Oscillometric Blood Pressure Estimation via Machine Learning Analysis of Pulse Transit Time

Oscillometry, a common noninvasive technique, traditionally estimates blood pressure (BP) from the pattern of arterial pulse amplitudes at different cuff pressures (CPs). However, accuracy can be limited with irregular pulses, such as in obese, arteriosclerotic, and atrial fibrillation patients. To address this, this article introduces a novel BP estimation approach that focuses on the timing of oscillometric pulses. We use Gaussian curve fitting (GCF) to decompose each pulse waveform into its forward and reflected components. This decomposition allows us to accurately extract pulse transit time (PTT) from the intervals between waveform components. Leveraging machine learning (ML), we model the relationship between systolic and diastolic BPs and PTT patterns relative to CP. We evaluated this method on two datasets: 425 wrist and 150 upper arm recordings. The results show improved accuracy compared to previous work, with the best results achieved using a random forest model, yielding mean absolute errors (MAEs) of 2.60 and 1.59 mmHg for systolic and diastolic BP in the first dataset and 4.97 and 2.80 mmHg in the second dataset. Our approach met the Association for the Advancement of Medical Instrumentation (AAMI) and British Hypertension Society (BHS) standards for both systolic and diastolic BP, offering enhanced accuracy and reliability for oscillometric BP devices and potential applicability in challenging patient populations.