Proceedings of the 2007 Winter Simulation Conference S. G. Henderson, B. Biller, M.-H. Hsieh, J. Shortle, J. D. Tew, and R. R. Barton, eds. A DISCRETE EVENT MODEL OF CLINICAL TRIAL ENROLLMENT AT ELI LILLY AND COMPANY Bernard M. McGarvey Nancy J. Dynes Burch C. Lin Wesley H. Anderson James P. Kremidas James C. Felli Eli Lilly and Company Lilly Corporate Center Indianapolis, IN. 46285, U.S.A. ABSTRACT Clinical trials constitute large, complex, and resource in- tensive activities for pharmaceutical companies. Accurate prediction of patient enrollment would represent a major step forward in optimizing clinical trials. Currently models for patient enrollment that are both accurate and fast are not available. We present a discrete event model of the pa- tient enrollment process that is accurate and uses relatively small CPU times. This model is now being used on a regu- lar basis to predict the enrollment of patients for large trials with around 13,000 patients and has led to significant re- duction in the time it takes to make these predictions. 1 INTRODUCTION The management of clinical trials used to determine the safety and efficacy of new pharmaceutical products repre- sents a major investment in resources for pharmaceutical companies. Such clinical trials consist of a number of complex and interdependent tasks, including: protocol de- velopment and approval, trial site selection, and patient identification, selection, and enrollment. Clinical trials also represent a major opportunity for pharmaceutical compa- nies to optimize the overall new drug approval process. Some of the issues that can occur in the clinical trial pa- tient enrollment process include: • Delay of study completion due to poor enrollment. Approximately 80% of clinical trials across the industry do not complete enrollment as planned, resulting in increased clinical operations expenses, cancelled trials, and loss of future revenues from delayed submissions. • Over-enrollment of patients to provide a safety factor. Each extra patient that adds to the cost of a trial without improving the statistical value of the analysis results in excess and unnecessary ex- penses. Subsequently such patients may be ex- cluded from the trial and so the resources used to enroll these people are wasted. • Mismatch of supply and demand for resources used in the trial. When resources such as clinical trial material are not synchronized with the avail- ability or location of enrolled patients, wasted ma- terials or delayed study completion can result. Felli et al. (2007) give a good discussion of current ef- forts to improve the patient enrollment process for clinical trials. In particular, the ability to predict the enrollment rate of patients in a clinical trial is presented as an important element of improving the clinical trial process overall. A theoretical model based on semi-Markov chain analysis developed at Eli Lilly (Felli et al. 2007) showed that mod- eling patient enrollment can lead to useful predictions. This model allowed the prediction of total patient enrollment rates based on the rates of enrollment at individual sites, anticipated patient drop out rates and so on. After this initial success, focus groups were created at Eli Lilly to evaluate the usefulness of the general model as a practical tool for the ongoing prediction of clinical trial enrollment. The main issue identified by these groups con- cerned the model execution time. Even for a relatively small clinical trial involving 700 patients over a 300 day simulation time horizon took about 2 ¾ hours on a desktop computer. Since the original vision for using the prediction model would allow users to do what-if scenario analysis as they sat at their desktop computers, such a long execution time was clearly unacceptable. Based on discussions with users the following goals for execution times were set: • small trials (<1000 patients) should take on the order of 5 minutes or less to simulate • large trials (>10,000 patients) should take on the order of 30 minutes or less to simulate. 1467 1-4244-1306-0/07/$25.00 ©2007 IEEE