结构化电子健康记录基础模型中的标记化权衡

Foundation models for structured electronic health records (EHRs) are pretrained on longitudinal sequences of timestamped clinical events to learn adaptable patient representations. The process of tokenization determines how these timelines are converted into discrete model inputs, influencing what information is preserved, the efficiency of encoding, and the relationships that need to be learned versus precomputed. However, the impact of tokenization design choices on downstream performance and computational efficiency remains largely unexplored. This study pretrained a transformer model on pediatric EHR data using a factorial design, varying tokenization along event encoding, time encoding, and workflow annotation. The effects of different tokenization strategies were analyzed by evaluating the area under the receiver operating characteristic curve.