Elsevier

Pattern Recognition Letters

Volume 138, October 2020, Pages 82-87
Pattern Recognition Letters

Improved ASD classification using dynamic functional connectivity and multi-task feature selection

https://doi.org/10.1016/j.patrec.2020.07.005Get rights and content

Highlights

  • We propose an efficient method for diagnosing ASD.

  • Dynamic functional connectivity features have complementary information.

  • We proposed an improved multi-task feature selection method.

  • A multi-kernel SVM method is applied.

  • Our method achieves classification accuracy of 76.8%.

Abstract

Accurate diagnosis of autism spectrum disorder (ASD), which is a neurodevelopmental disorder and often accompanied by abnormal social skills, communication skills, interests and behavior patterns, has always been a challenging task in clinical practice. Recent studies have shown great potential for using fMRI data to distinguish ASD from typical control (TC). However, it has always been a challenging problem to extract which features from fMRI data and how to combine these different types of features to achieve improved ASD/TC classification performance. To address this problem, in this study we propose an improved ASD/TC classification framework based on dynamic functional connectivity (DFC) and multi-task feature selection. Our proposed ASD/TC classification framework is evaluated on 871 subjects with fMRI data from the Autism Brain Imaging Data Exchange I (ABIDE I) via a 10-fold cross validation strategy. Experimental results show that our proposed method achieves an accuracy of 76.8% and an area under the receiver operating characteristic curve (AUC) of 0.81 for ASD/TC classification. In addition, compared with some existing state-of-the-art methods, our proposed method achieves better accuracy and AUC for ASD/TC classification. Overall, our proposed ASD/TC classification framework is effective and promising for automatic diagnosis of ASD in clinical practice.

Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental disorder and often accompanied by abnormal social skills, communication skills, interests and behavior patterns [16], [21], [38]. The patients with ASD are basically children from 2 to 9 years of age [30]. A recent survey from the Autism and Developmental Disabilities Monitoring (ADDM) Network in Centers for Disease Control and Prevention (CDC) show that about 1 in 59 children worldwide has ASD [4]. As a result of these symptoms, the patients with ASD have caused great economic burden for their families and society [2]. In the United States, the total annual costs for children with ASD is estimated to be between 11.560.9 billion [6]. However, since the etiology of ASD is unknown and there is no medical test (such as blood and urine tests) for the diagnosis of ASD, so far it has been a very difficult problem to accurately diagnose whether a person has ASD. Therefore, in order to treat the patients with ASD more effectively, it is urgent to develop an effective method for accurate diagnosis of ASD.

Nowadays, magnetic resonance imaging (MRI) [23] has provided a way for clinicians to examine the structural and functional changes associated with the development of diseases in vivo [12], [13], [24], and has been widely applied in clinical practice. At present, commonly used MRI mainly includes structural MRI (sMRI) and functional MRI (fMRI). Compared with sMRI, since fMRI [15] can measure the changes in hemodynamics caused by neuron activity at a series of time points for the whole brain, it has been widely applied in the research of brain dysfunction diseases. For example, in order to explore the feature representation related to ASD, [1] proposed a machine learning method based on fMRI data to perform ASD classification[36]; proposed a multi-view feature representation based on fMRI data to perform Schizophrenia classification and find the imaging markers related to Schizophrenia. Furthermore, commonly used fMRI mainly includes resting-state fMRI (rs-fMRI) and task-state fMRI (ts-fMRI). Compared with ts-fMRI, rs-fMRI attracts more attention because it is easier to acquire data. In this study we focus on rs-fMRI data.

With the rapid development of machine learning methods [5], recent studies have shown great potential for integrating fMRI data and machine learning methods to automatically diagnose brain diseases, such as Schizophrenia [36], [37], Alzheimer’s disease [7], [20], [26] and ASD [1], [19], [22], [32]. For example, [32] used rs-fMRI data as the inputs of graph convolutional network (GCN) to perform ASD classification, and obtained 70.4% accuracy. Recently, following the work of [19], [32] proposed an InceptionGCN method based on fMRI data to perform ASD classification task. However, although some results have been achieved by integrating fMRI data and machine learning methods for ASD classification, it has always been a challenging problem to extract which features from rs-fMRI data and how to combine these different types of features to achieve improved ASD/TC classification performance.

Based on the above analysis, to improve the performance of ASD/TC classification, in this study we propose a new ASD/TC classification framework based on dynamic functional connectivity (DFC) and multi-task feature selection as shown in Fig. 1. As can be seen from Fig. 1, our proposed ASD/TC classification framework mainly includes four steps. Firstly, we extract the time series of each brain region based on automated anatomical labeling (AAL) [35] atlas from rs-fMRI data of each subject. Secondly, we further extract the DFC by computing Pearson correlation coefficient (PCC) between paired brain regions via successive and non-overlapping time windows as feature representation of each subject. Thirdly, in order to obtain features that are more helpful from each feature set for ASD/TC classification, an improved multi-task feature selection method integrating elastic net and manifold regularization is proposed and denoted as MTFS-EM. Finally, a multi-kernel SVM learning method is adopted to perform the ASD/TC classification task by combining multiple selected feature sets. Our proposed ASD/TC classification framework is evaluated on 871 subjects (including 403 subjects with ASD and 468 TCs) with rs-fMRI data from the Autism Brain Imaging Data Exchange I (ABIDE I) via a 10-fold cross validation strategy.

Section snippets

Subjects and image preprocessing

The subjects involved in this study are provided by the Autism Brain Imaging Data Exchange I (ABIDE I1) [11]. The ABIDE I includes 539 subjects with ASD and 573 typical controls (TCs) from 17 international sites. For fair comparison, in this study we choose the same 871 subjects including 403 subjects with ASD and 468 TCs from the ABIDE I as some existing literature [1], [19], [32] as shown in Table 1. In Table 1, m ± std and M/F are

Experimental settings

Our proposed ASD/TC classification framework is performed via a 10-fold cross-validation strategy [27] and repeated 50 times. Before performing our proposed framework on ASD/TC classification, all parameters in the proposed framework are set. In the dynamic functional connectivity procedure, T is set to 4. In the multi-task feature selection procedure, α, γ and β are set to a range from 0 to 1 at a step size of 0.1, respectively, i.e., α={0,0.1,,0.9,1}, γ={0,0.1,,0.9,1}, and β={0,0.1,,0.9,1}

Impact of different number of time windows

In this section, we discuss the impact of different number of time windows (i.e., T) in our proposed framework for the performance of ASD/TC classification. In order to investigate the impact of different T in our proposed framework, we have done a series of experiments based on T={1,2,,9,10} for ASD/TC classification. Fig. 2 shows the accuracies of different T in our proposed framework for ASD/TC classification. It’s worth mentioning that when T=1, the functional connectivity between brain

Conclusion

In this study, we propose an improved ASD/TC classification framework based on dynamic functional connectivity and multi-task feature selection. Experimental results demonstrate that our proposed framework is effective for ASD/TC classification. This method paves the way to discriminative imaging markers for computer-aided diagnosis of ASD.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work is supported in part by the National Natural Science Foundation of China under Grant No.61802442, No.61877059, the Natural Science Foundation of Hunan Province under Grant No.2019JJ50775, No.2018JJ2534, the 111 Project (No. B18059), and the Hunan Provincial Science and Technology Program (2018WK4001).

References (40)

  • N. Tzourio-Mazoyer et al.

    Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain

    Neuroimage

    (2002)
  • D. Amendah et al.

    The economic costs of autism: a review

    Autism Spect. Disord.

    (2011)
  • A. Argyriou et al.

    Multi-task feature learning

    Advances in neural information processing systems

    (2007)
  • J. Baio et al.

    Prevalence of autism spectrum disorder among children aged 8 yearsutism and developmental disabilities monitoring network, 11 sites, united states, 2014

    MMWR Surveill. Summ.

    (2018)
  • C.M. Bishop

    Pattern Recognition and Machine Learning

    (2006)
  • A.V. Buescher et al.

    Costs of autism spectrum disorders in the united kingdom and the united states

    JAMA Pediatr

    (2014)
  • C.-C. Chang et al.

    Libsvm: a library for support vector machines

    ACM Trans. Intell. Syst. Technol. (TIST)

    (2011)
  • C. Craddock et al.

    The neuro bureau preprocessing initiative: open sharing of preprocessed neuroimaging data and derivatives

    Neuroinformatics

    (2013)
  • C. Craddock et al.

    Towards automated analysis of connectomes: the configurable pipeline for the analysis of connectomes (c-pac)

    Front Neuroinform

    (2013)
  • A. Di Martino et al.

    The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism

    Mol. Psychiatry

    (2014)
  • Cited by (44)

    • A review of methods for classification and recognition of ASD using fMRI data

      2022, Journal of Neuroscience Methods
      Citation Excerpt :

      In the case of the ABIDE dataset, it is difficult to extract useful features owing to the heterogeneity of subjects and the difference in data capture between sites. At present, fMRI data are mainly defined by FC (Sun et al., 2021b; Shi et al., 2021; Grana and Silva, 2021; Reiter et al., 2021; Spera et al., 2019), DFC (Liu et al., 2020a; Zhao et al., 2020), graph theory properties (Sun et al., 2021a; Chaitra et al., 2020; Wang et al., 2020; Kazeminejad and Sotero, 2020; Brahim and Farrugia, 2020), To achieve higher classification and recognition performance of ASD, future work will explore new algorithms to obtain more useful features, which can better identify biomarkers. To reduce model overfitting and improve its generalization capability, more data are input into the model to improve machine learning and deep learning performance.

    • Brain function and brain connectivity in autism

      2022, The Neuroscience of Autism
    View all citing articles on Scopus
    View full text