Deep learning approach for stroke volume variation estimation: retrospective modeling and prospective deployment in...

Area of Science:

• Anesthesiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Stroke volume variation (SVV) is a key indicator of fluid responsiveness in patients.
• Traditional methods for SVV estimation using arterial blood pressure (ABP) can be limited by noise and inter-individual variability.
• Deep learning (DL) offers a potential avenue for improving SVV estimation accuracy and reliability.

Purpose of the Study:

• To develop and validate a deep learning-based model, SVVNet, for estimating SVV from arterial blood pressure (ABP) waveforms.
• To compare the performance of SVVNet against traditional ABP-based pulse wave analysis (ABP-PWA) algorithms.
• To assess the clinical applicability of SVVNet through retrospective and prospective real-time evaluations in intraoperative settings.

Main Methods:

• A SVVNet model was trained on ABP waveforms from 2849 patients undergoing general anesthesia.
• Retrospective validation involved large cohorts to assess agreement, trending, noise robustness, and inter-individual deviation against ABP-PWA and other DL models.
• Prospective validation was conducted in real-time during 67 surgeries using a tablet-deployed model.

Main Results:

• SVVNet demonstrated higher agreement (MAE=0.70, ρ=0.92) than ABP-PWA (MAE=1.30, ρ=0.80) in retrospective external validation.
• The DL model exhibited greater robustness to signal noise (MAE=4.33 vs 30.87) and lower inter-individual deviation.
• Prospective real-time deployment achieved an MAE of 1.33 and a Pearson correlation of 0.87 against the ground truth.

Conclusions:

• The deep learning-based SVVNet model offers significant advantages over traditional ABP-PWA algorithms for SVV estimation.
• Prospective results confirm the clinical applicability and reliability of SVVNet in the operating room environment.
• SVVNet has the potential to enhance intraoperative monitoring and replace existing systems.