Automated EEG pathology detection based on different convolutional neural network models: Deep learning approach

https://doi.org/10.1016/j.compbiomed.2021.104434Get rights and content

Highlights

  • Visual evaluation of pathology EEG is laborious and often subjective.

  • Combination of deep CNN with SVM classifier in automated EEG pathology detection.

  • Attained maximum accuracy of 96.65% using SeizureNet-SVM-based system.

  • The proposed framework may aid the clinicians in diagnosing and for early treatment.

Abstract

The brain electrical activity, recorded and materialized as electroencephalogram (EEG) signals, is known to be very useful in the diagnosis of brain-related pathology. However, manual examination of these EEG signals has various limitations, including time-consuming inspections, the need for highly trained neurologists, and the subjectiveness of the evaluation. Thus, an automated EEG pathology detection system would be helpful to assist neurologists to enhance the treatment procedure by making a quicker diagnosis and reducing error due to the human element. This work proposes the application of a time-frequency spectrum to convert the EEG signals onto the image domain. The spectrum images are then applied to the Convolutional Neural Network (CNN) to learn robust features that can aid the automatic detection of pathology and normal EEG signals. Three popular CNN in the form of the DenseNet, Inception-ResNet v2, and SeizureNet were employed. The extracted deep-learned features from the spectrum images are then passed onto the support vector machine (SVM) classifier. The effectiveness of the proposed approach was assessed using the publicly available Temple University Hospital (TUH) abnormal EEG corpus dataset, which is demographically balanced. The proposed SeizureNet-SVM-based system achieved state-of-the-art performance: accuracy, sensitivity, and specificity of 96.65%, 90.48%, and 100%, respectively. The results show that the proposed framework may serve as a diagnostic tool to assist clinicians in the detection of EEG pathology for early treatment.

Introduction

The Electroencephalogram (EEG) is a graph that represents brain activity by plotting electrical potential measured at different points of the brain over time, first recorded in the year 1924 by Hans Berger [1]. These days, it is typically an inexpensive procedure requiring a mesh of non-invasive electrodes placed on the scalp to capture the signals, which is the reason it is still a popular choice for many use cases over relatively new technologies like Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans. EEG recordings can be used by neurologists to diagnose a range of conditions affecting the brain such as epilepsy, sleep disorders, etc. [[2], [3], [4]]. The diagnosis via EEG interpretation typically involves monitoring or recording EEG signals from multiple sessions for a long period of time. Data generated through this process is substantial and subsequently needs to be manually interpreted by a highly trained neurologist. This process is time-consuming, tedious, and prone to inaccuracy. Besides, the interpretation result can be subjective with low intra-rater and inter-rater reliability. Consequently, if some degree of the EEG interpretation is automated, it could assist in reducing the neurologist's workload by accelerating the inspection process.

Due to the ability of computers to process large amounts of data quickly, machine learning techniques have become a popular choice for the automation of EEG interpretation in recent times. In the diagnosis of a neurological condition, the first step is often to determine whether there are any abnormal patterns in the EEG recording [5]. This decision can help decide whether further investigation is necessary. It can also affect which medication is being prescribed. Currently, physicians typically follow an intricate decision tree to make this differentiation [5]. Hence, the motivation of our work is to have this first step of interpretation automated. The most common steps involved in the automation process are feature extraction and classification. Conventional handmade feature-based methods rely on the separation of these two modules, which may result in information loss during the feature extraction process. Moreover, EEG signals often have a low signal-to-noise ratio and high subject variability, for which handmade features may perform sub-optimally.

In the past decade, deep learning approaches have attained higher efficiency in many applications [[6], [7], [8]]. Indeed, features learned by deep networks have often proven to be more robust than handmade features. They allow the extraction of high-level features from the raw data without human intervention, unlike most traditional machine learning algorithms. Deep learning was successfully used in many EEG-related methodologies including affective state classification, diagnosis of various neurodegenerative diseases, motor imaginary task classification, sleep stage detection, drivers fatigue prediction, and epileptic seizure detection [[2], [3], [4],[9], [10], [11]]. Due to its success in many applications, this study proposes a system based on deep learning to identify pathology EEG. In particular, convolutional neural networks (CNN) followed by a support vector machine (SVM)-based classifier is used in the system. The CNN learns and extracts deep-learned features from time-frequency spectrum images. The reason we apply CNN to extract deep-learned features from images is that it has shown excellent results in image classification. Using the most popular models from literature, we could indeed achieve good performance. We investigate different models of CNN in the form of the Inception-ResNet v2, DenseNet, and SeizureNet. A public database, the Temple University Hospital (TUH) Abnormal EEG Corpus database [12], is utilized for the experiment. The contribution of this study is as follows: (i) the accuracy improvement provided by the usage of spectrum images with respect to previous works (see Table 1), (ii) application of time-frequency spectrum images for modelling signals in the context of pathology EEG detection, and (iii) the investigation of the proposed approach performance using different CNN models.

The paper is structured as follows. Section 2 contains the description of related previous works. Section 3 presents the TUH database, data preparation, preprocessing, and the proposed methodology, including the different CNN models. Section 4 discusses the experimental results and interpretations. Further, in Section 4, we view the results in comparison with the literature. Finally, the conclusions drawn from the work and some ideas for future research are given in Section 5.

Section snippets

Related works

In the literature, there are only a limited number of works that have utilized the TUH abnormal EEG corpus dataset for computerized pathology EEG detection. The first classification system for pathology EEG identification was proposed by Ref. [5]. For the system, some pilot experiments using Mel Frequency Cepstral Coefficients (MFCCs) features and the random forest (RF) were conducted to assess the feasibility. The system obtained an error rate of 31.7% for normal and pathology EEG

Scalp EEG dataset

The TUH abnormal EEG Corpus (v2.0.0) database [12,23] was utilized for training and testing in this study. It contains EEG records that are manually identified as either a clinical pathology or normal. The database was partitioned into two subsets, namely train and evaluation, and no patient recordings were present in both sets. Table 2 lists the number of patients and sessions employed in the training and testing set of TUH abnormal EEG corpus. There was a variation in the number of electrodes

Results and DISCUSSION

This section reports the major finding of this work for the detection of pathology EEG using DenseNet-SVM, Inception ResNet v2-SVM and SeizureNet-SVM algorithms. The CNN-based deep learning model was implemented using an open-source Keras library, scripted in Anaconda-Python (v3.5). The experiment was performed on a computer with Intel (R) Core (TM) i7-6700HQ CPU @2.59 GHz and NVIDIA GeForce 1080 Ti graphics-processing unit (GPU) on Ubuntu. Table 3 list the optimized hyper-parameters used in

Conclusion

In this work, we proposed an automated EEG pathology detection system based on different deep CNN models. The spectrogram images were fed into the CNN models to extract deep-learned features. Then, an SVM-based classifier was used for the final detection task. The maximum accuracy value of 96.65% for the T5-O1 channel and 91.30% for the F4–C4 channel is achieved using the SeizureNet-SVM model. The proposed model may assist neurologists in identifying pathology EEG recordings and normal brain

Declaration of competing interest

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Acknowledgments

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

References (33)

  • T. Dharamsi et al.

    Neurology-as-a-service for the developing world

    31st Conference on Neural Information Processing Systems

    (2017)
  • S. Lopez et al.

    Automated identification of abnormal adult EEGs

    IEEE Signal Processing in Medicine and Biological Symposium

    (2015)
  • T. Prasanth et al.

    Deep learning for interictal epileptiform spike detection from scalp EEG frequency sub bands

    42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)

    (2020)
  • S.L. Oh et al.

    A deep learning for Parkinson’s disease diagnosis from EEG signals

    Neural Computing and Applications

    (2018)
  • S.L. Oh et al.

    Deep convolutional neural network model for automated diagnosis of schizophrenia using EEG signals

    Appl. Sci.

    (2019)
  • Y. Roy et al.

    Deep learning-based electroencephalography analysis: a systematic review

    J. Neural. Eng.

    (2019)
  • Cited by (0)

    View full text