Optimized deep neural network models for blood pressure classification using Fourier analysis-based time-frequency...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Cardiovascular diseases (CVDs) pose a significant health risk, with hypertension being a primary cause.
• Effective blood pressure (BP) management through continuous monitoring and rapid diagnosis is crucial for CVD prevention.
• Photoplethysmography (PPG) signals offer a non-invasive method for physiological monitoring, but are susceptible to motion artifacts.

Purpose of the Study:

• To develop a real-time method for estimating BP from motion artifact-affected PPG signals.
• To propose a deep neural network (DNN) methodology for accurate PPG signal classification.
• To classify PPG signals into normotension, pre-hypertension, and hypertension categories.

Main Methods:

• Utilizing the Fourier Decomposition Method (FDM) to transform PPG signals into time-frequency (TF) spectrograms.
• Modifying the last three layers of pre-trained DNNs (GoogleNet, DenseNet, AlexNet) for PPG signal classification.
• Training and testing the framework on MIMIC-III and PPG-BP databases using a five-fold cross-validation approach.

Main Results:

• The proposed framework with the DenseNet-201 network achieved a test accuracy of 96.5%.
• FDM effectively computed TF spectrograms, separating motion artifacts and noise from PPG signals.
• DNNs trained with clean PPG signal features demonstrated improved generalized capability for real-time BP classification.

Conclusions:

• The DNN-based approach using FDM-generated TF spectrograms accurately estimates BP from noisy PPG signals.
• DenseNet-201 demonstrated superior performance in classifying BP states from PPG data.
• This method enhances the potential for real-time, non-invasive BP monitoring and hypertension screening.