Swarm ANN/SVR-Based Modeling Method for Warfarin Dose Prediction in Chinese

Abstract

In order to improve the predictive accuracy on warfarin dose, a modeling method using artificial neural network (ANN) and support vector regression (SVR) together with particle swarm optimization (PSO) is developed, which is denoted as PSO-(ANN/SVR). The procedure of PSO-(ANN/SVR) runs a population of ANNs and SVR to develop diverse candidate models, and a PSO is employed as a “shell” to optimize a group of ANNs and SVR by following the current optimum particles(i.e. the best ANN or SVR) in search space. We collected a dataset of 100 Chinese patients provided by The First affiliated Hospital of Soochow University, and divide it into training, validation and test set. (ANN/SVR) models are built on the training and validation set, and finally tested on test set. In the experiment, PSO-(ANN/SVR) is compared with five modeling methods in terms of mean squared error (MSE) and squared correlation (R ²). The experimental results show that the models developed by PSO-(ANN/SVR) present the best MSE and R ² in these cases. PSO-(ANN/SVR) achieved 11.9–48.9% reduction on MSE over the other methods with different variables in test set. It is noted that PSO-(ANN/SVR) models had a small decrease in R ² from training set to test set, and obtained a large R ² on both training and test set. This illustrates that the models of PSO-(ANN/SVR) have a good generalization.