A PSO-based rule extractor for medical diagnosis

One of the major bottlenecks in applying conventional neural networks to the medical field is that it is very difficult to interpret, in a physically meaningful way, because the learned knowledge is numerically encoded in the trained synaptic weights. In one of our previous works, we proposed a class of Hyper-Rectangular Composite Neural Networks (HRCNNs) of which synaptic weights can be interpreted as a set of crisp If-Then rules; however, a trained HRCNN may result in some ineffective If-Then rules which can only justify very few positive examples (i.e., poor generalization). This motivated us to propose a PSO-based Fuzzy Hyper-Rectangular Composite Neural Network (PFHRCNN) which applies particle swarm optimization (PSO) to trim the rules generated by a trained HRCNN while the recognition performance will not be degraded or even be improved. The performance of the proposed PFHRCNN is demonstrated on three benchmark medical databases including liver disorders data set, the breast cancer data set and the Parkinson's disease data set.