Practical Optimization of Deep Learning Models for Cuffless Blood Pressure Estimation From Photoplethysmography.

To improve deep learning-based cuffless blood pressure (BP) estimation from photoplethysmography (PPG) for continuous, non-invasive monitoring when cuff measurements are impractical, by assessing sampling frequency, systolic blood pressure (SBP) and diastolic blood pressure (DBP) derivation, respiratory rate (RR) effects, and errors at BP extremes. Approach: We analysed 205,850 ten-second paired PPG-arterial BP segments from the Medical Information Mart for Intensive Care (MIMIC) waveform database. A convolutional model with a multiresolution U-Net (MultiResUNet) translated PPG to arterial BP waveforms. Inputs at 125, 62.5, and 31.25 Hz were compared. SBP and DBP were derived by minimum-maximum extraction or by averaging the top 4 systolic peaks and bottom 4 diastolic troughs per segment. Performance was summarized using mean absolute error (MAE). Main Results: Inputs at 62.5 Hz reduced MAE versus 125 Hz (SBP 5.24 vs 6.34 mmHg; DBP 2.87 vs 3.27 mmHg), whereas 31.25 Hz increased error (SBP 16.09 mmHg; DBP 9.10 mmHg). Multi-peak averaging outperformed minimum-maximum extraction. MAE varied across respiratory rates, with generally lower errors observed at moderate respiratory rates, and higher errors observed at lower and higher SBP ranges. Significance: Sampling frequency, SBP/DBP derivation, and RR shape error patterns in PPG-only cuffless BP estimation, with higher errors at low and high SBP. These results inform device design for bedside and ambulatory monitoring. However, further validation using subject-level datasets is required to assess clinical applicability and compliance with established validation standards.