Cuffless blood pressure estimation methods: physiological model parameters versus machine-learned features.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Cardiovascular Monitoring

Background:

• Blood pressure (BP) monitoring traditionally relies on invasive methods.
• Non-invasive BP monitoring using photoplethysmogram (PPG) and electrocardiogram (ECG) signals is an active research area.
• Existing methods often use either physiological parameters or machine-learned features from these signals.

Purpose of the Study:

• To investigate crucial aspects of BP monitoring using PPG and ECG signals.
• To compare the performance of physiological parameters versus machine-learned features.
• To evaluate the impact of the ECG waveform, beyond its role as a time reference, on BP monitoring accuracy.

Main Methods:

• Extracted 27 physiological parameters from PPG and ECG signals.
• Utilized convolutional neural networks (CNNs) to extract deep-learned features from PPG alone and combined PPG-ECG segments.
• Evaluated feature sets and their combinations using various regression models on data from 200 subjects.

Main Results:

• The ECG waveform contains significant information that enhances BP monitoring accuracy.
• Machine-learned features extracted by CNNs demonstrated superior performance compared to traditional physiological parameters.
• Combining physiological parameters and machine-learned features did not yield additional improvements in accuracy.

Conclusions:

• CNN-based feature extraction provides concise and precise representations from ECG and PPG signals for effective BP monitoring.
• Machine-learned features are more effective than physiological parameters for non-invasive BP estimation.
• The ECG waveform's intrinsic characteristics are vital for improving BP monitoring accuracy.