Encoding Rich Frequencies for Classification of Stroke Patients EEG Signals

Abstract

The stroke, which is a sudden cut in the blood supply in the brain, has become a severe phenomenon. It has affected around 15 million people annually worldwide. Methods of stroke discovery and monitoring the patient's recovery are a long process, ranging from the analysis of medical images to frequent reporting of the patients for progress assessment. In this paper, we aim to process stroke patient EEG signals by a deep learning approach, and classify a given EEG signal into stroke/non-stroke. In particular, our model consists of several sub-modules which convert and re-model widely used signal processing techniques such as the Fast Fourier Transform (FFT), Convolution in the Frequency Domain and the Inverse Fast Fourier Transform (IFFT) to learnable and differentiable functions that are completely learned and optimized in an end-to-end manner by neural networks. We demonstrate that our model outperforms several baselines by learning rich frequency features through our proposed model. The proposed model could potentially assist a medical doctor in analyzing stroke brain images with high accuracy rates. It can also be useful for rehabilitation centers to monitor the progress of stroke patients.