A Signal Processing Framework for Cuffless Blood Pressure Estimation Using Convolutional Blind Source Separation on Multi-Wavelength Photoplethysmography Signals

Traditional pulse transit time (PTT)-based blood pressure (BP) estimation methods requiring multi-point sensors face significant limitations in accessibility and practicality. This study introduces a improved signal processing framework using wrist-based multi-wavelength photoplethysmography (PPG) to calculate arteriolar PTT for non-invasive BP estimation. By leveraging the Modified Beer-Lambert Law, we model the relationship between observed PPG signals and underlying physiological sources (including arterial blood volume changes) as a convolutional blind source separation (CBSS) problem. A time-frequency domain non-negative matrix factorization (NMF) model, optimized via iterative cost function minimization, enables separation of arterial blood volume change signals. The time delay between extracted arterial and superficial capillary signals defines arteriolar PTT, enable beat-to-beat BP estimation. Proof-of-concept validation in an 8 -subject intensive care unit (ICU) cohort yielded mean errors of 0.087 mmHg for systolic BP (SBP) and -0.737 mmHg for diastolic BP (DBP). Comparative analysis showed CBSS-derived arteriolar PTT exhibited better BP correlation than direct wavelength-based time-delay methods, validating CBSS for arterial component extraction and arteriolar PTT-based BP estimation.