Purpose: Effective clinical decision support (CDS) tools embedded in a practice's electronic medical record (EMR) depend on within-practice learning and information sharing. Cluster-randomized trials (CRTs) are the natural study design to evaluate EMR-centered CDS. In any CRT, we desire well-balanced clusters for fair and powerful comparisons. EMR data can be used to identify subjects and explicitly balance important characteristics before assigning practices to clusters. We used historical EMR data to balance prognostically important characteristics and outcomes at baseline in a CRT of EMR-centered CDS. Methods: The Diabetes Improvement Group – Intervention Trial (DIG-IT) compares the impact of two CDS alternatives on care and outcomes for 7,105 continuity diabetic patients in 10 diverse practices in an urban healthcare system. For the two years before the trial, we used the EMR to assemble practice-level data on pre-specified clinical, demographic, and utilization characteristics, including diabetes-related trends. We then examined balance in each possible clustering of the practices into two groups, selected the clustering that best balanced these characteristics historically, then randomly assigned interventions to groups. We use the intra-cluster correlation coefficient (ICC) to measure balance between clusters. Large ICC values indicate greater imbalance, implying substantial reductions in effective sample size. We report baseline balance of key covariates and outcomes in the analytic cohort identified at the trial's end. Results: In all, 2742 patients received usual EMR-based care and 4363 received the CDS interventions. Imbalances at study entry were small for most key covariates and outcome levels, with ICCs below 0.004 for hemoglobin A1c, LDL cholesterol, systolic BP, BMI, current smoking, age, gender, and rates of “no shows” and ED visits. Larger but modest imbalances (ICC near 0.02) appear in race/ethnicity and insurance status, and will require analytic adjustments. For DIG-IT's two main outcomes, glycemic control (A1c £ 7.0) and a guideline-based composite score, ICCs were < 0.0001 and 0.0025, translating to effective sample sizes of approximately 4300 (for A1c) and 700 (for the composite.) Conclusions: In a CRT of EMR-centered CDS effectiveness, explicitly balancing practice characteristics within study groups before assignment reduced substantially the potential for selection bias, the need for extensive covariate adjustment, and the impact on effective sample size. Increasing EMR adoption, especially in health systems with multiple practice sites, should facilitate better CRT designs and analyses.