A Hybrid Deep Learning Framework for Multi-Class High Dimensional Cancer Prediction
Abstract
Predicting medical disease patterns in high-dimensional features becomes more difficult in biomedical applications as medical datasets grow larger. Pathogenesis of several diseases relies heavily on micro-array data. MicroRNA interactions with disease are still challenging to anticipate experimentally. Computer methods previously used to classify probable gene-disease connections have also been identified as lacking. Without the identification of genes and semi-supervised learning, many current techniques cannot predict disease. In the interim, multiple additional approaches failed to highlight correlations for all diseases at the same time. Since existing gene-disease connections have been proven by biological experiment, an algorithm is needed to classify valid disease possibilities. As a solution to these challenges, a semi-supervised learning approach based on non-linear feature selection is proposed. The feature space is divided into k-correlated features using a hybrid correlation-based wrapper technique. For the first time, a deep neural network architecture for disease prediction has been devised and implemented. Real positivity and receiver operating characteristics reveal that the current model is more accurate than previous models (ROCs).