Automated Hypertension Detection Using ConvMixer and Spectrogram Techniques with Ballistocardiograph Signals

Area of Science:

• Biomedical Engineering
• Cardiovascular Health
• Artificial Intelligence in Medicine

Background:

• High blood pressure (HBP), or hypertension (HPT), is a critical health issue often termed the "silent killer" due to its severe, widespread damage and high mortality rate.
• Early and accurate diagnosis of HPT is essential for effective management and prevention of complications.
• Ballistocardiography (BCG) signals offer a non-invasive method for physiological monitoring.

Purpose of the Study:

• To develop and evaluate an automated system for diagnosing high blood pressure (HPT) using ballistocardiography (BCG) signals.
• To investigate the efficacy of deep learning models, specifically ConvMixer, for classifying hypertension from BCG-derived time-frequency representations.
• To compare the performance and efficiency of the ConvMixer architecture against established models like ResNet18 and ResNet50.

Main Methods:

• BCG signals were processed and transformed into the time-frequency domain using the spectrogram method, with optimized parameters for window type, length, overlap, and FFT size.
• Spectrogram images were then classified using the ConvMixer deep learning architecture, a model inspired by vision transformers and MLP-mixers.
• The ConvMixer model's diagnostic performance was benchmarked against ResNet18 and ResNet50 architectures.

Main Results:

• The ConvMixer architecture demonstrated high accuracy in diagnosing HPT, achieving 97.69%.
• Comparative analysis showed high accuracy for ResNet18 (98.14%) and ResNet50 (98.79%).
• The ConvMixer model exhibited significantly shorter processing times compared to ResNet18 and ResNet50, indicating computational efficiency.

Conclusions:

• The proposed automated diagnosis system using BCG signals and the ConvMixer architecture is effective for detecting high blood pressure.
• ConvMixer offers a promising, computationally efficient alternative for HPT diagnosis, balancing high accuracy with reduced operational time.
• This approach highlights the potential of leveraging advanced deep learning techniques on physiological signals for rapid and non-invasive cardiovascular health assessment.