Classification of cross task cognitive workload using deep recurrent network with modelling of temporal dynamics

Abstract

Cognitive workload assessment is important for optimizing multitasking abilities in the current information-rich world. A few good research studies assess cross task classification where a model trained for one task can efficiently handle other tasks. Building a reliable model for the estimation of cross task cognitive workload using electroencephalogram (EEG) signal is challenging as brain dynamics differ for various activities. The present study explores the wide visual aspects for inducing two levels of cognitive workload in the two self-design tasks. The present study addresses temporal dynamics of EEG signal using time windowing, time segment smoothing, and formation of variable duration time frames. The present study computes features from the statistical, morphological and non-linear domain for variable duration segments considering four clinical sub-bands. The neighbourhood component feature selection algorithm identifies the topmost common features of both tasks. Binary classification of cross task utilizes deep structure employing bidirectional long short-term memory (BLSTM) and LSTM. The result shows that the deep recurrent neural network (RNN) achieves the highest classification accuracy of 92.8% for the cross task in comparison to the traditional classifiers. The deep RNN also outperforms the existing literature studies for cross task classification. Result also reveals a significant improvement in classification accuracy with the usage of handcrafted features when supported by proper modelling of temporal dynamics of EEG signal. The present study exploits the capability of deep recurrent structure in handling long term dependencies of non-stationary signal and proves to be very efficient.