Calibration-free blood pressure estimation based on a convolutional neural network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Wearable devices require efficient blood pressure (BP) estimation models with limited computing power and susceptibility to signal noise.
• Traditional BP monitoring methods are often invasive or inconvenient for continuous, real-time tracking.

Purpose of the Study:

• To develop and evaluate a deep learning-based BP estimation model optimized for wearable environments.
• To assess the efficacy of using electrocardiogram (ECG) and photoplethysmogram (PPG) signals for BP estimation.
• To investigate the impact of noise and signal length on model performance.

Main Methods:

• A 3-layer convolutional neural network (CNN) was employed for BP estimation.
• Time-series ECG and PPG signals were preprocessed using differential and thresholding methods to reduce noise.
• Max-pooling techniques were applied to extract features from the input signals.
• The model was trained and validated on 2.4 million data samples from 49 intensive care unit patients (MIMIC database).

Main Results:

• The model achieved an average root mean square error of 3.41, 5.80, and 2.78 mm Hg for pulse pressure, systolic BP (SBP), and diastolic BP (DBP), respectively.
• Cumulative error percentages less than 5 mm Hg were 68% for SBP and 93% for DBP.
• ECG signals demonstrated superior performance in noise reduction compared to PPG signals, with lower mean absolute errors (9.72 mm Hg for SBP, 6.67 mm Hg for DBP).
• Input signal length did not significantly impact CNN performance.

Conclusions:

• Deep learning models, particularly CNNs, are effective for wearable BP estimation.
• ECG signals are more suitable than PPG signals for BP estimation in noisy wearable environments.
• Short sampling frames without calibration can be utilized for effective BP estimation.