An Event-Aware Dual Representation Model With Mixture-of-Experts for Serious Adverse Events Prediction in Clinical Trials

Download PDF
Open Webpage

Baoshuo Kan, Teng Wang, Hengdong Zhu, Rongjiao Liang, Enliang Yan, Fu Lee Wang, Tianyong Hao

Published: 01 Jan 2025, Last Modified: 11 Nov 2025IEEE Trans. Consumer Electron. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: With the growth of the healthcare consumer electronics market, the clinical trial industry is facing new opportunities and there is an increasing focus on safety to protect patients from potential risks. Serious adverse events (SAEs) in clinical trials may pose significant safety threats to patients and incur substantial economic losses. Therefore, predicting and preventing SAEs has become a critical issue in clinical trial research. However, the lack of standardized serious adverse event data limits studies to specific conditions (e.g., target diseases or trial phases), resulting in insufficient general applicability. To address this challenge, this paper aggregates trial protocol, drug, and disease information from multiple data sources to create a universal dataset for Serious Adverse Events prediction (SerAE). The dataset encompasses 10,643 clinical trials, 4,512 diseases, and 2,563 drugs. Furthermore, an Event-aware Dual Representation model with mixture-of-Expert (EDRE) for serious adverse event prediction is proposed to achieve universal prediction of SAEs. Specifically, drug molecules, target diseases, and trial eligibility criteria are encoded to acquire respective representations to construct dual representations, aiming to comprehensively utilize various factors that may trigger SAEs both before and during the trials. Subsequently, the dual representations and relevant knowledge of event categories are input into an event-aware module to aggregate salient event clues to enhance the representations. Finally, a mixture-of-experts classifier is designed to simulate expert interactions in multidisciplinary consultations, providing comprehensive predictions of SAEs. The experiments on SerAE demonstrate that EDRE significantly enhances the performance in predicting SAEs compared to state-of-the-art baseline models.