This study includes two phases of data collection and analysis. In Phase 1, we assessed data directly from the electronic medical record (EMR) and survey data. Phase 2 supplemented these analyses with health plan claims data for the members assessed in Phase 1. The study protocol was reviewed by the Allegheny Health Network (AHN) Institutional Review Board and was determined to be a clinical quality improvement project and not the human-subjects research, hence protocol was approved and requirement for informed consent was waived.
Phase 1: Management Practices
Inclusion and Exclusion Criteria of Participating Practices: All AHN primary care practices were included in the initial analysis, and 50 practices adopted HEDIS measures to monitor internal performance and quality improvement. HEDIS measures of all participating PCPs were collected using EMR and updated on a quarterly basis. Practices with fewer than 50 diabetic patients reported per quarter were excluded. A total of 44 practices with 19,059 clinic visits by patients with a diagnosis of diabetes were measured against these metrics based on their performance over a one-year period that included the second, third and fourth quarters of 2019 and the first quarter of 2020. HEDIS metrics for diabetes were aggregated over the 12-month period and included in the analysis. Using the scoring system approved by NCQA, each practice was assigned a score (100 being the maximum score) and then ranked based on their score. The top 25% of practices (11 practices) were grouped as top-performing, and the bottom 25% of practices (11 practices) were grouped as low-performing, which comprised the sample of providers used in the analyses.
Management practices: Five components of management practices (variables) were collected using a combination of EMR and facility information. The 5 components included (1) the rate of traditional endocrine consults; (2) the rate of electronic endocrine consults (e-consults); (3) the rate of CDCES referrals; (4) the utilization of insulin and noninsulin injectable medications (glucagon-like peptide-1 receptor agonists) among patients with diabetes; and (5) the PCP practice location and its proximity to the endocrine office (distance in miles). Selection rates were based on the ratio of selected management practices relative to PCP annual office visits of patients with diabetes. The 5 components chosen were based on their hypothesized effects on functional and clinical outcomes associated with diabetes management.
Data Analyses: Both descriptive and inferential statistics were computed for the management practices. For continuous variables, t tests were conducted. For rate- and proportion-based variables (all referral- and medication-related variables), generalized linear models were used. Patients’ visit volume across performance status groups (ratio of patients to practitioners) was tested using a negative binomial model.
Phase 2: Insurance claims data
Inclusion and Exclusion Criteria: We identified a total of 3,170 members with diabetes who were seen by the 79 physicians that belonged to the top 11 and bottom 11 performing practices. Using the criteria of (i) an approved claim with a primary diagnosis of diabetes (ICD 10: E08.xx, E09.xx, E10.xx, E11.xx, E12.xx, or E13.xx.) between April 1st, 2019, and March 31st, 2020, (ii) twelve months of continuous enrollment during the aforementioned timeframe, and (iii) receiving treatment from one of 79 AHN providers participating in the study, we identified 3203 members for inclusion.
Member Characteristics and Patient Load: Demographic data were obtained from members’ enrollment information. We defined health status using the Charlson Comorbidity Index (CCI), which quantifies long-term mortality in individuals with multiple comorbidities.7–8 The CCI for each member was calculated using ICD-10 codes appearing on members’ claims from the study period. Type of diabetes diagnosis (type 1, type 2, secondary, both type 1 and type 2) and insulin use were derived from the presence of corresponding diagnosis codes and national drug codes (NDCs) in a member’s claims throughout the 12-month period. In addition to member characteristics, patient-level data were used to calculate the number of patients within each practice and the number seen by each physician. Note that due to the nature of using claims data, these patient load estimates reflect the number of patients who are Highmark members and not the total number of patients.
Cost and Utilization: Allowed amounts (i.e., negotiated costs for service) were used to calculate the total cost of diabetes-related care and all other healthcare for the twelve-month study period. Diabetes-related cost of care was calculated as the sum of all claims costs that fell into one of the following categories: medical claims that had diabetes as the primary or admitting diagnosis (ICD 10 E08.xx – E13.xx), diabetes medication claims or claims for diabetes-related durable medical equipment (DME), i.e., glucose monitoring supplies. All claims not meeting the diabetes-related criteria were summed to create the total cost of other care. Details on medications, DME, and CPTs used for categorization of diabetes-related spending and utilization are provided in the Supplementary Materials (eTables 1–3). Diabetes-related hospital utilization was identified using a combination of diagnosis codes, claim type codes (e.g., inpatient vs. professional claims), and CPTs from individual claims. Three types of diabetes-related hospital utilizations over the 12-month period were identified (coded as any vs. none per member): hospital observations, inpatient admissions, and emergency department use (eTable 3). Determination of hospitalizations as diabetes-related followed the same coding strategy described above for costs.
Data Analysis: Descriptive statistics were computed for patient characteristics (e.g., demographics, type of diabetes diagnoses) as a function of performance status (high vs. low). Differences in patient characteristics were evaluated using Wilcoxon rank sum tests, chi-squared tests, and Fisher’s exact tests as appropriate. Cohen’s d and Cramer’s V are provided to quantify the magnitude of group differences. Follow-up analyses to determine factors that may underlie differences in practices’ performance status consisted of comparisons of the number of physicians within practices, the number of patients seen per physician, patient’s health status (CCI scores), diabetes-related care costs, and diabetes-related hospital utilization (stratified by type). These patient and care-related factors were compared using Wilcoxon rank-sum tests (continuous variables) and Fisher’s exact tests (categorical variables).
Cost and utilization measures that differed by performance status were compared using logistic and quantile regression to evaluate differences while controlling for patient factors. Quantile regression was selected to relax assumptions of normality and allow specification of the part(s) of the outcome distribution to predict using percentiles. For example, specifying the 50th percentile allows prediction of the median. Quantile regression was well designed for the cost data because the distribution was both skewed and bimodal. The percentiles chosen for modeling corresponded to the two modes for the high-performance group. In all models, the primary predictor of interest was performance status; control variables included age (centered), sex, CCI score, presence of type 1 diabetes diagnosis (yes/no), insulin use (any/none), and members spend on nondiabetes-related care (in $100 units).
Analyses were conducted using RStudio Workbench (version 1.4.1717-3) running R version 3.6.3 with an alpha level of 0.05 for significance. Analyses, tables, and figures were created using the following packages: ggplot2, ggridges, effsize, rstatix, quantreg, marginaleffects, performance, and gtsummary.9–16