Pulse Wave Analysis for BP Estimation
Pulse wave analysis (PWA) estimates blood pressure from single-site PPG waveform morphology features — systolic upstroke slope, pulse width, augmentation index, dicrotic notch timing, and second-derivative ratios — that encode arterial impedance and blood pressure information without requiring multi-site transit time measurement.
PWA extracts 15–30 morphological features from each PPG beat that correlate with arterial blood pressure. Key features include: systolic upstroke time (inversely correlated with systolic BP), pulse width at half-maximum (related to pulse pressure), diastolic decay time constant (related to arterial compliance), and the b/a ratio from the second derivative PPG (correlated with arterial stiffness). These features capture the hemodynamic imprint of blood pressure on the peripheral pulse waveform.
Machine learning models — gradient boosted trees (XGBoost), random forests, and deep neural networks — map these features to systolic and diastolic BP. Population-level models trained on >10,000 subjects achieve 8–12 mmHg MAE for systolic BP without individual calibration. Personalized models with single-point calibration reduce error to 5–8 mmHg. The CareAI study (2023) demonstrated a CNN processing raw PPG waveforms achieving 6.2 mmHg systolic MAE on an independent test set of 500 patients, approaching ISO 81060-2 criteria.
The fundamental limitation of single-site PWA is that peripheral PPG morphology depends on many factors beyond blood pressure: arterial stiffness, heart rate, stroke volume, peripheral resistance, and measurement site. Disentangling BP-specific features from confounders requires either large diverse training data or physical model constraints that encode cardiovascular physiology into the feature extraction pipeline.
Frequently Asked Questions
Can PWA work without any calibration?
Calibration-free PWA models achieve 8–12 mmHg MAE, below clinical standards but useful for relative trending. Models trained on very large, diverse populations show promise for calibration-free estimation within 5–7 mmHg for normotensive ranges.
Which PPG features are most predictive of blood pressure?
Systolic upstroke time, pulse area ratio (systolic/total), second-derivative b/a ratio, augmentation index proxy, and pulse width at 25% amplitude are consistently the top features in importance analyses across multiple studies.
How does PPG waveform quality affect BP estimation?
Signal quality directly impacts BP estimation accuracy. SQI < 0.8 increases BP estimation error by 30–60%. Quality-gated estimation that only reports BP from high-quality beats substantially improves clinical utility.