Predictive Modeling of Hypotension during General Anesthesia Using Machine Learning Algorithms on High-Frequency Physiological Time-Series Data

Intraoperative hypotension during general anesthesia represents a grave clinical complication that can result in severe adverse patient outcomes, including acute kidney injury, myocardial infarction, and increased postoperative mortality. Although prompt forecasting is absolutely essential for timely clinical intervention, conventional monitoring approaches often rely on periodic, intermittent measurements, which inherently delays the detection of critical hemodynamic deterioration. Furthermore, current predictive computational frameworks struggle with instantaneous forecasting and consistently fail to fully exploit the rich, dynamic information embedded within high-frequency time-series signals. These existing models also frequently suffer from poor explainability and limited generalizability when applied across heterogeneous patient cohorts in real-world clinical settings. To systematically address these critical issues, this study introduces a novel joint deep learning architecture integrating Long Short-Term Memory (LSTM) networks and advanced attention mechanisms to accurately forecast anesthetic hypotension utilizing high-frequency physiological parameters. The proposed methodological framework exhibits highly robust generalization capabilities, sustaining remarkably stable predictive performance when rigorously validated on several independent, external hospital datasets. On the primary evaluation cohort, the integrated model achieved an impressive accuracy of 0.87 ± 0.02 and an Area Under the Curve (AUC) of 0.91 ± 0.01, statistically outperforming traditional baseline models (p-value