A0748
Title: Leveraging longitudinal data for enhanced survival analysis using a novel deep transformer model
Authors: Pingzhao Hu - Western University (Canada) [presenting]
Abstract: Traditional statistical methods, while capable of analyzing the quality of life and toxicity data individually, often struggle with the efficiency and integration required for comprehensive multi-dataset analyses in clinical settings. This limitation hampers the long-term understanding of patients. This research develops an innovative model using a modified transformer encoder framework enhanced with an attention-free transformer (AFT). This model is engineered to concurrently process and correlate clinical variables with longitudinal assessments of quality of life and toxicity. Utilizing a transformer encoder architecture optimized with AFT, time-series clinical data is efficiently processed. This framework is integrated with Cox regression analysis, and measured via the concordance index. A key highlight of the model is its adeptness at integrating diverse datasets, including cross-sectional and longitudinal datasets, while adeptly managing variations in measurement times and frequencies across different subjects. Tested on a longitudinal dataset of 750 colorectal cancer patients over a two-year period, the model outstripped conventional analytical methods in predictive accuracy and integration, as evidenced by a substantial enhancement in the concordance index. The findings highlight the models potential to transform clinical decision-making processes, setting the stage for further exploration into its clinical implications and adaptability.
