Using direct explanations to validate a multi-layer perceptron network that classifies low back pain patients

Abstract

Using a new method designed by the first author, this paper shows how direct explanations in the form of a ranked data relationship can be provided to explain the classification of an input case by a standard multilayer perceptron (MLP) network. It is also shown how knowledge in the form of an induced rule can be discovered from the data relationship for each training case. The method is demonstrated for example training cases from a real-world MLP that classifies low back pain patients into three diagnostic classes. It is shown how the validation of the explanations for all training cases provides a way of validating the low back pain MLP network. In validating the network, a number of test cases apparently mis-classified by the MLP were found to have been correctly classified by the network and incorrectly classified by the clinicians.