Machine-learning-based diagnosis of schizophrenia using combined sensor-level and source-level EEG features

https://doi.org/10.1016/j.schres.2016.05.007Get rights and content

Abstract

Recently, an increasing number of researchers have endeavored to develop practical tools for diagnosing patients with schizophrenia using machine learning techniques applied to EEG biomarkers. Although a number of studies showed that source-level EEG features can potentially be applied to the differential diagnosis of schizophrenia, most studies have used only sensor-level EEG features such as ERP peak amplitude and power spectrum for machine learning-based diagnosis of schizophrenia. In this study, we used both sensor-level and source-level features extracted from EEG signals recorded during an auditory oddball task for the classification of patients with schizophrenia and healthy controls. EEG signals were recorded from 34 patients with schizophrenia and 34 healthy controls while each subject was asked to attend to oddball tones. Our results demonstrated higher classification accuracy when source-level features were used together with sensor-level features, compared to when only sensor-level features were used. In addition, the selected sensor-level features were mostly found in the frontal area, and the selected source-level features were mostly extracted from the temporal area, which coincide well with the well-known pathological region of cognitive processing in patients with schizophrenia. Our results suggest that our approach would be a promising tool for the computer-aided diagnosis of schizophrenia.

Introduction

Schizophrenia is diagnosed primarily using diagnostic criteria from the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), by asking patients a series of questions to elicit information such as duration of illness and clinical symptoms (American Psychiatric Association, 2013). The clinical symptom severity of schizophrenia is measured using clinical scales such as the Positive and Negative Syndrome Scale (PANSS) (Kay et al., 1988). Various diagnostic tools can help psychiatrists and clinical psychologists diagnose schizophrenia, but traditional clinical diagnoses might be sometimes inaccurate because schizophrenia patients sometimes intentionally hide their symptoms, and even experts sometimes have difficulty differentiating schizophrenia from other mental illnesses due to similar symptoms (Lindstrom et al., 1994, McGorry et al., 1995, Norman et al., 1996). Thus, many researchers have sought to develop objective, quantitative biomarkers that can enhance the overall accuracy of diagnosis with the aid of neuroimaging technologies. Among a variety of neuroimaging modalities, electroencephalography (EEG) is regarded as one of the most useful, thanks to its high temporal resolution and low cost. Many studies report disrupted cerebral information processing in schizophrenia, in the context of altered event-related potential (ERP) waveforms (Odonnell et al., 1995, Ozgurdal et al., 2008, Turetsky et al., 1998, Wang et al., 2003), disrupted functional connectivity patterns (Lynall et al., 2010, Winterer et al., 2003), and reduced source activity (Kawasaki et al., 2007, Kim et al., 2014, Pae et al., 2003, Wang et al., 2010).

Recently, an increasing number of researchers have attempted to differentiate patients with schizophrenia from healthy controls using machine learning (ML) methods with EEG biomarkers. Some of these studies used sensor-level biomarkers, such as ERP amplitude and latency, as features for classification (Neuhaus et al., 2013, Neuhaus et al., 2011). For example, Neuhaus et al. (2011) sought to identify schizophrenia using amplitudes/latencies of N100 and P300 that were evoked using visual and auditory oddball paradigms, respectively, and they reported a fairly high classification accuracy of 72.4%. Later, the same group (Neuhaus et al., 2013) reported improved classification accuracy of 79.0% using ERP components that were evoked in a visual target-locked paradigm. Previous studies including the ones described above used only sensor-level features such as ERP peak amplitudes and latencies extracted from raw EEG signals. However, these sensor-level features have an inherent limitation in that the signals can be distorted and smeared due to volume conduction (Nolte et al., 2004, Nunez et al., 1997, van den Broek et al., 1998), and thus potentially lose important information regarding underlying cortical activity. In fact, some EEG-based brain-computer interface (BCI) studies that classified different brain activity patterns during mental imagery tasks have reported increased classification accuracy using source-level features rather than sensor-level features (Ahn et al., 2012, Kamousi et al., 2007, Qin et al., 2004). However, these source-level features have not been widely applied to clinical applications, and have especially never been applied to machine-learning based diagnosis of schizophrenia.

In this study, we used both sensor-level features and source-level features for the differentiation of schizophrenia patients and healthy controls. We hypothesized that simultaneous use of both sensor-level and source-level features would enhance classification accuracy. To test this hypothesis, we used EEG data recorded while participants were performing an auditory oddball task, of which the results were known to be relatively consistent through a series of previous studies. The P300 amplitude evoked by the auditory oddball paradigm is significantly decreased in schizophrenia patients compared to healthy controls, and source activity is also reduced in patients with schizophrenia. In the present study, we compared classification accuracies for three different cases: ML with sensor-level features only, ML with source-level features only, and ML with combined two-level features.

Section snippets

Participants

Thirty-four patients with schizophrenia (20 males and 14 females) and 34 healthy controls (14 males and 20 females) were recruited for this study from the Psychiatry Department of Inje University Ilsan Paik Hospital. Patients who had diseases of the central nervous system, medical histories of alcohol and drug abuse, experience with electrical therapy, mental retardation, or head injuries with loss of consciousness were excluded from the study by initial screening interviews. The patients were

Maximum and average classification accuracy

The number of features varied from one to 20, for each of which the classification accuracy was evaluated. Table 2 summarizes the classification accuracies for three different feature sets with respect to the number of features used for the classification. This table also includes information on the ratio of sensor-level features to source-level features in the features selected from the combined feature set. A maximum classification accuracy of 88.24% was reported for the combined-level

Discussion

In this study, we demonstrated that simultaneous use of sensor-level and source-level feature sets could improve overall classification accuracy in the machine-learning-based diagnosis of schizophrenia. A maximum classification accuracy of 88.24% was obtained when the combined feature set was used, whereas the highest classification accuracies were 80.88% and 85.29% for sensor-level and source-level feature sets, respectively. The average classification accuracy of the combined feature set

Conflict of interest

All the authors declare that they have no conflicts of interest.

Contributors

Miseon Shim designed the study and wrote the manuscript. Han-Jeong Hwang developed the machine learning program used for the analysis. Do-Won Kim processed EEG data. Seung-Hwan Lee designed the study and wrote the protocol. Chang-Hwan Im supervised the study process and manuscript writing. All authors contributed to and have approved the final manuscript.

Role of funding sources

This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean Government (MSIP) (NRF-2015M3C7A1031969, 2014R1A2A1A11051796 and NRF-2015R1A5A7037676). All funding sources had no further role in study design; in the collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the paper for publication.

Acknowledgements

This research was supported in part by the Brain Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT & Future Planning (NRF-2015M3C7A1031969) and in part by the National Research Foundation of Korea (NRF) grants funded by the Korean Government (MSIP) (Nos. 2015R1A2A1A15054662 and NRF-2015R1A5A7037676).

References (46)

  • M.S. Kim et al.

    Neuropsychological correlates of P300 abnormalities in patients with schizophrenia and obsessive–compulsive disorder

    Psychiatry Res.

    (2003)
  • G. Kuperberg et al.

    Schizophrenia and cognitive function

    Curr. Opin. Neurobiol.

    (2000)
  • I. Lazzaro et al.

    Single trial variability within the P300 (250–500 ms) processing window in adolescents with attention deficit hyperactivity disorder

    Psychiatry Res.

    (1997)
  • D.H. Mathalon et al.

    Trait and state aspects of P300 amplitude reduction in schizophrenia: a retrospective longitudinal study

    Biol. Psychiatry

    (2000)
  • A.H. Neuhaus et al.

    Single-subject classification of schizophrenia by event-related potentials during selective attention

    NeuroImage

    (2011)
  • G. Nolte et al.

    Identifying true brain interaction from EEG data using the imaginary part of coherency

    Clin. Neurophysiol.

    (2004)
  • R.M.G. Norman et al.

    A study of the interrelationship between and comparative interrater reliability of the SAPS, SANS and PANSS

    Schizophr. Res.

    (1996)
  • P.L. Nunez et al.

    EEG coherency: I: statistics, reference electrode, volume conduction, Laplacians, cortical imaging, and interpretation at multiple scales

    Electroencephalogr. Clin. Neurophysiol.

    (1997)
  • S. Ozgurdal et al.

    Reduction of auditory event-related P300 amplitude in subjects with at-risk mental state for schizophrenia

    Schizophr. Res.

    (2008)
  • J.S. Pae et al.

    LORETA imaging of P300 in schizophrenia with individual MRI and 128-channel EEG

    NeuroImage

    (2003)
  • S. Pakarinen et al.

    Measurement of extensive auditory discrimination profiles using the mismatch negativity (MMN) potential of the auditory event-related (ERP)

    Clin. Neurophysiol.

    (2007)
  • J. Polich et al.

    Cognitive and biological determinants of P300 — an integrative review

    Biol. Psychol.

    (1995)
  • J. Polich et al.

    P300 latency reflects the degree of cognitive decline in dementing illness

    Electroencephalogr. Clin. Neurophysiol.

    (1986)
  • Cited by (0)

    View full text