Time-resolved modelling of the perioperative course
Using intraoperative vital-sign dynamics and the timing of events to tell when, during and after surgery, risk can be acted on.
Modern operating rooms record vital signs continuously, yet those signals are rarely used once the operation ends. This work treats the perioperative course as a time series. It asks how much risk is already visible at the end of surgery, and which complications belong to which time window.
From 10,719 surgical procedures, we turned arterial blood pressure, heart rate, oxygen saturation, temperature and end-tidal CO2 into interpretable summary, trend and distribution features. A model built on these features predicted postoperative infection at the end of surgery with an AUROC of 0.88, clearly better than preoperative variables alone. Its predictions stayed calibrated across the major procedure groups.
A follow-up analysis separates two windows that call for different responses: early death and later infection.
Publications
- Blatter, T. U., Wintsch, Y., Triep, K., Endrich, O., Guillen-Ramirez, H.*, & Beldi, G. End-of-surgery prediction of postoperative infectious complications from intraoperative vital-sign dynamics. npj Digital Medicine (2026). *Shared senior authorship. doi:10.1038/s41746-026-02707-1
- Lucas, K. L. et al. Two actionable windows: disentangling early mortality from late infection risks using time-resolved AI. British Journal of Surgery 113(Suppl. 5) (2026). Congress abstract. doi:10.1093/bjs/znag055.009