The Use of a Knowledge Discovery Method for the Development of a Multi-layer Perceptron Network that Classifies Low Back Pain Patients

摘要

Using a new method published by the first author this paper discovers the ranked class profiles of key inputs used by a multi-layer perceptron (MLP) network that classifies low back pain patients into three diagnostic classes. It is shown how the validation of the class profiles leads to the discovery of 4 mis-diagnosed training cases and 2 further cases which were not relevant. By interpreting the test cases mis-classified by the MLP and comparing them with the validated class profiles a number of test cases were also found to have been mis-diagnosed by the clinicians. It is shown how the class profiles were used to develop a more optimal network with approximately half the number of inputs and only a marginally reduced test performance.