A signature-based liver cancer predictive system

Abstract

The predictive system presented in this paper employs both SOM and Hopfield nets to determine whether a given chemical agent causes cancer in the liver. The SOM net performs the clustering of the training set and delivers a signature for each cluster. Hopfield net treats each signature as an exemplar and learns the exemplars. Each record of the test set is considered a corrupted signature. The Hopfield net tries to un-corrupt the test record using learned exemplars and map it to one of the signatures and consequently to the prediction value associated with the signature. Four pairs of training and test sets are used to test the system. To establish the validity of the new predictive system, its performance is compared with the performance of the discriminant analysis and the rough sets methodology applied on the same datasets.