Aims: Photoplethysmography (PPG) is a low-cost, non-invasive technology studied for blood pressure (BP) estimation. However, morphological feature consistency across diverse cohorts remains unclear. This study investigates the stability of correlations between PPG features and and systolic and dias tolic blood pressure (SBP/DBP) across four heterogeneous datasets.
Methods: Data include subsets from four sources: PPG-BP (219 subjects, ambulatory), VitalDB (285, intraoperative), MIMIC-III (225, ICU), and MIMIC-IV (63, ICU). Extracted segments (30s VitalDB/MIMIC; 2.1s PPG-BP) were paired with synchronized BP targets. PPG signals were bandpass-filtered (0.5–12 Hz) and screened for physiological plausibility. We extracted 41 pulse features, including timing, amplitude, second derivative ratios (SDPTG, a–e), and Pulse Decomposition Analysis (PDA). Segments were discarded if the median per-feature variability ratio (intra–segment to dataset-wide standard deviation) exceeded 0.3. Spearman correlation (ρ) between features and SBP/DBP was computed and ranked by mean |ρ| across datasets and min-max range.
Results: After applying the variability criterion, the algorithm retained 89.3% of segments for PPG-BP, 55.4% for VitalDB, 62.5% for MIMIC-III, and 70.9% for MIMIC-IV. Twenty-two features showed mean |ρ| > 0.3 with SBP across all datasets, and 15 with DBP. Stable SBP predictors include SDPTG ratios d/a (mean |ρ|=0.71, [0.58–0.81]), b/a (mean |ρ|=0.59, [0.53–0.65]), and (b−c−d−e)/a (mean |ρ|=0.47, [0.38–0.54]), alongside wave reflection index AIgh (mean |ρ|=0.55, [0.42–0.68]) and amplitude ratio Ygh (mean |ρ|=0.53, [0.42–0.61]). For DBP, temporal distances from the dicrotic notch (∆npf, ∆ngh; mean |ρ|:0.46–0.49) exhibited comparable stability. Systolic upstroke time (∆nup) correlated strongly in PPG-BP and VitalDB (|ρ|=0.63) but degraded on MIMIC-III (|ρ|=0.18). DBP predictor pulse slope (Spf) declined from |ρ|=0.53 (VitalDB) to 0.12 (MIMIC-III).
Conclusions: Stable and highly correlated features provide a solid foundation for generalizable BP estimation; however, cohort-specific variability continues to challenge the universal applicability of these predictive models.