Peripheral pulse multi-Gaussian decomposition using a modified artificial bee colony algorithm

Abstract

Time-domain analysis is one of widely used methods to acquire physiological information from peripheral pulse waves. It is crucial to precisely extract features from pulse waves so that classifications can be performed to distinguish differences with high accuracy. Curve fitting has been proved to be more effective than points-based detection for modeling pulse signal trends, and the Gaussian function is a popular model because its bell shape is like an individual pulse. The pulse decomposition problem is exactly an optimization problem, which can be solved by swarm algorithms. In this study, we use Gaussian-mixture functions to fit practical pulse waves by an artificial bee colony (ABC) algorithm. A modified ABC (MABC) algorithm is compared with classical ABC for performance evaluation. MABC is adapted from the classical ABC by modifications in searching strategies of employed bees, onlooker bees, and scout bees. It is concluded that MABC converges quickly under limited iteration times and is effective up to a dimension of 30 for a single complete cycle. We collected pulse waves from a group of 46 healthy volunteers and a group of 69 type-2 diabetic candidates in a clinic. A support vector machine (SVM) with radial basis functions (RBFs) was used for classification after pulse decomposition; the training and test data were obtained randomly at a 7:3 ratio. The largest accuracy averaged 89.1% by considering combinations of the first, third, and fourth center location values, and the first amplitude value of Gaussian functions.