Frequency Domain Deep Learning With Non-Invasive Features for Intraoperative Hypotension Prediction

Area of Science:

• Anesthesiology and Critical Care Medicine
• Biomedical Engineering
• Data Science in Healthcare

Background:

• Intraoperative hypotension poses a risk for postoperative organ dysfunction.
• Traditional methods rely on invasive arterial pressure, limiting biosignal integration and analysis of signal periodicity.
• Deep learning advances highlight the utility of frequency-domain information for biosignal analysis.

Purpose of the Study:

• To develop a deep learning approach integrating multiple biosignal modalities using frequency-domain information.
• To address limitations of conventional time-domain analysis in predicting intraoperative hypotension.
• To enhance the interpretability of deep learning models for clinical application.

Main Methods:

• Extracted frequency information using discrete Fourier transform, combined with time-domain data as input for a deep learning model.
• Incorporated interpretable deep learning modules to enhance result transparency.
• Utilized 75,994 data segments from 3,226 patients for model training and validation.

Main Results:

• The proposed frequency-domain deep learning model outperformed conventional time-domain methods.
• Achieved superior Area Under the Receiver Operating Characteristic curve (AUROC) performance with non-invasive data (0.898 vs. 0.853).
• Identified the 1.5-3.0 Hz frequency band as crucial for predicting hypotension events.

Conclusions:

• Frequency-domain analysis demonstrates high performance with invasive and non-invasive biosignals.
• The developed framework provides a novel perspective for predicting intraoperative hypotension.
• This approach offers significant performance improvements, particularly for non-invasive monitoring.