Figure 1.

EpICC combines Bayesian Neural Network (BNN) with uncertainty correction. BNN utilizes the gene expression data of feature genes for cancer classification. The BNN consists of 3 layers with the first layer consisting 250 neurons, the second layer containing 95 neurons and the final layer consists of output neurons. The number of output neurons is dependent on the number of classes to be predicted. The weights of the connections were initialized from prior probability distributions. We refined the weights over multiple iterations through the BNN. The output was used for uncertainty estimation. After estimating the uncertainty, we tested two different approaches for incorporating uncertainty to improve classification accuracy—uncertainty filtering and uncertainty correction. We thus obtained the filtered and the corrected outputs respectively which we used for cancer type and subtype prediction.