A new methodology for medical differential diagnosis purposes, based on the use of fuzzy clustered hybrid networks, is detailed in this paper. The study is focused on discriminating among the diseases of a certain class that includes normal hearts and seven types of cardiac abnormalities. Basically, the novelty of the proposed methodology derives from using a variant of fuzzy c-means clustered probabilistic neural networks designed to extract relevant features from data, and a multilayered feed-forward network to perform the classification task on the basis of the resulting feature values.

Following an introduction containing a series of comments about similar approaches published in the literature, the next two sections supply a brief medical description of each cardiac abnormality taken into account, and some details concerning the preprocessing steps performed on the electrocardiogram (ECG) signals. The next several sections present the feature extraction method based on auto-regressive modeling, the fuzzy clustering technique for feature reduction purposes, and a description of the neural architecture used in the classification step.

The performance of the learned classifier was analyzed in terms of classification accuracy, sensitivity, and specificity, experimentally evaluated for all classes. The comparative analysis against several similar approaches demonstrates that the probabilistic neural network with clustered features outperforms the other methods from the point of view of both speed and accuracy of classification. Moreover, the combined use of probabilistic neural networks and multilayer feed-forward networks shows outstanding accuracy in discriminating among the diseases considered in the study.