A framework for application of neuro-case-rule base hybridization in medical diagnosis

Abstract

Every approach to handling automation has its unique limitations. In the symbolic (rule base) approach, the brittleness of rules leads to the ineffectiveness of handling noisy data, but it derives its strengths in heuristic search. In the same vein, a case base reasoning paradigm is bedeviled with retrieval and adaptation problems. Neural Networks (NN) methodology suffers from intolerance of incremental insertion of new knowledge and limited explanation capability, but triumphs over other methods when it comes to adaptation using its generalization characteristics. Based on all these, a tight coupling of case base, rule base and neural networks methodologies is proposed for medical diagnosis. The case base provides the ‘desired’ outputs, which constitute an input to the neural networks. The results obtained from the trained neural networks assisted in formulating diagnostic rules, which form the rule base. Through the rule base, an inference engine that represents the hybrid is built. Data collected from three hospitals in Nigeria on hepatitis patients were used to test the functionality of the proposed system. The results obtained from the hybrid were compared with that obtained from the Multilayer Peceptron Neural Networks (MLPNN) training using NeuroSolutions 5.0 and found to covary strongly. The hybrid exhibits an explanation characteristic, a feature not found in neural networks.