Investigation on classification of motor imagery signal using Bidirectional LSTM with effect of dropout layers

While classification of neural signals is critical for various applications, practical uses of Electroencephalography (EEG) signals, for example wheelchair control, are sequential in nature. The Long Short Term Memory (LSTM) algorithm is known for its feasibility in analysing and learning from longer ranges of sequential data than recurrent neural networks (RNNs). This work has applied a novel LSTM model design on an EEG dataset which is larger than a reference work. The model in this work has achieved an accuracy of 83% on a dataset of 14 subjects using randomly shuffled tenfold cross validation. Whereas in reference work accuracy of individual subjects were considered. The test accuracy of the data was found to be higher for a training process due to application of dropout layers after LSTM layers. The implementation was realized using Tensorflow version 2.7.0 on a dataset contributed to Physionet which was of BCI2000 standard. Sampling frequency of the EEG signal was 160 Hz and the duration of epoch of the signal is 4 seconds. The implementation was realized using Tensorflow version 2.7.0 on a dataset contributed to Physionet which was of BCI2000 standard.