Blood Pressure Estimation Based on PPG and ECG Signals Using Knowledge Distillation.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Cardiovascular Monitoring

Background:

• Cuff-based and invasive blood pressure (BP) monitoring present significant limitations.
• Easy-access bio-signals offer a potential solution for non-invasive BP estimation.
• Photoplethysmographic (PPG) and electrocardiogram (ECG) signals are readily available bio-signals.

Purpose of the Study:

• To develop and validate a deep learning model for estimating systolic and diastolic blood pressure.
• To leverage knowledge distillation for training an efficient and accurate BP estimation model.
• To assess the model's performance against established medical standards.

Main Methods:

• A multistage deep learning architecture incorporating convolutional, bidirectional recurrent, and attention layers was designed.
• Knowledge distillation was employed, training a smaller student model using insights from a larger teacher model.
• The model was trained and validated on 1205 subjects from the MIMIC III database.

Main Results:

• The model achieved Grade A performance for both systolic blood pressure (SBP) and diastolic blood pressure (DBP) estimation, meeting AAMI standards.
• Post-knowledge distillation (KD), the model demonstrated a mean absolute error of 2.94 ± 5.61 mmHg for SBP and 2.02 ± 3.60 mmHg for DBP.
• The KD training method significantly reduced model parameters while enhancing predictive accuracy.

Conclusions:

• Knowledge distillation is a beneficial training strategy for blood pressure regression models.
• The proposed deep learning model offers a viable, non-invasive approach for blood pressure monitoring.
• The study highlights the potential of PPG and ECG signals for accurate BP estimation.