A new artificial intelligence model can predict dangerous drops in blood pressure during surgery up to 15 minutes before they occur, according to a study published in PLOS Medicine. The Transformer-based deep learning system was trained on data from 319,699 surgical cases at a tertiary hospital in China spanning a decade, and then independently validated using patient data from South Korea.
Intraoperative hypotension — when blood pressure falls to potentially harmful levels during a procedure — has long been a concern for anesthesiologists. The condition is associated with serious complications, including acute kidney injury. Yet existing prediction tools often depend on high-resolution waveform data that most operating rooms don't routinely collect.
The new model works differently, relying on standard continuous vital sign readings already captured during surgery. It achieved strong predictive accuracy at 5-, 10-, and 15-minute horizons, with area-under-the-curve scores ranging from 0.882 to 0.904 and recall rates above 88%. In simulated real-time alert scenarios using 10 representative cases, the system's predicted risk trajectories closely tracked actual fluctuations in mean arterial pressure.
Researchers also compared the Transformer model against XGBoost, another machine learning approach. The Transformer excelled at sensitivity and probability calibration, meaning it was better at correctly flagging patients who would experience hypotension and at estimating how likely those episodes were. XGBoost, meanwhile, showed higher overall accuracy and specificity. The two approaches reflect different clinical tradeoffs — catching every dangerous episode versus minimizing false alarms.
The study also reinforced why predicting these events matters. In a nested analysis, researchers found that cumulative exposure to low blood pressure during surgery was significantly linked to postoperative acute kidney injury and acute kidney disease. However, the authors cautioned that the research is retrospective. Prospective trials in multiple hospitals will be needed before the system can be deployed in real operating rooms.
