A Cuffless Blood Pressure Estimation Method Using Dimensionality Increasing and Two-Dimensional Convolution.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Current blood pressure (BP) monitoring methods, such as cuff-based non-invasive and vessel puncture invasive techniques, have limitations in real-time application and practicality.
• Accurate and continuous BP monitoring is crucial for managing hypertension and related cardiovascular diseases.

Purpose of the Study:

• To develop a novel, cuffless method for estimating blood pressure (BP) using 2D convolutional neural networks.
• To enhance the accuracy and efficiency of BP estimation by converting physiological signals into image representations.
• To validate the proposed method's performance and generalization capabilities on a clinical dataset.

Main Methods:

• Electrocardiography (ECG) and photoplethysmography (PPG) signals were transformed into 2D images using dimensionality increasing algorithms like recurrence plots and Gramian angular fields.
• Novel image representation techniques, fused Gramian angular field (FGAF) and combined Gramian angular field (CGAF), were proposed to optimize data and enhance signal relevance.
• 2D convolutional neural network models, including VGGNet, were trained on these images for BP estimation, with subsequent model pruning for size reduction.
• The pruned VGGNet model was fine-tuned and validated on the MIMIC-III dataset, and a prototype monitoring system was developed.

Main Results:

• The 2D convolutional models significantly improved BP estimation accuracy compared to 1D models, with VGGNet using Gramian angular difference field (GADF) showing a 38% accuracy increase.
• The proposed FGAF and CGAF methods reduced input data by 50% while maintaining high estimation accuracy.
• Minimum mean absolute errors for systolic and diastolic BP estimation reached 2.71 and 1.74 mmHg, respectively.
• A pruned VGGNet model maintained performance while reducing channel numbers by 60%, demonstrating generalization ability on the MIMIC-III dataset.

Conclusions:

• The proposed cuffless BP estimation method using 2D convolutional models offers a promising, accurate, and practical alternative to traditional BP monitoring.
• The novel image conversion techniques (FGAF, CGAF) effectively reduce data complexity and improve signal relevance for enhanced BP estimation.
• The developed system demonstrates the feasibility of real-time signal collection and BP estimation, paving the way for improved cardiovascular health management.