The Rochester Epidemiology Project: exploiting the capabilities for population-based research in rheumatic diseases

Abstract

The Rochester Epidemiology Project (REP) is a patient record-based database based upon a medical records-linkage system for all residents of the Olmsted County, MN, USA. This comprehensive system includes all health-care providers of patients resident in this geographically defined region. It uniquely enables long-term population-based studies of all medical conditions occurring in this population; their incidence and prevalence; permits examination of disease risk and protective factors, health resource utilization and cost as well as translational studies in rheumatic diseases.

Introduction

Comprehensive community-based health-care information systems are rare in the USA, making it difficult and costly to conduct true population-based epidemiological, health services research and natural history studies across the full severity spectrum of rheumatic diseases. Most of these studies include only patients attending specialized centres or specialty clinics, those agreeing to participate in registries, populations covered by a single payor or respondents to surveys. In contrast, the Rochester Epidemiology Project (REP) maintains a medical records-linkage system for all residents of the Olmsted County, MN, USA that enables long-term population-based studies of incidence, prevalence, risk and protective factors, health services utilization, cost-effectiveness and, more recently, translational studies in rheumatic diseases. This review describes the unique research capabilities of the REP in the context of rheumatic diseases, with a particular focus on recent advances in availability of bioinformatics and translational resources. Interested readers are also encouraged to refer to an earlier review of the REP resources published in 2004 [1].

The REP links medical records generated by different health-care providers over many years to specific individuals, maintains an electronic index of diagnoses and surgical interventions, and provides an ongoing census of individuals as they move in and out of the community over time. Basic components of the REP are illustrated in Fig. 1. Subjects can be followed through their outpatient office, urgent care, emergency department and hospitalization contacts with all local health-care providers (e.g. Mayo Clinic, Olmsted Medical Center and affiliated hospitals and private practitioners), allowing for the longitudinal follow-up of a well-defined population-based cohort that currently includes approximately 125 000 subjects. Since the REP was established in 1966, more than 700 000 subjects have been included in its medical record , and it has supported several studies and provided the basis on which close to 100 publications on the aetiology and outcomes of rheumatic diseases occurring in this population over the entire lifetime of the community residents, from birth to death, have been completed. These rheumatic diseases include RA [2–44], JRA [45–50], PsA [51–54], OA [55], SLE [56], GCA and PMR [56–72], gout [73, 74], SS [75], Behçet’s disease [76] and AS [77].

Fig. 1.

REP structure.

What is unique about the REP as compared with other rheumatic disease databases?

The REP includes all conditions that come to medical attention in a geographically defined population and links the diagnostic codes to the extensive details incorporated in complete (inpatient and outpatient) contemporary and archived medical records for each resident of the Olmsted County. Unique capabilities of the REP are summarized in Table 1.

Table 1.

Unique advantages of REP for population-based research in rheumatic diseases

Construction of population-based cohorts consisting of virtually all clinically recognized cases across the full spectrum of the disease

All diagnoses coded in REP index Assigned for clinical (not reimbursement) purposes Detailed Mayo adaptation of HICD-A (more informative than four- to five-digit ICD codes) High sensitivity Includes all diagnoses, regardless of setting Range of diagnosis codes used to identify potential cases is typically very broad Local residents have typically long duration of time over which condition can be clinically recognized High specificity Access to complete and detailed clinical information in provider-linked medical records Application of standardized research criteria for case validation; residency determination; and differentiation of incident cases from prevalent cases

Capacity for longitudinal follow-up of each member of every REP population-based cohort

Incident cases distinguished from prevalent cases Reduces problems of left censoring Affords estimates of effect of disease duration Essentially complete follow-up for mortality with REP death data

Active follow-up

Relatively high response rates Can estimate extent and consequences of response bias Access to detailed information in provider-linked medical records For persons who authorized use of records in research (rates of authorization typically 90–95%)

Passive follow-up is an hallmark of REP studies

Retrospective record review and HIPAA waiver Minimal participation bias Minimal loss to follow-up Excellent for clinically recognized morbidity (as a result of access to detailed information in provider-linked medical records) Information on clinical outcomes essentially complete (geographic isolation and small number of providers) Information available over extended periods (relatively low emigration rates; among emigrants, many return for care)

Health services utilization and cost data

Majority (>95%) of care provided by Mayo Clinic and two affiliated hospitals, Olmsted Medical Center and affiliated hospital Utilization and billing data are linked where each Olmsted County resident is assigned a unique ID across settings: ambulatory/ED/hospital inpatient and outpatient Utilization and billing data include Line-item detail on all goods/services: type and date provided, frequency and billed charge All plans (fee for service, managed care, Medicare, Medicaid and uninsured)

HICD-A: Hospital adaptation of the ICD: international classification of diseases; HIPAA: health insurance portability and accountability act; ED: emergency department.

Each of the several rheumatic disease databases in the USA or Europe has strengths and limitations. For example, prospective registry-based data or surveys are available for some specific conditions [78–80], but most such studies must actively recruit subjects, which is expensive and increasingly hindered by non-response bias at baseline and additional subject loss during follow-up [81–83]. Nationwide registries of larger populations in northern Europe link diagnoses, but do not include the comprehensive medical information from individual medical records. Health Maintenance Organization databases have passive access to diagnoses and procedures for thousands of patients with rheumatic diseases, but are not population based; may have overrepresentation from the working public; fail to include care given in the community but outside the health plan [84]; and do not provide large volumes of longitudinal data because members often change their insurers, resulting in high turnover [85]. US Medicare and Medicaid data cover very large populations and can be accessed with proper permission, but the data available are limited to administrative or billing data rather than the detailed medical records indispensable to attribute signs, symptoms, and costs to a condition and to study the course of a disease [86]. Moreover, these populations are limited to specific age and socio-economic groups, and the resulting administrative data may be further subject to biases driven by local billing practices [87, 88]. In contrast to these registry resources, the REP includes essentially the entire population of the Olmsted County, and provides longitudinal information and linkage to the medical records for inpatient and outpatient visits, regardless of the socio-economic status, insurer, health-care setting, provider or disease severity.

The REP population databases for various rheumatic diseases: RA as an example

The REP maintains a records linkage system for the complete population as defined in the classic epidemiological literature [89, 90]. Indeed, the REP covers >98% of all medical care provided to the residents, regardless of age, socio-economic status or insurance coverage, and the local health-care providers span the spectrum from primary outpatient care through tertiary and critical hospital care including all medical and surgical specialties. Although best known as a tertiary referral centre, Mayo Clinic has always provided primary and secondary, as well as tertiary, care to the residents of the Olmsted County. Furthermore, in contrast to many other registry-based cohorts in the USA, a patient does not need to be seen by a rheumatologist to be included in a REP-based study cohort. Indeed, many of the REP rheumatic disease study cohorts (i.e. RA, JRA, SLE, GCA, PMR, PsA and scleroderma) include patients who were almost exclusively managed by internists or other primary-care providers.

RA is an excellent example of an ongoing population-based cohort in the Olmsted County. The RA incidence cohort dates back to 1955 and it was first assembled using the 1958 ARA criteria. Later, the cohort was reassembled based on the 1987 ACR classification criteria [91]. Since then, the same criteria have been consistently applied to ascertain nearly 1200 incident RA subjects as of 1 January 2008.

An important challenge in designing studies of risk factors and outcomes is identification of an appropriate source population for selection of unexposed cohorts or controls. In the case of REP, this is a straightforward process. Since all RA patients come from the precisely defined and closely monitored the Olmsted County population, controls can easily be sampled from the same population (Fig. 2). The REP maintains an enumeration of the local population because >95% of the Olmsted County population has at least one contact with a health care provider every 2–3 years and their medical informations are continuously entered into the REP data systems. This complete enumeration also creates a framework for random sampling of the general population for prospective studies, complete with address, telephone number and other contact information.

Fig. 2.

Sampling of study cohort using the REP.

Studies across the full spectrum of care regardless of health-care provider or insurance

The REP resources afford essentially complete information on the entire medical history of the Olmsted County residents, extending from date of first contact with any health-care provider in the Olmsted County through death (or emigration). Very few other databases allow study of patients across all health-care delivery settings and payors. The REP resources include all primary, secondary and tertiary levels of care and, with few exceptions (some local orthodontists, optometrists, chiropractors, rehabilitation therapists, psychiatrists and home health-care providers), all health-care delivery settings (including office visits; specialty consults; emergency department, hospital inpatient and outpatient encounters; and nursing home stays). The ability to study individuals as they undergo transition across health-care delivery settings affords a more complete understanding of health-care utilization and costs [11, 12], risk factors and outcomes [92]. For example, in a recent study [92], we were able to identify all residents who had autoantibody testing (irrespective of rheumatic disease status) at any of the health-care providers in the Olmsted County, not just the Mayo Clinic.

Studies of long-term causal associations and outcomes of rheumatic diseases

Rheumatic diseases may develop at almost any point in life and are the result of genetic, epigenetic, developmental and early-life events and risk factors. Therefore, only studies considering long-term exposures/risk factors may be able to clarify causal associations. For example, to design studies linking pregnancy or early-life events to rheumatic diseases, a follow-up of >60 years may be needed. Indeed, the REP-based studies of rheumatic diseases typically have a very long duration of medical history, both before and after disease diagnoses. For example, the RA incidence cohort contains a median of 30.5 years (25th percentile: 15.2 years, 75th percentile: 42.8 years) of available medical history before, and 9.5 years (25th percentile: 5.0 years, 75th percentile: 16.7 years) after RA incidence date. Figure 3 illustrates the median duration of medical history, both before and after RA incidence date in our cohort by decades of age.

Fig. 3.

Median length of medical history before RA incidence date and median follow-up after RA incidence date according to age at RA incidence.

The extent of medical history also allows studies across generations. The REP maintains a database of birth records for births in the Olmsted County, which allows identification of mothers. In an attempt to examine the risk of heart disease among children of women with RA, we identified the children of 441 women with RA and 441 women of similar age and sex without RA. We were able to identify 381 children of women with RA and 414 children of women without RA and followed them retrospectively through their medical records from birth to date (mean follow-up of 40 years; Crowson CS, data not published). This time frame requires the existence of historical archives or databases that were established and maintained in the past and can be used now to test aetiological hypotheses. Similarly, to efficiently explore the long-term effectiveness and safety of drug treatments, surgical interventions or preventive measures, it is essential to combine historical documentation of exposures with current follow-up. In a study that spanned several decades [93], we demonstrated a significant reduction in joint-related surgeries for patients diagnosed between 1985 and 1994, as compared with 1965–74 and 1975–84. Another excellent example of this unique strength is our studies describing trends in RA mortality, providing compelling ecological evidence that the improvements in morbidity and quality of life in RA have not led to improvements in survival [14]. As the time frame involved often spans decades, short-term studies may be unable to measure the associations and effects. Population-based historical cohort studies have been the hallmark of the REP-based rheumatic disease cohorts, and the ongoing collection, coding and storage of contemporary medical or surgical data are essential for cohort studies in the future.

Feasibility of different study designs (a population-based approach): retrospective cohort studies and case–control studies of RA and its comorbidities as examples

The unique capability of the REP to enumerate the entire population allows innovative research designs and analytical approaches. For example, the ability to ascertain all cases of a certain disease in the community permits the assessment of population-attributable risks, an approach rarely feasible in other observational studies [94]. In another study examining the association between oral contraceptives and RA, we were able to demonstrate that the population-attributable risk was small and that the decline in RA incidence could not be explained by oral contraceptive use because the proportion of women exposed to oral contraceptives over this period of time was relatively small [27]. In addition, the availability of multiple concurrent population-based inception cohorts facilitates the use of case cohort designs, greatly augmenting statistical power in the observational epidemiology studies [95, 96]. This approach is being utilized in our current studies examining the risk of heart disease in RA.

REP facilitates studies with reduced risk of biases

Since the REP is based on a geographically defined population, it is not subject to the healthy beneficiary bias that may be encountered in managed-care systems, nor to the referral bias that may be present in registries based in tertiary-care hospitals and clinics. Indeed, there may be dramatic differences in the clinical spectrum and outcomes between referral patients and unselected community patients. Moreover, unlike US Medicare data, the REP data are not limited to persons aged ≥65 years, and the inclusion of the younger population is critical for investigating early vs mid-life determinants, outcomes and treatment of rheumatic diseases. For example, the REP-based cohort of JRA patients were followed up well into adulthood, and we were able to examine their BMD [50] and psychosocial status [48] in adulthood. Importantly, because medical record information is recorded at the time the disease or risk factor is observed, it is not subject to the potential recall biases present in self-reported data [97].

The REP also facilitates studies minimizing incidence-prevalence bias. Studies based on prevalence series include patients of varying disease duration and exclude those who died early of the condition. In contrast, the REP resources afford identification and follow-up of incident series that reflect the full range of disease severity and mortality. This was strikingly evident in a mortality review in RA where the pooled standardized mortality ration from inception cohorts (both, community/population and clinic based) were much lower than those in non-inception cohorts [98]. The same is true for population-based rates of various types of infections in RA patients, which are difficult to obtain in clinic-based settings and frequently cited as background rates [17]. Finally, the REP studies are able to reduce non-participation bias. Due to the passive access to medical record data for virtually the complete population, the samples of study participants obtained using the REP are frequently more representative of the general population than samples of subjects who participate actively in a study [99]. This may be particularly important in studies evaluating treatment effectiveness and safety.

Use of electronic medical records for ascertainment and monitoring of the anti-rheumatic treatment and safety surveillance: biologics as an example

In recent years, traditional technologies of data processing at the Mayo Clinic (i.e. manual data abstraction from medical records) have been enhanced by a variety of tools, in particular, natural language processing (NLP). These tools allow free text in the electronic medical system to be indexed, searched, retrieved and analysed using state-of-the-art computer science techniques. Electronic medical records (EMRs) coupled with NLP provide the unique capability for data mining within the REP resources.

The clinical Text Analysis and Knowledge Extraction System (cTAKES) is a pipeline that was developed at the Mayo Clinic and was released as an open source at www.ohnlp.org. The cTAKES processes clinical notes and identifies the following clinical named entities: drugs, diseases/disorders, signs/symptoms, anatomical sites and procedures. Each discovered named entity is assigned attributes for the text span, the ontology mapping code, the context (family history, current, unrelated to patient) and a negation indicator. For example, in the sentence ‘No evidence of unstable angina’, unstable angina is discovered as a named entity of type disease/disorder, the text attribute is populated with the spanned value, the associated code gets a value of 4 557 003 for a systematized nomenclature of medicine-clinical terms (SNOMED_CT) match, the status is assigned the value of ‘current’, and negation is set to true.

These capabilities have already been applied to musculoskeletal and cardiovascular research, where unstructured text of the EMR was analysed to identify patients with RA. Although case ascertainment for most studies is currently performed mainly by manual data abstraction from paper or electronic clinical reports, the increasing adoption of the EMR makes it possible to analyse the unstructured text of clinical reports in ways that were historically impossible or cost prohibitive with paper-based records. A number of projects conducted using the Mayo Clinic EMR system have also shown the advantages of free text over US international classification of disease-9 disease billing codes for case identification. In addition, these new capabilities will expedite the process of obtaining health outcomes. By first identifying potential health states, this new information technology will lessen the burden and expense associated with manual records review and data abstraction. Furthermore, risk factors such as smoking status [100] and alcohol use can be automatically extracted from each patient’s medical records.

EMR-based systems empowered with NLP/text mining applications have been shown to efficiently identify and monitor cohorts of drug-exposed subjects over extended periods. For instance, we have shown the usefulness of EMR for safety surveillance using biologics as an example [101]. This system can also be useful for non-specific outcomes, such as signs and symptoms that are not coded and cannot be identified in administrative databases. This information analysed in context of the medication use can serve for monitoring of the treatment effectiveness and safety surveillance. Clinical notes can be analysed for positive and negative evidence, as well as for the degree of probability of signs and symptoms, thus providing accurate identification of medical status. These data can be supplemented with detailed information on medical history and clinical characteristics.

Among the developing capabilities for information and data mining are: statistical classifiers (machine learning approaches); improved mapping from text to concepts in a standards-based ontology (e.g. SNOMED-CT); drugs and adverse outcomes; and disambiguation of term senses.

Together, these novel technologies for information extraction from the clinical narrative provide an opportunity for the effective use of the extensive resources of the REP medical records-linkage system for hypothesis generation and testing in clinical research and epidemiology. Richly annotated data across the EMR has the potential to unveil previously unexplored or rare patterns of association. Meystre et al. [102] provide an overview of the state-of-the-art of information extraction from the EMR.

Infrastructure for translational research

The REP also facilitates T2 and T3 translational studies [103]. The acquisition of human biospecimens and the subsequent cellular/molecular analysis of that tissue and the ability to link with phenotypic information are critical to early translational studies in rheumatic diseases. Since the beginning of the 20th century, the Mayo Clinic maintained a tissue repository. Paraffin blocks and sections from all biopsy, surgery and autopsy procedures are stored indefinitely and easily accessible. For example, by relying on autopsy material from this tissue repository, we were able to assess the extent and characteristics of coronary atherosclerosis in RA patients compared with controls [104]. A unique T2–T3 translational resource of the REP is the provider billing data for all the Olmsted County health-care providers. This resource captures ∼95% of all physician and hospital services provided to local residents dating back to 1987. Since this resource covers the entire Olmsted County population, it is possible to compare the costs for a cohort of persons with a particular disease with costs for persons unaffected by that disease or with costs for the general population. Several studies relied on this unique resource to demonstrate the direct medical, indirect medical and lifetime costs of RA; costs related to NSAIDs [105], costs of rheumatological care [15]; costs of osteoporotic fractures [106]; costs of OA [11, 55]; and cost of PMR [71].

Conclusions

The REP is a large, centralized medical data system including the lifetime health-care information of essentially all residents of the Olmsted County, MN, USA over the past five decades. The unique features of the REP provide unparalleled capabilities for the population-based research in rheumatic diseases. In particular, the resources of the REP enable the longitudinal population-based studies across the full spectrum of rheumatic diseases with long and complete follow-up regardless of health-care provider or insurance. This type of data resources will be particularly useful in the future for comparative effectiveness and patient centred outcomes research. These unique capabilities of the REP are continuously advanced with the novel technologies (particularly, of the data processing and translational research) providing the state-of-the-art research platform for epidemiological studies. The findings obtained through the REP significantly contribute to the overall knowledge of the epidemiology of rheumatic diseases and may provide guidance for other studies into the aetiology, outcomes and impact of rheumatic diseases. Thus, the REP can be viewed as an exemplar of the advanced and constantly improving research infrastructure, informing and guiding the epidemiological knowledge, particularly with regard to rheumatic disease.

Disclosure statement: The authors have declared no conflicts of interest.