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

Continuous arterial blood pressure (ABP) monitoring is essential for real-time hemodynamic assessment but is limited by the invasiveness of arterial catheterization. We propose a ResUNet-based deep learning framework that generates high-fidelity ABP waveforms from photoplethysmogram (PPG) signals, incorporating periodic calibration values applied every 2.5 minutes. Trained on intraoperative data from 4,687 surgical cases at Seoul National University Hospital, the calibrated model achieved a mean absolute error (MAE) of 5.69 mmHg with a standard deviation (SD) of 3.76 mmHg, outperforming the non-calibrated model (MAE: 9.47 mmHg [SD: 5.52]). It satisfied both AAMI and BHS standards, achieving Grade B for SBP and Grade A for DBP. This method enables accurate non-invasive ABP estimation and supports clinical deployment in settings where invasive monitoring is infeasible.