Disease surveillance and referral bias in the veterinary medical database

Abstract

The Veterinary Medical Database (VMDB) is a summary of veterinary medical records from North American veterinary schools, and is a potential source of disease surveillance information for companion animals. A retrospective record search from four U.S. university veterinary teaching hospitals was used to calculate crude disease rates. Our objectives were to evaluate the utility of the database for disease surveillance purposes by comparing the utility of two methodologies for creating disease categories, and to evaluate the database for evidence of referral bias. Summaries of the medical records from November 2006 to October 2007 for 9577 dogs and 4445 cats were retrieved from VMDB for all canines and felines treated at Kansas State University, Colorado State University, Purdue University and Ohio State University. Disease frequency, computed as apparent period-prevalence and as the percentage of veterinary visits, was compiled for 30 disease categories that were formulated by one of two methods. To assess the possible impact of referral bias, disease rates were compared between animals residing in zip codes within 5 miles of the hospitals (zone 1) and those animals living at more distant locations (zone 2).

When compared to zone 1 animals, disease conditions commonly associated with primary veterinary care were reduced by 29–76% within zone 2, and selected diseases generally associated with more specialized care were increased from 46 to 80% among zone 2 animals. The major differences in disease prevalence seen between zones suggests that substantial referral bias may exist, and that adjustment on the basis of geographical proximity to the university teaching hospitals may be useful in reducing this type of selection bias in the VMDB, thereby improve the accuracy of prevalence estimates and enhancing the utility of this database for purposes of disease surveillance.

Introduction

Passively collected disease surveillance data is prone to many types of biases. Nevertheless, field epidemiologists use admittedly imperfect databases to produce estimates of disease occurrence and disease trends that are useful for many purposes. For example, the Morbidity and Mortality Weekly Report (MMWR) is published weekly by the Centers for Disease Control and Prevention, yet multiple types of selection and informational bias heavily influence its disease rates that are based entirely on passively collected observational field data. For the purposes of disease surveillance, a reasonable estimate is often preferable to no estimation at all, even though observation surveillance data is clearly prone to many more biases than would be expected in data collected for a well-designed and controlled laboratory experiment.

Disease surveillance within companion animal populations is useful for many purposes, including veterinary service resource projections, directing teaching curriculums, detecting disease trends and also for animal health product development. Additionally, the 2007 pet food melamine exposures exemplified how monitoring of disease in an animal population might signal an early warning of a human disease health hazard (Brown et al., 2007, pp. 525–531; Puschner et al., 2007, pp. 616–624; Dobson et al., 2008, pp. 251–262). The founders of the Veterinary Medical Data Base (VMDB) realized these and other applications (VMDB, 2008). The VMDB was started in 1964 by the National Cancer Institute to abstract the medical records of the North American veterinary schools (Folk et al., 2002, pp. 405–410; Hahn et al., 2004, pp. 475–479). Currently, 26 veterinary schools have submitted data, but many are many years behind in their data submission. Medical records at each school are transcribed into SNOMED (Systematized Nomenclature of Medicine) codes and then centralized to a national database. Only basic demographics, dates, diagnosis and medical procedures are collected by VMDB. Patient number and owner identifiers link the VMDB database to each hospital's own medical records system, where researchers can retrieve more complete records upon approval by each hospital. However, the VMDB database manager reports that many schools are several years behind in SNOMED coding their medical records, and some institutions contribute data from only some of their clinical services. Nevertheless, the VMDB is the oldest and probably the most comprehensive pet animal health database in the U.S., and its utility for disease surveillance purposes and observational studies warrants evaluation.

Referral bias is a type of selection bias that affects estimates of disease frequency if the monitored clinics or hospitals have a predominance of patients that were referred for more specialized care, and may also exist if they have a relative deficit of cases receiving primary care (Salive, 1994, pp. 808–809; Sepkowitz, 1990, pp. 1629–1631; Froom and Froom, 1992). The extent of referral bias is expectedly different for every disease diagnosis, so no one measure of referral bias would be expected for all diseases or disease categories. Referral bias is, therefore, problematic if one's objective is to estimate disease frequency among the entire segment of the pet population that utilizes both primary and referral veterinary services.

Disease rates are the most basic of all the epidemiologic measures. Most clinicians realize that a clinical or laboratory value, e.g. BUN or body temperature, is useless in the absence of information regarding the expected or normal value in the healthy population. Similarly, disease rates for one particular veterinary clinic have little meaning unless compared to a large reference population. With up to 26 possible participating clinics, the VMDB database has the potential to serve as a reference population for national animal disease rates, if and when referral bias and other biases can be ameliorated. Private practitioners may wish to compare their practice caseloads to this reference population to identify areas for practice expansion or contraction of services, and to identify and monitor unexpected clusters of disease in their particular geographical area.

Our objectives were to evaluate the utility of the database for disease surveillance purposes by comparing the utility of two methodologies for creating disease categories, and to evaluate the database for evidence of referral bias.