Support vector machines for the prediction of aortic stiffness from the digital volume pulse waveform

Abstract

Cardiovascular disease (CVD) is currently the biggest single cause of mortality in the developed world hence its early detection and prevention is an important health issue. Aortic stiffness as measured by aortic pulse wave velocity (PWV) has been shown to be an independent predictor of CVD, however, the measurement of PWV is time consuming. Recent studies have shown that pulse contour characteristics depend on arterial properties such as arterial stiffness. This paper presents a method for estimating PWV from the digital volume pulse (DVP), a waveform that can be rapidly and simply acquired by measuring transmission of infra-red light through the finger pulp. PWV and DVP were measured on 461 subjects attending a cardiovascular prevention clinic at St Thomas' Hospital, London. The DVP contour was characterised by features such as the stiffness index, the time to the first peak and the time between the first and second peaks. Low stiffness, is defined as PWV<9 ms-1 and high stiffness as PWV>11 ms-1. When using traditional artificial neural network (ANN) approaches, the sensitivity and specificity to classify a specific set of DVP features into high and low PWV was 79% and 74% respectively. Using a non-linear Kernel based support vector machine (SVM) classifier with the same set of features, it is possible to achieve results of up to 87% sensitivity and 76% specificity on unseen data. Further, the use of support vector regression (SVR) techniques leads to a direct real-valued estimate of PWV with a standard deviation of only 2.0 ms-1 (compared with 2.5 ms-1 with ANN and 2.6 ms-1 with multi-linear regression). We therefore conclude that support vector machine based classification and regression techniques improve the prediction of arterial stiffness from the simple measurement of the digital volume pulse.