Methods for comparing community based/real-world chronic disease interventions using electronic health record data

Introduction: Statistical techniques applied to data from real world primary care patients offer a guide for treatment choices among patients with multiple chronic conditions.  The assembly of a large sample of type II diabetic patient data from primary care electronic health records allowed analysis of a broad range of patients across combinations of demographics, co-morbidity, and concurrent medications.

Objectives:  For this study we developed a database in a real world primary care population while applying methods that resulted in an unbiased sample.

Methods:  Propensity score matching and adjustment for non-randomization of treatments were employed; the analysis considered large numbers of covariates, interventions, and outcomes. This database of type 2 diabetes patients included clinical data, physical and lab measures, concomitant medications, and International Classification of Disease, 9^th revision coded medical problems. Methods of data cleaning included the review of frequencies to find extreme outlier values and combining similar categories to avoid small cell sizes.

Results: A cohort of 4,040 type 2 diabetes patients yielded statistically significant results when comparing the change in key indicator lab values across four oral treatment regimens. In covariate adjusted models, a diagnosis of hypertension made a significant contribution to predicting a change in HbA1c (p=0.0016) and the effective medication regimens differed among patients with and without hypertension.

Conclusions: Having a diagnosis of hypertension appeared to impact the change in HbA1c, which differed by treatment regimen.  The more effective regimen for T2DM appeared to be MET-TZD for patients with hypertension and MET-SU for patients without hypertension.