Purpose: In cluster-randomized trials (CRTs), identifying clusters that are balanced on key features before assigning them to study groups is critical for fair and powerful comparisons. The key measure of balance in CRTs is the intra-cluster correlation coefficient (ICC), with better values being closer to zero, and larger values associated with significantly reduced power relative to a randomized controlled trial of the same size (where treatment allocation occurs at the patient level.) This report demonstrates the use of EMR data in executing a well-balanced design for a CRT of clinical decision support (CDS).

Methods: The Diabetes Improvement Group – Intervention Trial (DIG-IT) compares the impact of alternative approaches to CDS on the care and outcomes of 15,000 diabetic patients cared for by 220 primary care physicians (PCPs) in 24 diverse clinical practices of 2 health care systems. DIG-IT interventions apply to practices, but outcomes (diabetes-related parameters, rates of appropriate tests, treatments, immunizations, and resource use) occur at the patient level. We considered all feasible groupings of System A's 10 practice sites into two clusters and System B's 14 sites into three clusters. Before assigning clusters to study groups, we assembled practice-level data from each System's EMR to examine these groupings along an array of baseline clinical, demographic, and utilization characteristics, including trends in diabetes-related parameters. Blinded to site identifiers, DIG-IT investigators reached consensus about the groupings that most effectively balanced these pre-specified characteristics. Interventions were then randomly allocated to clusters in those groupings.

Results: System A's 10 practices were split into two clusters of 5 practices (pre-intervention, 2085 and 2281 patients, respectively); System B's 14 practices were split into three groups of 4, 6, and 4 practices (2069, 3115, and 3185 patients). ICCs for most pre-specified characteristics within each System were below 0.005, including A1c trend and mean, last systolic blood pressure, current smoking, "no show" rates, and percent with ED visits in the past year. The achieved balance was superior to that of potential groupings by demographics alone, by geography, or unstratified practice-level randomization, and it also was excellent for characteristics not explicitly accounted for in the design.

Conclusions: Rich data available through EMRs provide opportunities for designing state-of-the-art community-based CRTs, including health care delivery trials of clinical decision support.