Applied machine learning for blood pressure estimation using a small, real-world electrocardiogram and...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Estimating non-occlusive blood pressure using machine learning on electrocardiography (ECG) and photoplethysmography (PPG) signals is an active research area.
• Real-world ambulatory ECG and PPG signals suffer from motion and noise artifacts, degrading the performance of established machine learning models trained on clean datasets.

Purpose of the Study:

• To improve the robustness of machine learning models for blood pressure estimation from noisy, real-world ECG and PPG signals.
• To evaluate the performance of four machine learning methods (Random Forest, Support Vector Regression, Adaboost, Artificial Neural Networks) on a small, self-sampled dataset.

Main Methods:

• Applied four machine learning regression techniques to a dataset of 54 subjects.
• Utilized pulse arrival time, morphological and frequency PPG parameters, and heart rate variability as features.
• Evaluated model performance using root mean square error (RMSE) and mean absolute error (MAE).

Main Results:

• Support Vector Regression (SVR) demonstrated the best performance in estimating blood pressure from noisy data.
• SVR achieved a Mean Absolute Error of 6.97 mmHg, meeting the British Hypertension Society's Level C criteria.
• The study confirmed the feasibility of using ambulatory ECG-PPG signals from mobile devices for blood pressure estimation.

Conclusions:

• Machine learning models, particularly Support Vector Regression, can effectively estimate blood pressure even with noisy ambulatory ECG and PPG signals.
• The findings suggest that mobile, discrete devices can be utilized for robust, non-invasive blood pressure monitoring.
• This approach holds promise for improving remote and continuous blood pressure tracking.