Perioperative hypoxaemia early warning based on high-frequency waveform fusion and deep learning

Area of Science:

• Anesthesiology and Critical Care Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Current intraoperative hypoxaemia prediction models using structural parameters have limitations like low sampling frequency and high data attrition.
• High-frequency waveform data offers potential for enhanced predictive accuracy and robustness.

Purpose of the Study:

• To develop and evaluate a deep learning-based hypoxaemia prediction model utilizing long-term, high-frequency waveform data.
• To address the shortcomings of conventional models and improve predictive accuracy and robustness for patient safety.

Main Methods:

• Extracted respiratory, blood oxygenation, and ECG waveforms from the VitalDB database.
• Constructed numerical and image-based datasets using waveform-derived variables, Gramian Angular Summation Field (GASF), and four-quadrant matrix methods.
• Employed a Long Short-Term Memory (LSTM) algorithm with a 5-minute learning window and 1-13 minute prediction windows for evaluation.

Main Results:

• Optimal model performance was observed within a one-minute prediction window.
• The 'Multimodal image model' showed superior accuracy (0.932) and AUC (0.953), but singlemodal models using raw waveform data were sufficient.
• Singlemodal waveform-only models demonstrated enhanced computational efficiency and met clinical prediction requirements.

Conclusions:

• High-frequency waveform data significantly enhances stability and performance in hypoxaemia prediction models.
• Waveform processing methods like GASF and four-quadrant matrix are effective.
• Singlemodal models utilizing only raw waveform data offer a computationally efficient and clinically viable approach for perioperative hypoxaemia prediction.