Employing machine learning techniques for non-invasive blood pressure classification using photoplethysmography...

Area of Science:

• Biomedical Engineering
• Data Science
• Cardiovascular Health

Background:

• Effective blood pressure (BP) regulation is crucial for health, but traditional measurement methods are often invasive or cumbersome.
• Continuous, non-invasive BP monitoring is needed to address challenges posed by traditional methods and manage cardiovascular health risks.
• Photoplethysmography (PPG) signals offer a potential avenue for non-invasive BP assessment.

Purpose of the Study:

• To enhance non-invasive blood pressure (BP) monitoring using photoplethysmography (PPG) signals and machine learning (ML).
• To develop and validate an accurate ML model for classifying BP into four distinct categories.
• To improve computational efficiency and model robustness through feature selection and data balancing techniques.

Main Methods:

• Analyzed PPG data from a diverse cohort (ages 21-86) with varying health statuses.
• Applied four feature selection strategies and five ML classification models with k-fold cross-validation.
• Implemented data balancing techniques to address dataset imbalance and improve model performance.

Main Results:

• An ensemble-based extra trees classifier (ETC) with SelectFromModel achieved 91.13% accuracy initially.
• After data balancing, test accuracy improved to 93.25%, with F1 scores exceeding 90% for most BP categories.
• The study demonstrated high efficacy in classifying BP categories non-invasively.

Conclusions:

• The developed non-invasive BP monitoring system using PPG and ML shows significant potential for clinical application.
• This method offers a more convenient and continuous alternative to traditional cuff-based BP measurements.
• Further research can explore broader applications and integration into wearable health devices.