I-PNN: An Improved Probabilistic Neural Network for Binary Classification of Imbalanced Medical Data

The modern development of Medicine relies heavily on effective data mining However, many practical diagnostic tasks operate on small data samples with an asymmetric number of instances of different classes. This paper considers the binary classification task in the case of a short unbalanced set of medical data. The authors improved the implementation of the Probabilistic Neural Network (IPNN). It is based on a new method of forming the outputs of the PNN's summation layer, which, as in the analog, retains the condition of ensuring a complete system of events (formation of a set of probabilities of belonging to each class that in the sum equal 1). However, in contrast to analogs, this method takes into account the uneven representation of all classes in a stated data set, which provides the ability to effectively solve classification tasks in the case of an unbalanced dataset. The authors substantiate the proposed approach and describe all the steps of algorithmic implementation of the proposed I-PNN. Modeling of I-PNN operation using a well-known unbalanced medical dataset was performed. The optimal parameters of I-PNN operation are selected. An experimental increase in the accuracy of the proposed I-PNN (up to 5% based on Fl-score) compared to the existing PNN was found. All these advantages create many prerequisites for the practical use of I-PNN in the case of processing a short set of unbalanced data in various areas of medical diagnostics.