The Nearest Shrunken Centroid (NSC) method, with Prediction Analysis for Microarrays being its most well known implementation, has been widely used as a classification method for high dimensional biomedical data. A threshold value must also be provided in this method as input and normally, this is selected manually on a “trial and error” basis by executing the NSC method many times using a number of predetermined shrinkage threshold values. The optimal value is then obtained by minimizing the cross-validated error on the training data. This process can be time-consuming and the optimal threshold value may be limited by the granularity of the predetermined values. In this paper, an approach incorporating the NSC method and a multi-objective evolutionary algorithm, Non-dominated Sorting Algorithm 2, is proposed for obtaining the optimal shrinkage threshold value automatically. The NSC method acts as the fitness evaluator in the evolutionary process. Multiple objectives can be incorporated for determining the threshold values and a number of optimal solutions are obtained, each on the basis of tradeoffs between the objectives. By providing multiple potential solutions, it allows biomedical experts to better explore the joint behaviors of features in their data. The proposed approach also overcomes limitations normally associated with single objective approaches; a single optimum and the need to determine weightings associated with various objective functions in an aggregated objective function. The proposed approach was evaluated using the Alzheimer's Disease, Colon and Leukemia cancer dataset.

NSC-NSGA2: Optimal search for finding multiple thresholds for nearest shrunken centroid

Data mining techniques have been used widely in many areas such as business, science, engineering and medicine The techniques allow a vast amount of data to be explored in order to extract useful information from the data One of the foci in the health area is finding interesting biomarkers from biomedical data Mass throughput data generated from microarrays and mass spectrometry from biological samples are high dimensional and is small in sample size Examples include DNA microarray datasets with up to 500,000 genes and mass spectrometry data with 300,000 m/z values While the availability of such datasets can aid in the development of techniques/drugs to improve diagnosis and treatment of diseases, a major challenge involves its analysis to extract useful and meaningful information The aims of this project are: 1) to investigate and develop feature selection algorithms that incorporate various evolutionary strategies, 2) using the developed algorithms to find the “most relevant” biomarkers contained in biological datasets and 3) and evaluate the goodness of extracted feature subsets for relevance (examined in terms of existing biomedical domain knowledge and from classification accuracy obtained using different classifiers) The project aims to generate good predictive models for classifying diseased samples from control

/pdf/evolutionary-approaches-for-feature-selection-in-biological-4sa69zav1t.pdf

Evolutionary approaches for feature selection in biological data

sIdentifying individuals with early mild cognitive impairment (EMCI) can be an effective strategy for early diagnosis and delay the progression of Alzheimers disease (AD). Many approaches have been devised to discriminate those with EMCI from healthy control (HC) individuals. Selection of the most effective parameters has been one of the challenging aspects of these approaches. In this study we suggest an optimization method based on five evolutionary algorithms that can be used in optimization of neuroimaging data with a large number of parameters. Resting-state functional magnetic resonance imaging (rs-fMRI) measures, which measure functional connectivity, have been shown to be useful in prediction of cognitive decline. Analysis of functional connectivity data using graph measures is a common practice that results in a great number of parameters. Using graph measures we calculated 1155 parameters from the functional connectivity data of HC (n=36) and EMCI (n=34) extracted from the publicly available database of the Alzheimers disease neuroimaging initiative database (ADNI). These parameters were fed into the evolutionary algorithms to select a subset of parameters for classification of the data into two categories of EMCI and HC using a two-layer artificial neural network. All algorithms achieved classification accuracy of 94.55%, which is extremely high considering single-modality input and low number of data participants. These results highlight potential application of rs-fMRI and efficiency of such optimization methods in classification of images into HC and EMCI. This is of particular importance considering that MRI images of EMCI individuals cannot be easily identified by experts.

https://www.biorxiv.org/content/biorxiv/early/2021/03/05/2021.03.04.433989.full.pdf

Evolutionary optimization in classification of early-MCI patients from healthy controls using graph measures of resting-state fMRI

A general problem faced in classification is a small number of samples compared with the large number of genes. In this case, the feature selection (FS) process becomes a challenging task to improve the classification performance by reducing the dimensionality. Particle swarm optimisation (PSO) is a powerful method for solving FS problems. The main ingredient of the binary PSO is its transfer function that allows mapping a continuous space to a discrete space. A sigmoid function was used to update positions in BPSO. However, due to its way of updating positions, this function is not very effective to dodge local minima and speed up the convergence. Thus, this paper suggests an enhanced BPSO, through the FS approach. The study employs a V type transfer function and a special method for updating positions. In addition, a hybrid FS method that integrates the information gain (IG) as a filter approach with the wrapper approach (EBPSO) is proposed. In a hybrid model, feature subsets are ranked based on their significance in a decreasing order. The EBPSO is applied to these subsets and then FS is performed. The proposed algorithm is tested on six microarray datasets and the results verify its superiority. Also, the results of the proposed hybrid FS model, supports its effectiveness because the model produced a small feature subset that showed high classification performance.

An Enhanced Binary Particle Swarm Optimization (EBPSO) algorithm based a V-shaped transfer function for feature selection in high dimensional data

Feature set selection is the process of identifying a subset of features which produces the result same as the entire set. The feature set selection helps in clustering the datasets. In this paper, a Highly Correlated Feature set Selection (HCFS) algorithmis proposed for clustering the data. This algorithm helps in selecting features based on its relevancy and redundancy factors. All the selected features are finally clustered based on how they are correlated with each other. The main objective of this paper is to identify the feature subsets which will improve the classification performance by constructing minimum spanning tree (MST) between the features.The HCFS algorithm works in two steps. In the first step, the features are divided into clusters using the spanning tree construction process. In the second step, the cluster representatives are selected using Frequent Pattern Analysis (FPA) technique to form the effective feature set which reduces the time required for query evaluation process. The redundant and irrelevant features are removed based on their Symmetric Uncertainty (SU) values. This effectively improves the efficiency of data clustering process.

Highly correlated feature set selection for data clustering

In this paper, a hybrid approach incorporating genetic algorithm and rough set theory into Feature Selection is proposed for searching for the best subset of optimal features The approach utilizes K-means clustering for partitioning attribute values, the rough set-based approach for reducing redundant data, and the genetic algorithm for searching for the best subset of features A set of six attributes was obtained as the best subset using the proposed algorithm on the colon cancer dataset Classification was carried out using this set of six attributes with 23 classifiers from WEKA (Waikato Environment for Knowledge Analysis) software to examine their significance to classify unseen test data In addition, the set of 6 genes found by the proposed approach was also examined for their relevance to known biomarkers in the colon cancer domain

Incorporating genetic algorithm into rough feature selection for high dimensional biomedical data

In this paper, a hybrid approach incorporating the Nearest Shrunken Centroid (NSC) and Genetic Algorithm (GA) is proposed to automatically search for an optimal range of shrinkage threshold values for the NSC to improve feature selection and classification accuracy for high dimensional data. The selection of a threshold value is crucial as it is the key factor in the NSC to find significant relative differences between the overall centroid and the class centroid. However, selecting this threshold value via “trial and error” in empirical approaches can be time-consuming and imprecise. In the proposed NSC-GA approach, shrinkage threshold values for the NSC are encoded as genes in chromosomes that are evaluated using a fitness measure obtained from the classifier in the NSC. The proposed approach automatically searches for the optimal threshold for the NSC by utilizing GA. The proposed approach was evaluated using a number of data sets; Alzheimer's disease, Colon and Leukemia cancer datasets. Experimental results indicated that the proposed approach finds the optimal range of shrinkage thresholds for each dataset, subsequently leading to a higher classification result and involving a smaller number of features when compared to previous studies.

NSC-GA: Search for optimal shrinkage thresholds for nearest shrunken centroid

High-throughput technologies such as microarrays and mass spectrometry produced high dimensional

biological datasets both in abundance and with increasing complexity. Prediction Analysis for Microarrays

(PAM) is a well-known implementation of the Nearest Shrunken Centroid (NSC) method which has been

widely used for classification of biological data. In this paper, a hybrid approach incorporating the Nearest

Shrunken Centroid (NSC) and Memetic Algorithm (MA) is proposed to automatically search for an optimal

range of shrinkage threshold values for the NSC to improve feature selection and classification accuracy.

Evaluation of the approach involved nine biological datasets and results showed improved feature selection

stability over existing evolutionary approaches as well as improved classification accuracy.

Vinh Quoc Dang

Papers

NSC-NSGA2: Optimal search for finding multiple thresholds for nearest shrunken centroid

Incorporating genetic algorithm into rough feature selection for high dimensional biomedical data

NSC-GA: Search for optimal shrinkage thresholds for nearest shrunken centroid

Gene Selection using a Hybrid Memetic and Nearest Shrunken Centroid Algorithm