Machine Learning Methods for the Prediction of Intraoperative Hypotension with Biosignal Waveforms

Area of Science:

• Anesthesiology
• Artificial Intelligence
• Medical Informatics

Background:

• Intraoperative hypotension (IOH) is a critical factor linked to adverse postoperative outcomes, including myocardial infarction, acute kidney injury, and mortality.
• Predicting and preventing IOH is crucial for improving patient safety during non-cardiac surgery.

Purpose of the Study:

• To develop and validate machine learning (ML) and deep learning (DL) models for predicting intraoperative hypotension.
• To utilize a combination of intraoperative biosignals and personalized clinical information for accurate IOH prediction.

Main Methods:

• A retrospective observational study using the VitalDB open dataset, including 2611 patients undergoing non-cardiac surgery.
• Development and validation of ML (Gradient Boosting Machine) and DL (CNN-RNN) models using four waveforms (arterial blood pressure, ECG, PPG, capnography) and clinical data.
• Model performance evaluated 5 minutes prior to a hypotensive event.

Main Results:

• Both Gradient Boosting Machine and the hybrid CNN-RNN model achieved a high Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.94.
• The models demonstrated strong predictive accuracy, with an overall accuracy of 0.88 for predicting IOH.
• The models successfully predicted IOH using waveform data and clinical covariates.

Conclusions:

• Validated ML and DL models can effectively predict intraoperative hypotension using biosignal and clinical data.
• These predictive models hold promise for real-time IOH detection in operating rooms.
• The study's findings are expected to contribute to reducing the incidence of IOH and associated patient morbidity.