|
Abstract:
|
Real-time projections of analysis times based on pooled blinded time-to-event data of on-going trials have important operational and logistical implications, such as planning for the timing of data safety monitoring committee meetings, interim and final analyses. Data mining of clinical trial data provides the opportunity to develop and apply new predictive methods; here we do so for prediction of analysis times. A proportional hazards (PH) model often forms the basis for the predictions, but PH models are often ill-equipped to address non-PH conditions. The delayed separation of Kaplan-Meier curves and durable long-term responses are examples of non-PH, and have been observed in event-driven immuno-oncology studies. Prediction of analysis times using established methods for non-PH scenarios often suffer a lack of accuracy, precision, and applicability. We propose and investigate new approaches and compared them with an existing methodology across a variety of scenarios drawn from synthetic data, and from actual Immuno-Oncology trials data at BMS. The comparisons suggest improved accuracy, precision, and applicability over conventional methods in non-PH scenarios.
|
