Fuzzy preprocessing and artificial neural network classification for the diagnostic interpretation of the resting ECG

An hybrid system for the interpretation of the 12 leads resting ECG has been designed and tested. A preprocessor based on fuzzy set theory has been embedded in an artificial neural network classifier. A well documented ECG database, produced at the University of Leuven and used for evaluating other ECG diagnostic programs has been chosen for training and testing the hybrid system. It contains 3266 12 lead ECG records clinically tested. A layer of Radial Basis Functions used as fuzzy activation functions, are embedded in a neural network architecture for fuzzy preprocessing every input pattern. Several networks have been implemented and tested by varying learning rules, system architecture, and the various parameters of the hybrid system. A pruning technique is also applied to reduce the system size. Several classification strategies have been generated by varying the roles of the network output units, and have been evaluated as well. The results show that this combination of fuzzy and neural techniques is effective even for small structures. The best system reached the 69% of average sensitivity and 94% of average specificity on the test set. Thus it shows very promising performances, since they are comparable with those of traditional elaborated systems. This hybrid architecture has a further interesting feature to be deeply investigated since in principle it allows the interpretation/explanation of the results obtained from a neural network.