted within the DNPR are summarized subsequently and in Table 8 . We also address the special methodological problems that relate to studies of temporal health trends. The nationwide coverage since 1978 provides sample sizes that permit studies of rare diseases, disease complications, and effects in subgroups of patients (effect modification and interactions). Of note, very rare diseases may still be difficult to study because of the relatively small size of the Danish population. 202 Appropriate population-based study designs can reduce selection biases in cohort studies for three reasons. First, the Danish population has a relatively stable and homogeneous demography with regard to race and religion. Second, the universal health care system (and small private hospital sector) 40 prevents selection bias arising from selective inclusion of specific hospitals, health insurance systems, income levels, or age groups. Third, virtually complete follow-up of all patients (with no unrecorded dropouts) is possible because the Danish Civil Registration System records vital status and migrations on a daily basis. 20 Still, the cohort represented in the DNPR is only unselected for diseases that always require hospital treatment. For diseases that can be treated in general practice, cases included in the DNPR to some degree represent a selected patient group, with either high severity of the disease in question (eg, herpes zoster infections, obesity, diabetes, and hypertension) or severe comorbidity leading to a lower threshold for hospital admission compared with patients without comorbidity (eg, pneumonia in transplant patients vs in young otherwise healthy adults). Although it is obvious that registration and retrieval of patient information from the DNPR must be based on correct SKS codes, this task is not always easy. The SKS includes many codes that might not be mutually exclusive from a clinical point of view. For many diagnoses, it is thus necessary to be aware of potential differences in registration practice among hospital departments 24 and over time. 122 , 170 , 203 Before engaging in extensive retrieval and analysis of data, it is therefore important to consult clinicians from the relevant specialty to learn about current and previous coding practices. As an example, atrial fibrillation and atrial flutter have separate codes at the four-digit level. However, a large proportion of all diagnoses for atrial fibrillation or flutter are registered as “not elsewhere specified” (Danish, uden nærmere specifikation). Since ~95% of all I48 codes correspond to atrial fibrillation and only 5% to atrial flutter, 104 use of the unspecified code will increase the sensitivity of the DNPR-based definition of atrial fibrillation but reduce its specificity. Hence, DNPR studies on risk of atrial fibrillation are often limited by considering atrial fibrillation and flutter as one disease entity. 174 Another example is ICD-10 diagnoses of stroke (I60–I64). Approximately one-third of the cases are registered as unspecified stroke (I64), 204 and among these, two-thirds are ischemic strokes. 91 Inclusion of unspecified diagnoses will increase sensitivity but reduce specificity of stroke subtypes. The introduction of the Diagnosis-Related Group system in 2002 29 , 30 regarding payment to public hospitals may have resulted in more complete registration. However, it may also have affected coding practices for some diseases and certain types of treatments. Private hospitals and clinics are potential sources of underreporting. 40 Although it has been mandatory for private health care providers to report all activities since 2003, and the Danish Health and Medicines Authority runs information campaigns to promote registration, 38 registration from private hospitals and clinics remains incomplete. 17 , 41 Private hospitals offer services paid by taxes due to the rules of “free hospital choice” or as part of an agreement with a region, as well as services paid privately either by insurance companies or private parties. 21 , 40 Services paid for by private parties have the highest degree of incomplete registration. In contrast to validity, the completeness of diagnoses is often higher in the DNPR than in the clinical registries. 89 , 100 , 164 , 205 , 206 This higher completeness is expected since many clinical registries receive data from the DNPR. Another reason is that the law requires the national clinical registries to cover only 90% of patients with a given condition. 207 Moreover, the degree of completeness varies among and within clinical registries over time. 164 , 208 Nonrandomized studies are susceptible to confounding by known and unknown factors. 209 Therefore – irrespective of data source – the potential for confounding always needs to be addressed in the study design or analysis. The DNPR provides an opportunity to obtain information on many potential confounders, particularly comorbidities. 58 , 210 The possibility of identifying such covariables from patients’ history of hospital encounters (back to 1977) rather than short-fixed historical windows may also result in less biased estimates. 211 Still, it should be kept in mind that incomplete registration of some diagnoses and missing data on other characteristics (eg, lifestyle risk factors 212 ) may leave substantial residual and unmeasured confounding. As data in the DNPR currently span almost four past decades, the registry is a unique data source to monitor long-term temporal trends in use of diagnostic procedures (eg, cardiac CT angiography), 164 treatments (eg, use of implantable cardioverter-defibrillators), 213 and disease incidence (eg, myocardial infarction). 27 , 170 Related particularly to disease incidence, however, a number of methodological problems must be considered. First, the DNPR only covers patients with disease episodes associated with hospital contact and thus not necessarily the total number of patients with a given disease (as described previously). Second, lack of information on deaths occurring outside the hospital among persons with no previous hospital contact for a given disease may lead to underestimation of both the disease incidence and the disease-specific mortality. This problem is particularly important for acute critical events such as myocardial infarction. 170 Still, it should be noted that a person is not considered legally dead in Denmark before a physician has confirmed clear signs of death. Thus, all patients dying in an ambulance or otherwise arriving at a hospital with no signs of life are also admitted and registered in the DNPR (even when no resuscitation is attempted at the hospital). Data linkage to the Danish Register of Causes of Death 198 may help to provide a more complete picture of the incidence of acute fatal events not included in the DNPR. 170 Third, it may be difficult – or even impossible – to identify incident diagnoses of chronic diseases in older patients because of immigration or the lack of hospital data before 1977. Thus, events occurring prior to 1977 are left censored if individuals are enrolled in a study and left truncated if they are not. 214 On the other hand, the DNPR enables reconstruction of individual life and health trajectories of persons born in 1977 or later. Fourth, defining incidence by “the first occurrence of the disease in the registry” leads to overestimation of incidence in the period immediately following the initiation of the DNPR, after initiation of a screening program, or after introduction of new registry codes, due to misclassification of “backlogged” prevalent cases as incident cases. Because this problem decreases with the passage of time after 1977 or with the number of screening rounds, a “washout period” before identification of incident cases may reduce the error. This source of error is less important when examining diseases of short duration, such as infections. The transition from ICD-8 to ICD-10 in 1994 and inclusion of outpatients and ED diagnoses in 1995 may similarly introduce artifacts in long-term incidence trends. Exemplifying this problem, the incidence of alcoholic cirrhosis showed no clear trend for men or women of any age from 1988 to 1993 but apparently increased by 32% in 1994 and by an additional 10% when including outpatient and ED visits. 122 Fifth, changes in classification systems and diagnostic criteria and use of more sensitive diagnostic methods over time (diagnostic drift) may hamper the interpretation of secular trends in incidence. As an example, a transient increase in the observed rate of hospitalization with myocardial infarction in Denmark between 2000 and 2004 was likely attributable not to the true increase of occurrence but to new diagnostic criteria introduced in 2000, which included troponin as the main diagnostic biomarker. 170 , 215 Similar time-trend biases have been observed for the incidence of primary liver cancer 203 and advanced stages of lung cancer, the latter leading to an apparent improvement over time in stage-specific prognosis. 216

Flowchart for the systematic review of validation studies. Notes: The literature search was performed on July 20, 2015 using the following search string in 1) PubMed: “Danish National Patient Registry” OR “Danish National Registry of Patients” OR “Danish National Hospital Register” OR “Danish National Health Registry” OR “Danish National Patient Register” OR “Danish Hospital Discharge Registry” OR “Danish National Hospital Registry” OR “Danish Hospital Registers”; and 2) the Danish Medical Journal: “Landspatientregisteret”. Bibliography of validated administrative data, diagnoses, treatments, and examinations in the Danish National Patient Registry Notes: The ordering corresponds to the SKS browser, ie, ICD-10 for diagnoses and NOMESCO for surgery; ICD codes without and with capital letters refer to ICD-8 and ICD-10 codes, respectively; reflects the reviewed number of records in the DNPR (ie, the denominator in calculations of PPV). Among obstetric variables, we included only validation results based on >20 diagnoses; confidence intervals were calculated using Wilson’s score method; information not specified in validation papers, but confirmed through correspondence with authors. Unspecified and unconfirmed data are listed as not available (n/a); recalculation of confidence intervals using Wilson’s score method not possible due to insufficient data; confidence limit equals 100 due to rounding. Abbreviations: A, primary diagnosis; AC, anticoagulant therapy; B, secondary diagnosis; COPD, chronic obstructive pulmonary disease; CT, computed tomography; d, day; DVT, deep venous thrombosis; DANMONICA, Danish Monitoring Trends and Determinants in Cardiovascular Disease project; DCR, Danish Cancer Registry; DNPR, Danish National Patient Registry; DS, discharge summaries; echo, echocardiography; ECG, electrocardiography; ED, emergency department; GP, general practitioner; HBV, hepatitis B virus; HCV, hepatitis C virus; HIV, human immunodeficiency virus; ICD, International Classification of Diseases; ICU, intensive care unit; IN, inpatient contact; LABKA, Clinical Laboratory Information System Database; mo, month; MR, medical records; MRI, magnetic resonance imaging; MS, multiple sclerosis; n/a, not available; NPV, negative predictive value; OUT, outpatient contact; PD, Pathology Registry; PE, pulmonary embolism; PPV, positive predictive value; PR, Prescription Registry; PSA, prostate specific antigen; Se, study sample sensitivity; Sp, study sample specificity; ultrasound, ultrasonography; y, year(s); V-P, ventilation-perfusion; VTE, venous thromboembolism; NOMESCO, Nordic Medico-Statistical Committee; wks, weeks.