The Opportunity for Research

The dream of a computer-based (electronic) medical record is almost as old as the first HMO [1]. With large medical staffs working in multiple medical facilities, the daily problem of missing medical charts led naturally to a vision of electronically stored records that could be accessible at any time from multiple locations. It was also recognized early on that the same data elements needed to maintain electronic records could easily be reconfigured into clinical research databases [2] and that such databases (in the context of defined populations) would create remarkable research opportunities.

Progress toward developing practical electronic medical records has quickened in the 1990s because advances in computer technology and networking strategies have led to more efficient, affordable systems than were previously possible. Equally important, the growth of managed health care has increased competition among HMOs, thereby heightening the need for efficient clinical services and information systems that support these services. The adoption of disease management strategies [3] and outcome-based approaches to quality assurance [4], increasing sophistication of employer-purchasers when requesting data on health care and clinical outcomes [5], and emergence of accreditation organizations and report cards [6] make accurate, readily available clinical data on various aspects of medical care a necessity.

Researchers are probably fortunate that the decision of many managed care organizations to make large investments in clinical information systems was driven by factors other than an interest in research. Most research projects are self-limited, with budgets that are designed to cover the cost of time-limited primary data collection. By contrast, medical records require ongoing, detailed data on entire populations. Quality-assurance activities also require that data collection continue over time. Therefore, researchers are in the position of being end users of databases that are often larger, more thorough, and more durable than any that would have been designed for research programs. Once these databases are developed, they become available for numerous studies of health care services and epidemiology.

From the perspective of researchers studying health care services, the potential exists for capturing the complete health care experience of large, representative populations (including health care utilization patterns, costs of care, and clinical outcomes) in readily retrievable databases. However, most of these databases continue to lack sufficient data on comorbid conditions and disease severity to allow comparisons between populations or to adjust for subtle differences between clinical subgroups. The epidemiologist is similarly drawn to large, defined populations with accurate information on follow-up, clinical events, and to some extent, exposure patterns, risk factors, and other predictors. Again, a concern arises about the adequacy of information on important exposure patterns and covariates in computerized databases.

When additional data are needed, two features of managed health care settings can be advantageous. Older, integrated managed care organizations can access complete paper medical records dating back 20 or more years. These paper records provide previously recorded information on exposure patterns (such as records on prescription drugs or screening tests) and many risk factors (for example, data on blood pressure levels or comorbid conditions) and therefore avoid the risk for recall biases. When medical records are inadequate, most managed care organizations have the ability to conduct mail or telephone surveys of beneficiaries to assess exposure patterns, behavioral risk factors, disease severity, functional status, and quality of life.

Currently Available Data: Description and Limitations

To optimally conduct research or quality-assurance activities, information on beneficiaries needs to be converted into a vast cohort study in which data on membership, demographic characteristics, clinical information utilization, and cost accumulate and are stored over time. These data elements are stored in distinct databases that have been developed at different times and often by different departments. Some important data sets, such as cancer registries or state files on birth and death certificates, are developed and maintained outside managed care organizations. Linkage across databases and over time is therefore needed to fully characterize individual beneficiaries and patient groups.

The descriptions that follow reflect automated data presently available at the Division of Research, Kaiser Permanente, Northern California. Databases with similar characteristics are found in many integrated managed care organizations throughout the United States, although compatibility across organizations is minimal.

Membership Data

Accurate information on dates that members join a health care plan and terminate coverage together with demographic information are essential for characterizing and following patient cohorts. Because patients can leave an HMO at any point, most studies on utilization, cost, clinical outcomes, and epidemiology require construction of person-time denominators from membership data. Departures from a health care plan are as likely to reflect changes in employer or a change in the current employer's choice of health care insurance as they are an employee's decision to change plans. Nevertheless, observation ends when this departure occurs. The data that members join a health care plan also serves as a useful risk adjuster in that recent beneficiaries may be a particularly healthy subgroup, whereas any selection effects tend to decrease over time [7].

Important data elements that are still missing in many databases on HMO membership are race or ethnic group and socioeconomic status. Although these variables can be critical for understanding differences in clinical outcomes or utilization patterns, health care plans have until recently been reluctant to request such information. For research purposes, HMO-based research projects have used geocoding (which links individual addresses to U.S. census data on racial or socioeconomic characteristics of a beneficiary's census tract or block group) to obtain proxy measurements of these variables [8-11].

Data on Utilization and Costs

Utilization databases include counts and descriptions of units of service. These databases are generally archived versions of administrative systems used in the provision that contain information on the delivery of health care services (for example, outpatient appointments, registration, pharmaceutical agents, and diagnostic testing) and the services and costs involved in health care referrals and claims for out-of-plan care. Because they must support clinical activities, these systems tend to be highly accurate. When databases on utilization and accounting are linked, costs can be calculated for a unit of health care service and across categories of service at the patient, provider, or medical facility level. Software that can effect these linkages is operational at many HMOs and has been used to report the costs of illnesses [12, 13] and to add an empirical cost-effectiveness component in randomized and observational studies [14]. The assignment of costs is an assumption-laden process that often results in dubious values for specific units of service. Nevertheless, within a single HMO, costs can serve as a useful common denominator for comparing total utilization patterns between selected groups of members.

Clinical Data

Information on both inpatient and outpatient diagnoses, pharmaceutical records, laboratory results, and results of diagnostic tests are among the data elements commonly available in automated databases. Inpatient diagnoses have generally been available for years as part of state- and Medicare-man-dated reporting systems. Databases on outpatient diagnoses are not yet universally available in prepaid systems of health care, are generally newer than databases on hospitalization, and often have not been extensively validated. Clinical data on out-patient diagnoses depend on the willingness of health care providers to record accurate information and are subject to inaccuracies if providers believe that their income or work load may be affected by what they record. In general, these records do not capture the reason (or reasons) for a visit with sufficient accuracy to attribute the visit or its costs to a single diagnosis. Nevertheless, when these records are present, researchers are able to identify large cohorts of patients with specific diseases.

Measures of comorbid conditions or case mix have been developed on the basis of both outpatient diagnoses [15, 16] and outpatient use of prescription drugs [17]. These measurements have been shown to predict subsequent utilization at least as effectively as self-reported health status [18] and can be readily calculated for entire populations from automated data. Each system captures comorbid conditions and, to a limited extent, disease severity, which is shown by the ability of the systems to predict mortality as well as utilization patterns [19].

Disease Registries

Cancer registries have been maintained or used by researchers in HMOs for several decades, often with the assistance and collaboration of regional Surveillance, Epidemiology, and End Results (SEER) registries. Systems with comprehensive clinical information, particularly databases on outpatient diagnoses, have encouraged the creation of patient registries for diseases and conditions other than cancer. Registries have been developed for illnesses that are considered costly or difficult to manage, such as diabetes mellitus, HIV infection, and end-stage renal disease. The large size of these registries, coupled with availability of data on comorbid conditions and the possibility of long-term follow-up, offers unprecedented opportunities for describing the natural history of these illnesses, understanding the predictors of disease complications, and studying the effectiveness of new therapies or disease management strategies.

Mortality Data

Because patients can die outside a health care facility without seeking medical attention, HMOs are not uniformly aware of the occurrence, timing, or cause of a beneficiary's death. By periodic linkage of membership data to state or national death certificate information, matches can be identified and validated and the necessary information incorporated into a mortality database. Because of the large number of deaths, manual verification is generally not feasible. Software programs that contain algorithms to examine possible matches and assign weights for each match are used [20]. Matches can be categorized as nearly certain, unclear, or highly improbable. For most studies, only those classified as unclear need to be manually reviewed.

Future Improvements

Data on the general health status of HMO members and disease severity of patients with specific conditions are not collected routinely for entire managed care populations because their utility for clinical or quality-assurance purposes has not been demonstrated. However, the increasing enrollment of Medicare population by HMOs may lead to greater appreciation of the need to assess functional status [21]. Periodic assessment of data on general measurements of health and functional status, such as the SF-36 [22], would be extremely useful for adjusting case-mix differences across health plan populations or as an index of disease severity or clinical outcomes in research studies. These measurements have been shown to be related to the presence of clinical illness and are predictive of clinical outcomes [22] and subsequent use of health care [23].

Although validated indexes of disease-specific severity and functional status now exist for many acute and chronic conditions, few of these indexes are routinely measured and incorporated into clinical databases. However, exploratory work has evaluated the routine collection of such information for several populations, including infants who are admitted to neonatal intensive care units [24] and patients in outpatient mental health care settings [25]. Participation of HMOs in national efforts to develop such databases as the National Registry of Myocardial Infarction [26] is increasing largely because data on disease severity and clinical outcomes collected for such registries enhance quality assurance and research activities.

The construction of many databases requires completion of paper forms, coding, data entry, and data cleaning. Data sets are often unusable for weeks to months after clinical events. This limitation reduces the utility of these databases for clinical decision support, quality assurance, and timely research. Future improvements will move toward real-time collection, transmission, storage, and processing of data by way of wide area networks and Internet applications.

Research Centers within Managed Care Organizations

Managed care organizations range from the older, highly integrated staff and group model HMOs, in which all enrolled members are assigned to a single provider group, to newer, looser networks, in which a single HMO (insurer) enrolls members and has contracts with several to many physician groups. The creation of large, well-described cohorts of patients is more difficult when members of one insurer are cared for by multiple medical groups or when patients of a large medical group represent multiple HMOs. Records on membership data are generally maintained by insurers, whereas much of the data on utilization, clinical outcomes, and diagnoses are the property of physician groups and may be considered proprietary by these groups. As physician groups in the newer organizations become more organized and accept more risk, their need to conduct studies on utilization management and quality assurance [27] should provide motivation to develop more complete information systems.

To date, the most active and successful HMO-based research operations are found in the older managed care organizations, particularly in the nonprofit group or staff model HMOs. Active research units have flourished in several of these organizations for 30 or more years and have made substantial contributions to medicine in the United States [28]. In part, the nonprofit status fosters an orientation toward research as a contribution to the public good [29]. However, the overlap of interests between insurer and the physician group has been critical in fostering development of the comprehensive databases needed for research [30].

In 1995, the research units of 10 HMOs convened to consider the advantages of forming a loose research network. Positive aspects of the venture included opportunities to pool data to increase sample sizes and population diversity, to establish the generalizability of observations made in single settings, to obtain external funding, and to exchange ideas among scientists who are working on similar research in similar environments. The current 12 members of The HMO Research Network are shown in the (Table 1). These centers have in common the production of research with the intent of publishing results regardless of findings. Much of this public domain research is funded by external agencies (including the National Institutes of Health, the Agency for Health Care Policy and Research, and the Centers for Disease Control and Prevention) and by private foundations (such as the Robert Wood Johnson Foundation). Network members have pooled efforts to produce important information on various topics, including shared decision making [31] and physician-patient staffing ratios [32].

A substantial amount of this research is conducted in collaboration with investigators from academic settings, including schools of medicine and public health. These collaborations may be initiated by a research center that is seeking the specific expertise of a university-based scientist, by an academic investigator who is seeking access to high-quality data on representative populations, or as a result of successful competition by both university-based and HMO-based research centers for multicenter grants or contracts. These collaborations should be distinguished from simple data sharing, which is seldom in the interest of either a research unit or parent HMO.

Successful collaborations with industry also occur frequently when the interests of an HMO and the funding organization coincide. Such projects require careful negotiation to protect the legitimate rights of both parties and the integrity of the research project.

Examples of Studies on Utilization, Clinical Outcomes, and the Cost of Illnesses

Among the research contributions of these units are examples of descriptive epidemiology [33, 34], etiologic studies [35, 36], studies on the efficacy and effectiveness of treatment regimens [37-40], cost-effectiveness studies [12-14, 41], and examinations of issues related to organizational change [42] and methods for improving assessments of population risk and capitation rates [18, 23]. The effectiveness of treatment programs, prevention strategies, health education, and provider training interventions have been evaluated with randomized trials where feasible and quasi-experimental or observational methods when necessary.

The two studies to be described were accomplished by linking multiple databases at the Kaiser Permanente Medical Care Program, Northern California. They are presented as examples of research projects generated by the need for quality assurance. In each study, both patterns of utilization and clinical outcomes are reported.

Effects of Emergency Department Copayment

In the latter part of 1992, several large employer-purchasers demanded introduction of a $25 or $35 copayment for emergency department visits for covered employees and family members, with a concomitant reduction in premiums. Concerned about the possibility of adverse effects on patients, Kaiser Permanente commissioned and funded a study that compared use of the emergency department and other outpatient departments and clinical outcomes before and after institution of the copayment. From the membership database, the 30 600 members subjected to the copayment and two control groups that totalled more than 90 000 members were identified. Proxy measures of socioeconomic status were obtained by linking (geocoding) home addresses to U.S. census data on block groups. By linking membership and copayment information to databases on utilization and mortality, researchers demonstrated a reduction of 15% in emergency department use by beneficiaries of the copayment group compared with use by those in the two control groups [43]. Increased use of other outpatient services was not detected. Diagnoses were stratified into four levels of urgency, and a decrease was primarily shown to be in the less-emergent diagnostic categories. There was no evidence of increases in hospitalization rates (overall or for a set of potentially preventable hospitalizations), no differences in mortality, and no differences in the rate of death from myocardial infarction. Therefore, this study suggested that relatively small copayments for emergency department services are effective and safe for insured populations younger than 65 years of age.

Area Variation in Use of Coronary Angiography and Revascularization

Remarkable variation in medical practice can be found within and between health maintenance organizations. For many years, Kaiser Permanente, Northern California, was aware of considerable variation in angiography practices among the 16 medical centers in the region. Between 1990 and 1992, the use of angiography during the 3 months that followed myocardial infarction differed by more than twofold (range, 30% to 77%) among these centers.

By using hospital discharge files and the mortality database, 6851 patients admitted for myocardial infarction during this period were identified and followed for as long as 4 years. Clinical outcomes included death from heart disease, reinfarction, and rehospitalization. Secondary discharge diagnoses recorded at the index hospitalization and all discharge diagnoses from the previous 5 years were used to create a comorbidity index adapted for claims data [44]. Through treatment of the medical center as the unit of analysis, a weak inverse association was demonstrated between the facility's rate of angiography or revascularization and clinical outcomes [45]. However, in an analysis that reviewed the charts for a subgroup of 1109 patients, detailed data on disease severity and information on the clinical course and results of diagnostic tests were collected. Using this information, researchers stratified patients according to RAND necessity criteria for angiography [46] and repeated the analyses separately in each stratum. The strong ecologic association of angiography rate with clinical outcome seemed to be confined to patients who met necessity criteria and to be explained by receipt of received angiography. Therefore, this quality-assurance study provided an explanation for the failure of many area variation studies to find associations between medical practice and clinical outcomes: only a subgroup of the population should be expected to benefit from a practice or procedure. The information on disease severity and diagnostic testing necessary to identify this subgroup was not available in automated records, which necessitated costly, lengthy review of patient charts.

Conclusion

Top Conclusion Author & Article Info References

Managed care organizations provide an excellent setting for conducting research on epidemiology and health care services. For business reasons, these organizations are committed to developing comprehensive, high-quality information systems. The opportunities for conducting relevant research in areas of clinical effectiveness, cost-effectiveness, public policy, and epidemiology are just beginning to be tapped. As managed care organizations grow, research by these organizations increasingly emulates public health research. The future will undoubtedly witness a continued increase in collaborations between HMO-based research centers and researchers from academia and government.