A novel hybrid differential evolution strategy applied to classifier design for mortality prediction in adult critical care admissions

Abstract

The optimisation of classifier performance in pattern recognition and medical prognosis tasks is a complex and poorly miderstood problem. Classifier performance is greatly affected by the choice of artificial neural network architecture and starting weights and biases — yet there exists very little guidance in the literature as to how to choose these parameters. Recently evolutionary artificial neural networks have been proposed to mitigate some of these problems; however, whilst evolutionary methods are extremely effective in finding global optima, they are notoriously computationally expensive (often requiring tens of thousands of function evaluations to arrive at a solution). This paper proposes a novel hybrid adaptive approach to the optimisation of artificial neural network parameters where the global search capabilities of differential evolution and the efficiency of local search heuristics (such as resilient back-propagation for artificial neural network training) are combined. A state-of-the-art adaptive differential evolution algorithm, JADE, has been chosen as the basis for this hybrid algorithm due to its proven effectiveness in optimising high dimensional problems. The performance of this hybrid adaptive differential evolution algorithm is then demonstrated in the design of a classifier for mortaUty risk prediction in a critical care environment, where the optimised classifier is shown to outperform the current state-of-the-art in risk prediction.