Blood Pressure Estimation from Photoplethysmogram Using a Spectro-Temporal Deep Neural Network

Area of Science:

• Biomedical Engineering
• Cardiovascular Physiology
• Artificial Intelligence in Medicine

Background:

• Hypertension detection relies on accurate blood pressure (BP) monitoring.
• Cuff-based BP devices have limitations, including user aversion and inability to measure during activity.
• Photoplethysmogram (PPG) signals offer a potential non-invasive, cuffless method for BP estimation.

Purpose of the Study:

• To investigate the feasibility of estimating blood pressure (BP) solely from photoplethysmogram (PPG) signals.
• To develop and evaluate a novel deep neural network for cuffless BP monitoring.
• To assess the impact of model personalization on BP prediction accuracy.

Main Methods:

• Analysis of over 700 hours of high-quality PPG and arterial BP waveforms from the MIMIC III database (510 subjects).
• Utilized a spectro-temporal deep neural network with residual connections, incorporating PPG signals and their derivatives as input.
• Employed a leave-one-subject-out cross-validation strategy to evaluate model performance.

Main Results:

• The developed neural network accurately modeled the relationship between PPG and BP.
• Achieved mean absolute errors of 9.43 mmHg for systolic BP and 6.88 mmHg for diastolic BP.
• Demonstrated that personalized models significantly improve BP prediction accuracy compared to general models.

Conclusions:

• Cuffless BP estimation using PPG and deep learning is a promising approach.
• Model personalization is crucial for achieving high accuracy in PPG-based BP monitoring.
• Publicly releasing code and subject lists aims to establish a baseline for future research in this area.