They developed a transformer-based deep learning model to predict real-time intraoperative hypotension (IOH) from dynamic time series of vital signs. They trained the model on 319,699 surgical cases from a Chinese hospital (2013–2023) and externally validated with data from South Korea. It achieved high performance in 5-, 10-, and 15-min predictions with AUCs of 0.904, 0.892, and 0.882, respectively, at a recall of ≥88.3%. Compared to XGBoost, it had better recall (5-min recall 0.891 vs. 0.737) and calibration (expected error 0.0083 vs. 0.0373), although XGBoost had higher specificity (0.913 vs. 0.723). IOH burden (cumulative MAP ≤65 mmHg) is significantly associated with postoperative acute kidney injury (AKI) and acute kidney disease (AKD), e.g. OR 1.10 for AKI at 60 mmHg·min (95% CI [1.02, 1.19]; p=0.012). The predicted risk of IOH in the simulations closely followed the variation in MAP. The study is retrospective, requiring prospective multicenter validation.