A non-linear Kernel feature subset selection based semi-supervised framework for medical disease prediction

Abstract

 As the size of the medical datasets increases, the prediction of disease patterns also increases in biomedical applications in the high dimensional features. Micro-array data play a significant role in the pathogenesis of the various diseases. Many existing methods are not applicable to the prediction of disease without identified genes and semi-supervised learning. Meanwhile, several other approaches have struggled to prioritize associations for all diseases at the same time. Therefore, the development of an algorithm that can identify reliable candidates for disease using existing gene-disease associations verified by the biological experiment is essential to solve these problems effectively. To overcome these problems, a semi-supervised learning model based on non-linear feature selection is proposed to overcome the problem of prediction of disease. The wrapper-based hybrid correlation approach is used to partition the space of the function into k-related features. Ultimately, the deep neural network architecture was designed and implemented to enhance the prediction of disease on datasets of large dimensions. Experimental results have shown that the current model is more reliable compared to existing models in terms of true positivity and receiver operating characteristics (ROC).