A Novel Clustering-Based Algorithm for Continuous and Non-invasive Cuff-Less Blood Pressure Estimation

TL;DR

This paper introduces a clustering-based machine learning approach to improve the accuracy of cuffless, continuous blood pressure estimation from physiological signals, achieving significantly better results than previous models.

Contribution

The paper presents a novel clustering step combined with regression models to enhance blood pressure estimation accuracy, addressing limitations of prior methods.

Findings

MAE of 2.56 for SBP and 2.23 for DBP with 5 clusters and GBR

Significant accuracy improvement over non-clustered models

Effective use of the MIMICII dataset for validation

Abstract

Extensive research has been performed on continuous, non-invasive, cuffless blood pressure (BP) measurement using artificial intelligence algorithms. This approach involves extracting certain features from physiological signals like ECG, PPG, ICG, BCG, etc. as independent variables and extracting features from Arterial Blood Pressure (ABP) signals as dependent variables, and then using machine learning algorithms to develop a blood pressure estimation model based on these data. The greatest challenge of this field is the insufficient accuracy of estimation models. This paper proposes a novel blood pressure estimation method with a clustering step for accuracy improvement. The proposed method involves extracting Pulse Transit Time (PTT), PPG Intensity Ratio (PIR), and Heart Rate (HR) features from Electrocardiogram (ECG) and Photoplethysmogram (PPG) signals as the inputs of clustering and regression, extracting Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP) features from ABP signals as dependent variables, and finally developing regression models by applying Gradient Boosting Regression (GBR), Random Forest Regression (RFR), and Multilayer Perceptron Regression (MLP) on each cluster. The method was implemented using the MIMICII dataset with the silhouette criterion used to determine the optimal number of clusters. The results showed that because of the inconsistency, high dispersion, and multi-trend behavior of the extracted features vectors, the accuracy can be significantly improved by running a clustering algorithm and then developing a regression model on each cluster, and finally weighted averaging of the results based on the error of each cluster. When implemented with 5 clusters and GBR, this approach yielded an MAE of 2.56 for SBP estimates and 2.23 for DBP estimates, which were significantly better than the best results without clustering (DBP: 6.27, SBP: 6.36).