To date, continuous cuff-based non-invasive blood pressure (NIBP) monitoring remains limited. NIBP provides intermittent readings for systolic blood pressure (SBP) but misses short-term SBP fluctuations. Newly developed approaches that use the photoplethysmogram (PPG) for estimating SBP generalize poorly without calibration due to subject-specific vascular properties. To overcome these limitations, we reformulate PPG-based SBP estimation as estimating the change in SBP (ΔSBP) between two time points.
Methods. We implemented a learning-based approach that compares pairs of 20-second PPG segments to estimate ΔSBP. Our approach was developed and evaluated on 12,279 patients, with a held-out test set comprising 2,457 patients, collected from the VitalDB and MIMIC-III datasets. Root mean squared error (RMSE), Pearson correlation coefficient (r), and Cohen's kappa (κ) were used for performance quantification.
Results. On the held-out test sets, our approach achieved RMSE of 14.53/11.37 mmHg, r of 0.79/0.65, and κ of 0.57/0.44 for PPG-based ΔSBP (VitalDB/MIMIC-III), across segment pairs separated by 1 min to multiple hours. Errors occurred primarily in cases where morphological analyses identified only minor changes in the waveform despite large ΔSBP values, especially in MIMIC-III. Model predictions correlated more strongly with SBP than with heart rate, indicating reliance on pulse wave morphology. Long-term analysis demonstrated stable ΔSBP estimation over 28 hours.
Conclusion. Our results support the potential of morphology-based relative blood pressure estimation using PPG. The introduced approach could complement cuff-based monitoring by filling gaps between intermittent measurements, particularly during sleep, where cuff inflations are disruptive.