Efficient Adverse Event Forecasting in Clinical Trials via Transformer-Augmented Survival Analysis

With the in-depth application of generative AI in clinical statistical processes, the TFL (Table-Figure-Listing) automation platform and macro system have significantly shortened the reporting cycle and improved data quality with unsupervised anomaly detection, laying a clean data foundation for adverse event risk modeling for time-event prediction. On the basis that AI-driven TFL automation and outlier cleaning have significantly improved data quality, we propose Segmented Relative Positional Encoding-Transformer Survival Network (SRPE-TSN): this method only introduces the key improvement of "segmented relative time embedding" on existing Transformer survival models such as SurvTRACE. The longitudinal event sequence of patients was divided into learnable time periods according to clinical milestones, and the relative position information was used to guide multi-head attention to focus on risk signals at different time scales, so as to take into account both right-censored processing and long-term dependency capture. SRPE-TSN increased the 12-month adverse event time-dependent AUC from 0.71 to 0.80 on data from four phase III oncology and cardiovascular trials.