Deep Learning-Based Arterial Blood Pressure Waveform Generation Using Photoplethysmography for Non-Invasive...

Area of Science:

• Biomedical Engineering
• Medical Devices
• Artificial Intelligence in Medicine

Background:

• Continuous arterial blood pressure (ABP) monitoring is crucial for hemodynamic assessment.
• Current invasive methods using arterial catheterization pose risks and limitations.
• Non-invasive alternatives are needed for widespread clinical application.

Purpose of the Study:

• To develop and validate a deep learning framework for non-invasive ABP waveform generation from photoplethysmogram (PPG) signals.
• To improve the accuracy and reliability of non-invasive ABP monitoring.
• To assess the clinical feasibility of the proposed method.

Main Methods:

• A ResUNet-based deep learning model was developed to derive ABP waveforms from PPG signals.
• Periodic calibration values were incorporated into the model every 2.5 minutes.
• The model was trained and validated on a large dataset of 4,687 intraoperative cases from Seoul National University Hospital.

Main Results:

• The calibrated deep learning model achieved a mean absolute error (MAE) of 5.69 mmHg (SD: 3.76 mmHg) for ABP estimation.
• This significantly outperformed the non-calibrated model (MAE: 9.47 mmHg, SD: 5.52 mmHg).
• The model met AAMI and BHS standards, achieving Grade B for systolic blood pressure (SBP) and Grade A for diastolic blood pressure (DBP).

Conclusions:

• The proposed deep learning framework enables accurate, non-invasive estimation of arterial blood pressure.
• Periodic calibration enhances the fidelity of ABP waveforms generated from PPG signals.
• This technology supports clinical deployment in various settings, especially where invasive monitoring is not feasible.