Multi-stage feature selection in identifying potential biomarkers for cancer classification

• Main Authors: Wong, Yit Khee, Chan, Weng Howe, Nies, Hui Wen, Moorthy, Kohbalan
• Format: Conference or Workshop Item
• Language: English; English
• Published: Institute of Electrical and Electronics Engineers Inc. 2022
• Subjects: QA75 Electronic computers. Computer science; QA76 Computer software; T Technology (General); TA Engineering (General). Civil engineering (General)
• Online Access: http://umpir.ump.edu.my/id/eprint/39085/1/Multi-stage%20feature%20selection%20in%20identifying%20potential%20biomarkers.pdf; http://umpir.ump.edu.my/id/eprint/39085/2/Multi-stage%20feature%20selection%20in%20identifying%20potential%20biomarkers%20for%20cancer%20classification_ABS.pdf; http://umpir.ump.edu.my/id/eprint/39085/; https://doi.org/10.1109/ICICyTA57421.2022.10037807

Biomarkers are indicators that show the disease state or its progression of certain health conditions. Identification of biomarkers greatly raises the probability of earlier diagnosis and could be further applied in developing effective treatment for the disease. Besides conducting laboratory analysis, potential biomarkers also can be identified by analysing gene expression data through feature selection and machine learning. Many algorithms have been applied and introduced in this area, yet the challenge of high dimensionality of gene expression data remains and it could lead to the existence of noise that could negatively impact the analysis outcome. Therefore, this study aims to investigate and develop a better feature selection to identify potential biomarkers from gene expression data and construct a deep neural network classification model using these selected features. Thus, a multistage feature selection, namely CIR is proposed, that composed of Chi-square, Information Gain and Recursive Feature Elimination. The dataset used in this study consists of the integration of seven ovarian cancer gene expression datasets from GEO database. Both selected genes and classification model are evaluated through biological context verification and classification performance respectively. The proposed method shows improvements over the existing methods in terms of accuracy (+2.2294%), precision (+8.1415%), recall (+2.2294%), Fl-score (+4.5494%) and AUC scores (+0.2302). The proposed CIR method successfully identified eight genes that could be potential biomarkers for ovarian cancer, including WFDC2,S100A13, PRG4, NRCAM, OGN, B3GALT2, VGLL3, and GATM which are further verified through literature.