Modern surgical care generates heterogeneous data spanning continuous intraoperative physiology, postoperative laboratory measurements, and complex procedural classifications. A key challenge is how to represent and integrate these data in a way that preserves clinically meaningful structure while enabling robust and interpretable learning.
We present a unified framework for perioperative risk modelling focused on postoperative infectious complications. Using large-scale electronic health record data, we show that representations of intraoperative vital-sign dynamics derived from irregular and noisy time series enable accurate prediction of postoperative infections already at the end of surgery, substantially outperforming models based on static preoperative variables alone. These representations capture physiologic instability through distributional, trend, and entropy-based features and remain interpretable using explainable machine learning.
We further extend prediction into the postoperative phase by modelling laboratory time series with imputation strategies that account for missing-not-at-random mechanisms driven by clinical decision-making. Integrating laboratory kinetics enables earlier and more accurate detection of infection than clinician-initiated treatment and supports procedure-specific laboratory testing recommendations.
Finally, we introduce MAP-CARE, a multilingual embedding framework that aligns national surgical procedure classifications into a shared semantic space using large language models. This representation layer enables cross-language retrieval and surgery-agnostic learning, providing a foundation for generalizable perioperative risk prediction.
