A 1D-CNN prediction model for stroke classification based on EEG signal

Stroke is an acute cerebrovascular disease with high mortality and disability. Computer-aided interventional diagnosis is a necessary measure to improve the efficiency of stroke diagnosis by using modern advanced medical instruments and machine learning methods. Electroencephalogram (EEG) as a diagnostic means, is a test that measures the electrical activity of the brain through electrodes attached to the scalp to find changes in brain activity. EEG detection has the advantages of low cost, simple and easy to implement, and no physical harm and psychological stress to patients. Studies have shown that EEG signal might be useful in diagnosing stroke. By using machine learning methods, EEG signals can be used to classify stroke patients and normal subjects, or subtypes. Stroke is generally divided into two types: ischemic stroke and hemorrhagic stroke. How to classify ischemic and hemorrhagic strokes based on stroke patients' EEG data by constructing prediction model is the main purpose on this paper. In recent years, researchers have developed many technologies in the field of stroke classification prediction based on EEG signals, using a variety of machine learning methods to ensure the improvement of prediction accuracy. The typical methods usually extract the time domain, frequency domain or spatial domain features of EEG signals before establishing a stroke classification model. However, the quality of the extracted features cannot be guaranteed in stroke patient or subtype classification. In addition, EEG feature extraction is usually computationally expensive. The main goal of this paper is to propose a novel classification prediction model using an end-to-end deep neural network that avoids the process of manual feature extraction. This paper proposes a one-dimensional convolutional neural network (1D-CNN) classification model based on stroke EEG signal. The model includes four convolutional blocks, a global average pooling layer, a dropout layer, and a SoftMax layer. Each convolution block consists of two convolution layers and a pool layer for extracting features and reducing the number of parameters. A one-dimensional convolution kernel is used in order to match the characteristics of EEG one-dimensional time domain signal. The model can automatically extract the features of stroke EEG signal for classifying stroke by using convolutional layers. The EEG data of clinical stroke patients collected from the neurology department of a hospital are used in the experiments. Long Short-Term Memory (LSTM) model is also used as a benchmark to achieve end-to-end prediction for verifying the proposed model performance. The experimental results show that the proposed 1D-CNN prediction model has good prediction performance, with an accuracy of 90.53%, a precision of 87.90%, a sensitivity of 91.60%, and a specificity of 89.65%. It is much higher than the prediction result of LSTM model.