ATTEMPTED MOVEMENT CLASSIFICATION OF SPINAL CORD INJURED PATIENTS COMBINING CNN AND LSTM NETWORK

Abstract

Electro encephalography (EEG) can b e used to classify attempted hand movements of Spinal Cord Injured (SCI) patients for improving their quality of life. EEG classifiation with Brain-Computer Interface (BCI) allows individuals who are suffring from the most severe motor disabilities to control and direct electromechanical devices. Practical applications of BCI system require improved classifiation p erformance for attempted hand movements of SCI patients. This research aims to develop a hybrid CNN-LSTM architecture for multichannel EEG signal classifiation, optimize its hyp erparameters, and validate p erformance metrics for improved classifiation p erformance. EEG data acquired from SCI patients go through fitration, downsampling, and artifact removal, followed by the generation of Time-Frequency Representation (TFR) of EEG data. Spatial enco ding is done by arranging TFR data in a 2D array corresp onding to spatial layout of electro des. Spatial enco ded TFR data is then fed to CNN-LSTM (Convolutional Neural Network – Long Short Term Memory) N etwork to obtain the fial classifiation output. Sp ectral, spatial, and temp oral information is vital in EEG classifiation. Novelty in the design of Hybrid CNN-LSTM network architecture is that it can learn to extract sp ectral, spatial, and temp oral information and then use these learned information to improve fial classifiation p erformance. Hybrid CNN-LSTM architecture achieved a classifiation accuracy of 92.36% using 10% of the dataset for training and 90% of the dataset for testing. This result shows 47.363% increased classifiation accuracy as compared to related study while also having improved generalizability. H ybrid CNN-LSTM network for EEG classifiation is able: to deal with artifacts in EEG data without signifiant loss in classifiation p erformance; to extract sp ectral information, spatial information, and temp oral information of valuable neural impulses from EEG data. The steps of EEG classifiation used in this research can b e used not only for attempted movements of SCI patients but also for other neurological diseases, neuroscience applications, mental workload, neuromarketing, and biometrics.