An electronic health record-enabled obesity database

Obesity is a multi-factorial disease that has reached epidemic proportions in the US, with about two thirds of adult Americans overweight (body mass index [BMI] > 25 kg/m²), one third obese (BMI >30 kg/m²), and about 5% with class III levels of obesity (BMI > 40 kg/m²) [1]. Obesity is associated with increased risk for a variety of co-morbidities, higher lifetime health care expenditures, and a greater risk for mortality [2]. Increasing BMI is associated with further increases in disease burden and risk of mortality [3] and the prevalences of both class III (BMI > 40 kg/m²) and super (i.e., BMI >50 kg/m²) obesity are rising faster than other classes of obesity [4]. Dietary modification and physical activity are effective at decreasing obesity-related disease risk and severity [5], although only modest weight loss is usually achieved and few individuals are able to maintain weight loss over long periods. In contrast, Roux-en-Y gastric bypass (RYGB) surgery commonly results in a more substantial and sustained weight loss, and has thus emerged as a more effective therapy for long-term weight loss in morbidly obese patients [6], and more recently, as a surgical therapy for type 2 diabetes [7]. However, the degree of weight loss with RYGB is variable [8, 9] and the factors determining long-term success are not known. The effect of RYGB of various clinical outcomes is also not well characterized.

To conduct research on RYGB outcomes, significant resources are required for the long-term follow-up of patients as well as for clinical characterization. Economic efficiencies can be gained through leveraging existing clinical infrastructure. One such resource that has only recently been exploited for research is the electronic health record (EHR). Body mass index, calculated from height and weight, is perhaps the most commonly used measure of obesity and is often measured during clinical encounters and entered into EHRs. A large number of other variables, such as laboratory test results, the presence of co-morbid conditions, and medication use are also often recorded in EHRs. The potential availability of such data suggests that EHRs may be a rich resource for RYGB and obesity research.

We have developed a research database within the context of a Weight Management Clinic at a large integrated health system through utilization of an EHR fed data warehouse. A large number of clinical variables on patients undergoing RYGB surgery were extracted and assembled into research datasets. To illustrate the potential utility of the data, we estimated the length of post-operative follow-up after RYGB surgery. The efficient electronic acquisition of these data serves as a new paradigm for RYGB and obesity research.

All patients enrolled in the Geisinger Bariatric Surgery clinical program are offered participation in the ongoing obesity research program. The results presented below were obtained from 2028 patients who had completed RYGB surgery between January 1, 2004 and July 2, 2010. The overall consent rate for this group was greater than 90%. The program continues to accrue patients at a similar rate.

Length of follow-up

The amount of time a patient is followed after a weight loss intervention is extremely important since the percentage of patients that experience weight gain recidivism is high. We used date of surgery and dates associated with follow-up BMI measurements to calculate the percentage of patients lost to follow-up over time using Kaplan-Meier curves (Figure 2 and Additional file 1: Table S6). Because patients with lower degrees of obesity may achieve desired weight loss goals sooner than patients with more extreme levels of obesity [17, 18], we stratified patients into three BMI ranges, 35–39, 40–49, and 50+ kg/m². As can be seen in Figure 2, patients in the 35–39 kg/m² range had similar or higher rates of follow-up data until 48 months when they had a 77% follow-up rate versus 79% and 74% for the other two groups (log rank p-value = 0.077), although the sample size is small for the lighter patients. At 1 year all three BMI groups had 92% or greater rates of patients with follow-up data, and at both 24 and 36 months only patients in the highest BMI category were less than 80% (75%). Overall rates of follow-up dropped less than 10% per year over the course of the follow-up period.

We found no significant differences among year of surgery in length of follow-up for 2004, 2005, 2006, and 2007 (log-rank p-value = 0.262). The effect of surgeon on follow-up was also evaluated. There were five primary surgeons that accounted for 96% of the surgeries, with 2 of the 5 who accounted for 66% of the surgeries. These two surgeons were compared with each other and with the other surgeons as a single group. No significant difference was found in length of follow-up (p = 0.130). We also analyzed the type of surgical approach (open versus laparoscopic), which was also not related to length of follow-up (p = 0.123).

The EHR is a longitudinal electronic record of patient health information generated by clinical encounters in a variety of care delivery settings and includes patient demographics, progress notes, problem lists, medications, vital signs, past medical history, immunizations, laboratory data, and radiology reports. EHRs make use of relational database structures and utilities to access and display data that facilitate medical care and clinical decision-making, substituting for traditional paper-based “charts”. EHRs may also help address the long-standing problem of the long lag time that exists before evidence-based medical knowledge is used in clinical care [19]. In addition, the EHR is a robust source of data that can be exploited for research. It provides a means to characterize patients through exploiting existing clinical data, rather than re-capitulating such data via research activities.

In research protocols, physical measures such as height and weight are generally collected by research assistants trained in specific protocols. The tools for those measurements are typically calibrated instruments to reduce error. In contrast, clinical measurements that are recorded in the EHR may be taken by a variety of health care professionals whose procedures may be less stringent, and with equipment that may differ by location. For example, weights may be fully clothed or gowned; heights may be with shoes or without. Some of this type of error can be mitigated by large sample sizes, but some systematic errors cannot. If, for example, the majority of heights are measured with shoes, BMI calculations will be systematically biased downwards. In this study, key height and weight data were extracted from the EHR from measurements obtained in the Weight Management clinic in which a research-like standardized process using calibrated instruments was performed by trained personnel in the same clinic. This standardization also improves the rate of measurement of height, which is measured less frequently than the measurement of weight in EHRs [20].

Data on medication usage and co-morbidities may also present potential problems when extracting from EHRs. We used medication reconciliation and not physician medication orders, sometimes used as a proxy for medication usage. However, not all filled prescriptions will be used by patients. International Classification of Diseases, ninth revision, (ICD-9) codes, which are used for billing, insurance, and documentation in EHRs, as well as for tabulation of clinical statistics and quality analyses, were used as surrogates for the presence of co-morbid conditions. We did not attempt to document the accuracy of individual ICD-9 codes using supporting data such as laboratory results or medication orders. However, the percentages of patients assigned specific ICD-9 codes was similar to the percentage of patients prescribed corresponding medications for a number of major disorders including diabetes and biguanides, depression and SSRIs, and osteoarthosis and NSAIDs. Less closely correlated were hypertension and hypertension medications likely because of the use of tailored polypharmacy to treat individual patients. In addition, the percentage of patients prescribed proton pump inhibitors was higher than the number with GERD likely because of prescribing for other gastrointestinal disorders.

Laboratory data are among the most robust data available from EHRs. We were able to extract a large number of laboratory variables on most patients. Two major reasons for why patients lacked data were that the tests were either never ordered or were obtained at an outside laboratory and the results transmitted in non-electronic form, i.e. paper or scanned portable document format (pdf) copy. This contrasts with survey data in which surveys were offered by providers to each patient, whether in scannable paper format or through on-line access. Since they were not “prescribed” through formal physician order entry, the relative rate of return of this data was less than other types of data domains. Nevertheless, substantial numbers of several survey instruments were obtained.

Despite the presence of clinical data in EHRs that are readily accessed by providers on individual patients, such data can be logistically difficult to extract for research use. Some research groups have been developing natural language processing approaches to obesity-related EHR data [21–23]. We used a data warehouse that mirrored the EHR and greatly facilitated practical access to data. Other obesity research studies have utilized data warehouses [24]. The Veterans Health Administration stores height and weight measurements that have been entered into the EHR in the national Corporate Data Warehouse (CDW). Similar to the Geisinger CDIS data warehouse, the CDW was developed to allow access to data and tools for several purposes including research. Although weight, height and other data are deposited at frequent intervals, the CDW does not yet contain data on laboratory measures, procedures, and diagnoses.

Consensus guidelines have been recommended for determining weight loss following bariatric surgery using operative weight and follow-up weights [25], including the need to define operative weight as the weight at admission or just before surgery. Our data was based upon weights obtained the day of admission based upon the clinic code used for admitting patients for RYGB surgery. Our length of follow-up was also above the recommended minimum rate of 61%.

EHR data can be a valuable source of data for obesity research, although the availability and integrity of different data types can vary substantially. Access to a data warehouse can greatly enhance the efficiency of data collection over direct extraction from the EHR. EHR derived data can be used for a variety of research and clinical uses such as for determining length of follow-up.