Beyond Thresholds: Learning Perioperative Risk from Intraoperative Physiological Trajectories

Sadaf Sohrabi, Shehan Hettiaratchy, Aldo A. Faisal

Keywords: Perioperative Risk Prediction, Intraoperative Physiology, Clinical Time Series, Surgical Outcomes, Electronic Health Records, Hidden Markov Models, Representation Learning

TL;DR: Intraoperative physiological trajectories improve perioperative risk prediction beyond static and threshold-based models, with gains largest in patients whose static risk profile appears reassuring.

Abstract: Clinical risk scores for surgical patients are computed once from preoperative variables and consequently fail to capture the physiological dynamics that unfold during surgery. We ask whether intraoperative vital-sign time series, routinely recorded but rarely used for real-time risk stratification, provide predictive information that static models fail to capture. Using 130,960 operations from the INSPIRE perioperative dataset, we compare four modelling arms across progressively richer feature regimes (static baseline, physiological summaries, HMM-derived trajectory features, and their combination) and evaluate performance across four postoperative outcomes using grouped cross-validation. We find that learned trajectory representations capture complementary prognostic structure to threshold-based summaries for ICU-related outcomes. Predictive signal is detectable as early as 15 minutes after anaesthesia onset, and gains are largest in patients whose static risk profile underestimates true intraoperative vulnerability. These results demonstrate that temporal modelling of multivariate intraoperative time series provides clinically meaningful information that threshold-based approaches systematically miss.

Submission Number: 54