PURPOSE: Clinical trials often exceed their anticipated enrollment periods, and study sites often do not meet accrual goals. We previously reported the development and validation of a single-site accrual prediction model. Here, we describe the expansion of this methodology at 16 cancer centers (CCs) and compare an overall model versus site-specific models.
METHODS: This retrospective cohort study used data from treatment and supportive care intervention studies permanently closed to accrual between 2009 and 2015 at 16 United States-based CCs. Center and ClinicalTrials.gov data were used to generate both site-specific and random effects mixed models (random effect: institution). Accrual predictions were generated from each model and compared with the accrual prediction of the disease team (DT).
RESULTS: Sixteen institutions submitted 5,787 eligible trials (range, 93 to 697 trials per institution). Local accrual ranged from 363 to 6,716 participants; 1,053 studies (18%) accrued no participants. Actual average accrual was 8.5 participants (median, four participants). Site-specific models predicted accrual at 99% of actual and correctly predicted whether a study would accrue four or more participants 73% of the time versus DT prediction of 58%. Correlation at the category level was 30%; model sensitivity and specificity were 83% and 62%, respectively. The overall model predicted accrual 93% of actual and correctly predicted accrual of four or more participants 66% of the time, with a correlation at the category level of 28%.
CONCLUSION: Both regression models predicted clinical trial accrual at least as or more accurately than DT at all but one center. Site-specific models generally performed slightly better than the random effects model. This study confirms the previous finding that this method is an accurate and objective metric that can be easily implemented to improve clinical research resource allocation across multiple centers.