Classification of ADHD with fMRI data and multi-objective optimization

Highlights

• A novel multi-objective optimization classification scheme is proposed.

• The scheme uses a cost sensitive three objective model to handle the class imbalance problem.

• A preferred subset of pareto optimal classifiers can be obtained based on the decision maker's preference.

• Results show that the proposed scheme performs considerably better than some traditional methods.

Abstract

Background and objective

Dataset imbalance is an important problem in neuroimaging. Imbalanced datasets would cause the performance degradation of a classifier by utilizing imbalanced learning, which tends to overfocus on the majority class. In this paper, we consider an imbalanced neuroimaging classification problem, namely, classification of attention deficit hyperactivity disorder (ADHD) using resting-state functional magnetic resonance imaging.

Methods

We propose a multi-objective classification scheme based on support vector machine (SVM). Our scheme addresses the imbalanced dataset problem by using a three objective SVM model with the positive and negative empirical errors being handled explicitly and separately. Moreover, an interactive multi-objective method incorporating the decision maker's preference is adopted, thus a preferred subset of pareto optimal classifiers for decision making can be obtained.

Results

The proposed scheme is assessed on five datasets from the ADHD- 200 consortium. Numerical results show that the proposed multi-objective scheme considerably outperforms some traditional classification methods in the literature.

Conclusion

The proposed multi-objective classification scheme avoids hyper-parameter selection, it effectively addresses dataset imbalanced problem from algorithm level. The scheme can not only be used in the diagnosis of ADHD but also in the diagnosis of other diseases, such as Alzheimer and Autism etc.

Introduction

Attention deficit hyperactivity disorder (ADHD) is a common childhood disorders which lasts to adulthood in most cases. It is defined as a neurodevelopmental disorder in the fifth edition of diagnostic and statistical manual of mental disorders, mainly characterized by attention deficits, excessive activity and behavioral impulses [1]. The worldwide pooled prevalence of ADHD is reported to be 3.4% in children and adolescents [2]. So far, the etiology and pathogenesis of ADHD is still not clear [3], and the diagnosis of ADHD currently relies mainly on the subjective experience of doctors. Therefore objective diagnosis and effective treatment of ADHD is one of the significant topics in the field of neuroscience.

Medical imaging technologies such as electroencephalography, functional near-infrared spectroscopy, magnetic resonance imaging (MRI), and functional magnetic resonance imaging (fMRI) have been used for computer-aided diagnosis of ADHD. see, e.g., [4], [5], [6], [7], [8], [9]. As a method of fMRI, resting-state fMRI (rs-fMRI) has shown prominent advantages in the pathological analysis of psychiatric diseases. Various feature extraction, selection and classification methods have been used in rs-fMRI based disease diagnosis. Castellanos et al. [10] found functional connectivity (FC) information of fMRI can be a prominent feature for ADHD diagnosis. Du et al. [11] used graph kernel principal components analysis (PCA) to extract features and proposed a discriminative subnetwork to classify ADHD. Miao and Zhang [12] discussed the classification of ADHD with fMRI data, and used relief algorithm to obtain a subset of fractional amplitude of low-frequency fluctuation features. Itani et al. [13] proposed a multi-level approach based on decision trees for ADHD classification. Qureshi et al. [14] computed the global connectivity maps of fMRI and used hierarchical extreme learning machine to classify ADHD. Riaz et al. [15] integrated non-imaging and imaging data and used a machine learning framework to study alterations of functional connectivity between ADHD and normal control (NC) subjects. Considering the data imbalanced property, they generated synthetic minority class samples using synthetic minority over-sampling technique (SMOTE) [16]. It needs to be noted that recently multi-objective evolutionary computation algorithms, such as multi-objective particle swarm algorithm [17] and multi-objective self-adaptive particle swarm algorithms [18] etc. have shown some success in feature selection (see [19] for a survey) and can also be used for rs-fMRI feature selection.

Most of the classification algorithms mentioned above make the assumption of well-balanced training datasets and equal misclassification costs. However, dataset imbalance is a critical problem in ADHD rs-fMRI datasets. Due to imbalanced learning, imbalanced datasets may lead to overfocus on the majority class, hence degrade the performance of a classifier. There are many approaches proposed to handle the imbalanced dataset problem (also called the class imbalance problem), see [20], [21], [22], [23]. These approaches basically fall into two major categories: data level approaches and algorithm level approaches. The main idea of data level approaches to handle the imbalance is to resample the training set, whereas algorithm level approaches normally adopt the idea of introducing unequal misclassification costs in decision making process. In ADHD classification, data level approaches such as SMOTE have been used to handle the dataset imbalance. However, by performing random oversampling the minority or under-sampling the majority class, these strategies for creating balanced training datasets may lead to suboptimal performance [24].

Considering the multi-objective nature of classification problems, Shao et al. [25] have proposed a bi-objective classification method for the classification of ADHD. However, dataset imbalance is not considered. In this work, incorporating a decision maker's preference, we propose a novel reference point based three objective optimization classification scheme. The main contributions of our work are as follows.

• We propose using a cost sensitive three objective classification model which is based on SVM to handle the ADHD dataset imbalance problem.

• From a practical viewpoint, an interactive multi-objective optimization method incorporating decision maker's preference information is proposed. A preferred subset of pareto optimal classifiers can be obtained, thus a classifier with the best performance can be selected.

The rest of the paper is organized as follows. In Section 2, we first introduce the acquisition and preprocessing of data. Then the three objective classification scheme for ADHD diagnosis is proposed. As a main part of the scheme, a three objective classification model based on L₁-norm SVM is introduced and a reference point based multi-objective optimization method is proposed. Section 3 shows some computational experiments. An interactive multi-objective decision making example and some comparison results of the three objective classification scheme and some other methods are given. In Section 4, we give some further discussions about the results. Finally we draw the conclusion in Section 5.