Evolving fuzzy medical diagnosis of Pima Indians diabetes and of dermatological diseases.

Objective: This paper reviews a methodology for evolving fuzzy classification which allows data to be processed in online mode by recursively modifying a fuzzy rule base on a per-sample basis from data streams. In addition, it shows how this methodology can be improved and applied to the field of diagnostics, for two popular medical problems.

Method: The vast majority of existing methodologies for fuzzy medical diagnostics require the data records to be processed in offline mode, as a batch. Unfortunately this allows only a snapshot of the actual domain to be analysed. Should new data records become available they require cost sensitive calculations due to the fact that re-learning is an iterative procedure. eClass is a relatively new architecture for evolving fuzzy rule-based systems, which overcomes these problems. However, it is data order dependent as different orders of the data result into different rule bases. Nonetheless, it is shown that models of eClass can be improved by arranging the order of the incoming data using a simple optimization strategy.

Results: In regards to the Pima Indians diabetes dataset, an accuracy of 79.37% was obtained, which is 0.84% lower than the highest in the literature. The proposed optimization strategy increased the accuracy and specificity of the model by 4.05% and 7.63% respectively. For the dermatology dataset, an accuracy of 97.55% was obtained, which is 1.65% lower than the highest in the literature. In this case, the proposed optimization strategy improved the accuracy of the model by 4.82%. The improved algorithm has been compared to other existing algorithms and seems to outperform the majority.

Conclusions: This paper has shown that eClass can effectively be applied to the classification of diabetes and dermatological diseases from discrete numerical samples. The results of using a novel optimization strategy indicate that the accuracy of eClass models can be further improved. Finally, the system can mine human readable rules which could enable medical experts to gain better understanding of a sample under analysis throughout the traditional diagnostic process.