Blood pressure estimation and classification using a reference signal-less photoplethysmography signal: a deep...

Area of Science:

• Biomedical Engineering
• Cardiovascular Health Monitoring
• Artificial Intelligence in Healthcare

Background:

• Hemodynamic parameters like blood pressure (BP) are crucial for cardiovascular system function assessment.
• Continuous BP monitoring aids in early detection of cardiovascular diseases (CVDs), but motion artifacts often impede accuracy.
• Existing methods may require multiple sensors (e.g., ECG and PPG), limiting practical application.

Purpose of the Study:

• To develop an optimized deep learning model for simultaneous estimation of systolic blood pressure (SBP) and diastolic blood pressure (DBP) from a single photoplethysmography (PPG) signal.
• To classify BP stages (normotension, prehypertension, hypertension) using the same network, robust to motion artifacts.
• To enable accurate, non-invasive BP monitoring suitable for wearable devices.

Main Methods:

• Utilized a Convolutional Neural Network (CNN) for automatic feature extraction from PPG signals.
• Employed the superlet transform to convert 1-D PPG signals into 2-D time-frequency (TF) spectrograms, separating true signals from motion artifacts.
• Trained and validated the model on the MIMIC-III dataset.

Main Results:

• Achieved Mean Absolute Errors (MAE) of 2.71 mmHg for SBP and 2.42 mmHg for DBP.
• Demonstrated high classification accuracy: 96.79% for SBP and 98.94% for DBP stages.
• The superlet transform effectively mitigated motion artifact interference.

Conclusions:

• The proposed optimized deep learning framework accurately estimates and classifies BP from single-channel PPG signals, even in the presence of motion artifacts.
• The method is less complex and suitable for deployment on resource-constrained wearable devices.
• This approach offers a promising, non-invasive alternative for continuous cardiovascular health monitoring.