Robust Feature Selection for BP Estimation in Multiple Populations: Towards Cuffless Ambulatory BP Monitoring | JoVE Visualize

Area of Science:

Biomedical Engineering
Cardiovascular Health
Machine Learning in Healthcare

Background:

Current blood pressure (BP) estimation methods lack accuracy and adaptability for ambulatory diagnosis, especially in at-risk populations, often due to limited sample sizes.
Photoplethysmography (PPG) signals offer a non-invasive approach, but robust feature extraction for BP estimation remains a challenge.

Purpose of the Study:

To develop and validate a novel algorithm for cuffless blood pressure estimation using PPG signals and demographic features.
To enhance the adaptability of BP estimation models to population shifts using advanced feature selection techniques.
To compare the performance of in-clinic versus ambulatory BP monitoring models.

Main Methods:

An algorithm was developed for BP estimation using PPG signals and demographic data.
Markov Blanket (MB) feature selection was employed to identify informative and transmissible features.
The algorithm was validated on the Aurora-BP database, comprising ambulatory, wearable, cuffless BP measurements from over 500 individuals.
Gradient Boosting was selected as the optimal machine learning regression method.

Main Results:

MB feature selection identified temporal, frequency, and demographic features as most important for BP estimation, while statistical features were deemed non-significant.
The proposed MB feature space demonstrated superiority over generic models, achieving a mean absolute error of [Formula: see text] for systolic BP and [Formula: see text] for diastolic BP.
Ambulatory models showed significantly lower performance compared to in-clinic models, with estimation errors dropping by [Formula: see text] for systolic and [Formula: see text] for diastolic BP.

Conclusions:

The study provides a robust algorithm for BP estimation from PPG signals, adaptable to population shifts.
The findings highlight the importance of causal features identified through MB selection for resilient BP estimation.
Significant performance disparities between in-clinic and ambulatory BP monitoring models were quantified, emphasizing the need for specialized ambulatory models.