Related Experiment Videos
Transformer-based deep learning model for real-time prediction of intraoperative hypotension using dynamic
Shouqiang Zhu1, Weihai Shi1, Haichuan Qian2
1Department of Anesthesiology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
Background:
The clinical importance of transient intraoperative hypotension (IOH) remains debated, and existing models often rely on high-resolution waveform data that are not routinely available.
Methods And Findings:
We developed a Transformer-based deep learning model to predict IOH in real time using continuous vital sign time-series data. The model was trained on 319,699 surgical cases from a tertiary hospital in China (2013-2023) and externally validated using an independent dataset from South Korea. Model interpretability was explored through a real-time alert simulation using 10 representative surgical cases from the internal validation cohort, comparing predicted IOH risk trajectories with measured mean arterial pressure (MAP). To assess clinical relevance, a nested cohort study evaluated the association between IOH burden (cumulative MAP ≤65/60/55 mmHg in mmHg·min) and postoperative acute kidney injury (AKI) and acute kidney disease (AKD). The Transformer model achieved strong prediction performance at 5-, 10-, and 15-min horizons (AUCs 0.904, 0.892, 0.882; recall ≥88.3%). Compared with XGBoost, the Transformer had higher recall (internal 5-min recall 0.891 versus 0.737) and substantially better probability calibration (expected calibration error 0.0083 versus 0.0373). XGBoost showed higher overall accuracy and specificity (internal 5-min specificity 0.913 versus 0.723). External validation confirmed comparable discrimination across models and generalizability, with attenuated calibration differences. In alert simulations, predicted IOH risk closely corresponded to MAP fluctuations. IOH burden was significantly associated with postoperative AKI and AKD (MAP ≤65 mmHg: OR per 60 mmHg·min 1.10 (95% CI, [1.02, 1.19]; p = 0.012) for AKI; 1.26 (95% CI, [1.19, 1.33]; p < 0.001) for AKD). The study is limited by its retrospective design, and prospective, multicenter validation is needed to confirm real-time applicability and generalizability of the model.
Conclusions:
IOH burden is associated with increased risk of postoperative AKI and AKD. The Transformer model prioritizes sensitivity and calibration for short-term IOH prediction, whereas XGBoost emphasizes accuracy and specificity, reflecting different operating characteristics. Prospective, real-time evaluation is needed before clinical implementation.