Population and setting
The primary data source is the Health Care Services Division Disease Management Evaluation Database (DMED), created to monitor patients enrolled in several initiatives of LSU HCSD's disease management programs. The study population was extracted from all 89,567 LSU HCSD diabetes patients with a diagnosis code ICD-9: 250.xx seen in one of the HCSD's eight hospital EDs between 1998 and 2006. After excluding patients with type 1 and other types of diabetics, patients with only one visit, prisoners, and some cases with obvious errors, the resulting data set contained 30,097 type 2 diabetic (ICD-9: 250.x0 and 250.x2) patients with two or more visits in the study period.
Type 2 diabetes patients (n = 30,097) consisted of three groups: the 6.5% of patients who used only emergency department services; the 8.8% who used only the LSU HSCD DM clinics; and 84.6% of patients (n = 25,475) who used both the DM clinics and the ED sites of care. We kept only the last group of patients for this study. Within the study group, we eliminated an additional 10,176 patients who received some type of care in 1998 because we did not know when their DM treatment was initiated. Only the 8,596 patients whose first records appeared in 1999 were retained for the study.
The 8,596 patients in the study group had a total of 220,719 clinic visits and 60,189 ED visits between 1999 and 2006. The ED visits were classified as urgent and less-urgent, based on the ICD-9 codes and review by two nurses who both agreed on the less-urgent ED classification (n = 28,5440) for the visit. Patient ED visits that occurred before the patient's first diabetes-related visit (chief complaint diagnosis codes, ICD-9: 250 to 250.93) were not counted, as well as ED visits that occurred on weekends because the clinics were not open. We kept only data from 6,412 patients who were 45 years and older. The resultant data set contained 119,695 outpatient clinic visits and 16,249 less-urgent ED visits after the first diabetes-related visit occurring on weekdays. This study was approved by Tulane University's Institutional Review Board (IRB#C0344).
Measures
After removing all urgent ED visits, type 2 diabetic patient visits were classified as less-urgent ED (Y = 1) or clinic visits (Y = 0). Ten independent variables were used in the analyses based on other studies [12, 13] including patient age, health plan, duration in the DM program, and facility size where services were received. For the analysis, facilities were classified as large versus small size based on the facility beds (over 100 or below 100 beds). Other variables included the Charlson Comorbidity Index (CCI); a normal A1c rate over the past 12 months based on laboratory test results (normal A1c rate); 12 month adherence to clinic schedules (adherence rate); and the experience of a prior year hospitalization.
Patients with a range of comorbid conditions had each condition assigned a score from 1, 2, 3 and 6 based on Charlson's study [14]. A higher final score means more or more severe comorbidities. We then summed each patient's scores and assigned a total score to represent his/her comorbid conditions as CCI. The observed A1c test results were grouped into three levels (<7% normal, 7-9% borderline and > = 9% high) to calculate the index of normal A1c rate. (= Σ normal level/(Σ normal level + Σ borderline level + Σ high level) in past 12 months).
Adherence to annual patient diabetes-related clinic visit schedules was based on American Diabetes Association suggestions [15], which were separated into three levels of adherence: none (1 point), midpoint (2 points), and high (3 points). For instance, patients who had no diabetes-related clinic visits within the past 12 months got 1 point; patients with 1 or 2 visits, and where the time between the first and second visit was less than 6 months got 2 points; patients with at least 2 visits, where one of the visit periods was longer than 6 months, received 3 points.
Data Analysis
This study is a longitudinal, retrospective analysis of clinic and less-urgent ED visits from 1999 to 2006. Generalized Estimating Equation (GEE) regression methods for binary responses are appropriate to analyze longitudinal data, especially models with time-dependent variables and repeated measures on the same case. GEE methodology examines the relationship between the occurrence of less-urgent ED visits as compared with clinic visits based on a number of predisposing and enabling factors. The GEE regression model for binary responses identifies those factors that can be altered to reduce the unnecessary use of the ED by diabetic patients participating in the LSU HCSD diabetes disease management program. First, we analyzed the relationship between the outcome and each single unadjusted effect. Then, we computed a full model with all predictors to assess the adjusted variables. All analyses were conducted using SAS 9.12 (SAS Institute, Cary, NC).