Multicollinearity Analysis for Cuffless Blood Pressure Estimation Regression Algorithms

Abstract

The estimation of blood pressure in a cuffless manner has been a major study of interest for past decade. Blood Pressure has shown significant correlation with features derived from PPG and ECG signals; this has led to development of multiple regression algorithm for estimating blood pressure using the derived features. The aim of our study is to analyze the existence of multicollinearity in multivariable regression algorithms proposed by earlier studies. Continuous bio-signals constituting Arterial Blood Pressure (ABP), Plethysmography (PPG) and Electrocardiogram (ECG) of 240-minute length for 250 subjects was collected from MIMIC-III database. Six different features including pulse transit time and heart rate was extracted from the data and subjected to multicollinearity analysis using variation inflation factor and correlation coefficient. High multicollinearity was observed for algorithms with more than 2 independent variables indicating that such algorithms suffer from inherent problem and could lead to unstable regression coefficients and algorithm. Principal Component Analysis is proposed in our study to tackle the problem of multicollinearity in blood pressure estimation algorithm. This method would remove the dangers of multicollinearity without any loss of information from features. Further analysis showed that top half of the principal components explained more than 90.0% of the feature variance. Thereby it was concluded that multivariable regression algorithms for blood pressure estimation suffer from multicollinearity and this needs to be addressed before developing an algorithm.