Validity of EHR-Derived Phenotypes for Outcome Ascertainment in the PRESERVE Trial: Implications for Embedding RCTs in Healthcare Systems

Randomized controlled trials (RCTs) are crucial for determining effective interventions but are costly, time-consuming, and complex. Electronic health records (EHRs) offer the potential to enhance research using clinical data, reducing costs and improving efficiency. We evaluated the accuracy of EHR-derived phenotypes for endpoint identification versus RCT-outcome ascertainment in the PRESERVE trial (NCT01467466), comparing sodium bicarbonate vs. sodium chloride and oral acetylcysteine vs. placebo in preventing serious adverse outcomes and contrast-associated acute kidney injury (CIAKI) after angiography. PRESERVE outcomes were compared with EHR data from the Veterans Affairs Corporate Data Warehouse. Real-world data (RWD) phenotypes captured serious adverse events within 90 days post-angiography, including death, dialysis need, and persistent kidney decline. Sensitivity, specificity, and percent agreement were calculated using trial outcomes as the gold standard. Among 4,196 PRESERVE patients, 95% had outcomes assessed. RWD outcomes showed mixed accuracy: 100% sensitivity/specificity for death; high sensitivity for primary composite events (71.2%), dialysis (75.0%), and hospitalization (78.1%); and low sensitivity for CIAKI (45.8%) and persistent kidney decline (40.5%). Substituting RWD outcomes into primary analysis produced conclusions matching the RCT conclusions. EHR phenotypes can substitute for traditional outcome collection, given adequate planning and validation. Despite variable accuracy, RWD outcomes yielded similar conclusions to RCTs, supporting their use for trial efficiency and reducing costs.