Feature selection methods on gene expression microarray data for cancer classification: A systematic review

Abstract

This systematic review provides researchers interested in feature selection (FS) for processing microarray data with comprehensive information about the main research directions for gene expression classification conducted during the recent seven years. A set of 132 researches published by three different publishers is reviewed. The studied papers are categorized into nine directions based on their objectives. The FS directions that received various levels of attention were then summarized. The review revealed that ‘propose hybrid FS methods’ represented the most interesting research direction with a percentage of 34.9%, while the other directions have lower percentages that ranged from 13.6% down to 3%. This guides researchers to select the most competitive research direction. Papers in each category are thoroughly reviewed based on six perspectives, mainly: method(s), classifier(s), dataset(s), dataset dimension(s) range, performance metric(s), and result(s) achieved.

Introduction

Nowadays, classifying microarray datasets, which are called (large-scale biological data analysis), is a popular and attractive area of study for many researchers, as applying microarray technology is one of the most important applications in molecular biology for cancer detection [1]. It depends on developing more effective classification models that can be used for classifying any unseen microarray data after training the model over a specific training dataset. Detecting and classifying cancer, using microarray gene expressed data, have posed a huge challenge for researchers in the field of computer science, as this kind of datasets contains a small number of examples versus a huge number of genes. However, many of these genes are considered irrelevant or redundant, and they must be removed by using an efficient FS method for improving the performance of classification. Therefore, researchers have employed much effort in coming up with more effective FS techniques that can increase classification's accuracy and decrease the computation time using a smaller number of genes in diagnostic and prognostic prediction of tumor cancer [2].

FS is the process of selecting the most relevant and efficient features for improving classification's performance in high dimensional datasets [3]. Filters, wrappers, embedded, and hybrid methods are the main types of FS methods, but there is a new kind of FS methods that has been recently developed, which is called ensemble [4].

Filters use statistical measures for evaluating features against a class label. There are two categories of filters, mainly: ranking-based (univariate) and search-space-based (multivariate). The first category selects features that have higher ranks based on a specific threshold value, where ranks are provided according to the relationships between each feature and the specified class label for removing the most irrelevant features. In contrast, the second category takes care of the relationships within features. Therefore, it can remove irrelevant features in addition to redundant ones [5].

Wrappers depend on a classifier evaluation to select features as it selects feature sets that satisfy best results based on a fitness value for a classifier. This kind of methods consists of three parts: search algorithm, classifier, and fitness function [6], while embedded methods select features that automatically improve classification performance as a part of the learning stage [2]. See Fig. 1.

Ensemble methods have recently appeared for obtaining stability which does not exist in many FS methods. This will be achieved by aggregating the results of different feature subsets which were generated either by using the same FS method on various training data (homogeneous ensemble FS approach) or by applying various FS methods over the same training data (heterogeneous ensemble FS approach) [7]. See Fig. 1 (d and e).

Each approach has its negative and positive properties. For example, filter-based FS methods need less computational complexity and avoid over-fitting problems more than wrapper and embedded-based FS methods. In contrast, wrappers and embedded-based FS methods provide better accuracy than filters. Hybrid methods have advantages of both wrappers and filters as they result in achieving better accuracy than filters, and need less computational cost than wrappers. Ensemble methods are the most flexible of FS methods in high dimensional data, and are the least prone to over-fitting problems [2].

In literature, there are many reviews that concern FS in microarray data subjects, such as the works of [[8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23]], but to the best of our knowledge, there are no reviews done on categorizing these studies based on the main purpose of the reviewed papers.

In this systematic review, we categorize studies done recently on FS for microarray data processing into nine directions based on the main objectives of each research. Where some of these studies developed new FS methods which are (filter, wrapper, embedded, ensemble, distributed, parallel), other studies selected some FSM from the literature and compared their performance over the same environment. On the other hand, there are some researchers who conducted surveys on the subjects of FS.

Each research in each direction was reviewed based on methods, datasets, classifiers, performance metrics, datasets dimension's range used, and results obtained. Finally, we summarized some of the general observations that we noticed along with our conclusions.

The main contributions of this systematic review can be summarized as follows:

• It analyses 132 research papers from Elsevier, Springer, and IEEE publishers in the literature published in the last seven years about FS in microarray data processing according to their main objectives and categorizes these researches into nine main directions.

• It summarizes FS directions that received the highest, the middle, and the lowest research attention in the recent seven years for guiding researchers who plan to pursue research on FS for microarray data processing to choose their research direction.

• In each direction, each research paper is reviewed based on methods, datasets, classifiers, performance metrics, datasets dimensions range used, and results obtained. This aims to provide researchers with valuable information about the related work of their selected directions if they have already specified their research direction in FS for microarray data classification field as displayed in Fig. 2

The rest of this paper is organized as follows: Section 2 presents a background. Section 3 displays the methodology. Section 4 displays and discusses the main nine directions of FS in microarray data researches in the recent seven years that we proposed in this survey. The development of FS publications over the recent seven years is displayed in Section 5. In Section 6, the main observations are identified, and in Section 7, the conclusion and future work are presented.