A hybrid filter-based genes selection stochastic model for cancer prediction and classification based on gene expression data

The study of high-dimensional genes expression data has made an important role in disease diagnosis and cancer type classification. Nevertheless, due to the problem of curse of dimensionality, dimension reduction methods, particularly feature selection (FS) methods, are crucial for eliminating redundant genes and improving disease classification. Stochastic FS models are vital class of FS models in the analysis of genes expression data. Despite the availability of some stochastic FS models, each model has its own limitations. In this study, we propose a hybrid stochastic FS model that embedding multivariate filtering in a hidden Markov-model (HMM) framework. It addresses the problem of redundancy and applicable for binary and multi-class problems, therefore, we named Generalized Multivariate FS based HMM (GMFS-HMM). On colon data, the new GMFS-HMM model outperforms the state-of-the-art HMM model, achieving the highest accuracies (>90%) for Randon Forest (RF) and across 10 to 30 selected genes. Gene enrichment analysis of colon data further validated this performance of GMFS-HMM. Moreover, the applicability and performance of the GMFS-HMM was also demonstrated on three multi-class datasets, including high-dimensional RNA-seq Pan-cancer data. The analysis on benchmark datasets illustrates that GMFS-HMM has improved classification accuracy and applicability on binary as well as multi-class datasets.