A1699
Title: Dynamic time-to-event prediction with ML/DL: Addressing competing risks in clinical outcomes
Authors: Li Tang - St. Jude Children's Research Hospital (United States) [presenting]
Abstract: With the growing availability of longitudinal data from electronic medical records (EMR), integrating this data with baseline variables for dynamic prediction of clinically significant time-to-event outcomes has become a research priority. Previous studies have explored machine learning and deep learning models such as random survival forest, DeepSurv, and Transformers to incorporate multivariate longitudinal data into survival prediction. While these approaches show promise, the impact of competing risks common in survival analysis remains underexplored. Our study addresses this gap by evaluating these models under competing risks through simulations. We found that prediction accuracy is sensitive to landmark times, prediction windows, and interactions between variables. We then applied these refined methods to predict acute graft-versus-host disease (aGVHD) in pediatric oncology patients undergoing allogeneic hematopoietic cell transplantation. Using high-dimensional longitudinal EMR data, our approach significantly improves GVHD risk prediction, enabling early identification of high-risk patients and facilitating tailored prophylaxis strategies. We demonstrate the potential of AI-driven clinical risk prediction and underscore the need for careful application to ensure these tools truly enhance patient care.
