Dynamic Prediction of Overall Survival for Patients with Advanced Non-Small Cell Lung Cancer

• Activity Number: 372 - SPEED: SPAAC SESSION IV
• Type: Topic-Contributed
• Date/Time: Thursday, August 12, 2021 : 12:00 PM to 1:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Xuechen Wang* and Benjamin Haaland and KATHLEEN KERRIGAN and Sonam Puri and Wallace Akerley
• Companies: University of Utah, School of Medicine, Department of population health science and University of Utah, School of Medicine, Department of population health science and University of Utah, Huntsman Cancer Institute and University of Utah, Huntsman Cancer Institute and University of Utah, Huntsman Cancer Institute
• Keywords: Time-dependent clinical factor; landmarking; dynamic prediction; IPCW AUC; model-based calibration

Abstract

For patients with advanced cancer, it is standard clinical practice to estimate prognosis in order to inform treatment decisions. However, clinical estimates of prognosis are not precise and rapidly become out-of-date if clinical factors which evolve over time are not directly incorporated. With the development of and access to EHR-derived data, dynamic prediction methods that use structured data are critical to accurately and robustly characterize the evolving relationship between patient factors and outcomes, handle a shrinking at-risk population over time, and make predictions at any follow-up time. Joint modeling and landmarking are commonly used for dynamic prediction. However, the joint modeling approach requires strong model assumptions and substantial computation, and is not applied as widely as landmarking in practice. Using data on patients with advanced NSCLC from the Flatiron Health EHR-derived de-identified database, we propose a spline smoothed landmarking approach to dynamically estimate survival probabilities. Inverse probability of censoring weighted AUC improves dramatically compared to models based on only baseline information.