A Deep Learning Approach to the Classification of EEG Normality

The relatively low cost and noninvasive nature of EEGs have made them a popular tool in the identification of disease over the past few decades [1]. Traditional interpretation of EEG signals for clinical purposes have been largely dependent on the input of trained professionals, which often results in extensive lag time between the duration in which an EEG is taken and the production of an EEG report [2]. Additionally, said interpretation can be highly subjective, increasing the possibility of a misdiagnosis that could potentially be life- threatening in some scenarios.

In the past decade, complex machine learning algorithms have had numerous impacts on the field of medicine, as models with capabilities to diagnose and provide prognosis with incredible accuracy have emerged [1]. The recent creation of large clinical databases with patient readings and information offers extensive data which could be utilized for the development of deep learning models.
In the general procedure for a scalp EEG, the workflow is as follows: electrode placement, data acquisition, EEG analysis, and report generation (Fig. 1). The EEG report consists of important information such as the demographics and clinical history of the patient and a description of observed patterns [3].
One particular portion of the report classifies the reading as “normal or abnormal” based on the professional’s observation of the EEG activity during the session. Our work centers around this aspect of the report – specifically, we aim to use novel deep-learning techniques to classify EEG signals as abnormal or normal.

While other approaches to normality classification of EEG signal data typically center around non-deep or non-time series models [1, 2, 4], our model utilizes a time-series LSTM RNN. To our knowledge, an LSTM RNN approach has never been applied in the context of EEG normality classification [5]. We compare the performance of our model to various non-deep model, such as k-Nearest Neighbor, Random Forest, and Logistic Regression classifiers.