Deep Learning Based Recurrent Neural Network Model for Stress Detection in EEG Signals

Automatic ElectroEncephalogram EEG classification for Stress detection represents a crucial interest, simultaneously with the increasing deaths caused by depression and psychological effects.Accurate automatic classification of EEG signals represents a complex task, requiring the use of sophisticated algorithms.In this light, we focus through this work on achieving the automatic stress detection from EEG signals, to help clinicians to get the true diagnosis in an early stage.At this light, we opt through this paper to the implementation of a proposed Recurrent Neural Network RNN model for automatic stress detection.The proposed work employs a pre-processing combined with Recurrent Neural Network models such as Gated Recurrent Unit (GRU).We have applied the FFT transformation on EEG signals, available from Kaggle.The EEG classification results have reached 97.23% for train, 93.68% for validation and 88.86% for the test process, by implementing GRU based SGD optimizer networks.To get more accurate results, Adam optimizer has been implemented, achieving results equal to 99.53%, for the train, 94.98% for the validation and 89% for the test process.Moreover, stress emotions have been well detected as demonstrated by the confusion matrix results.Finally, accuracy and loss curves show promising results for both training and validation and the error rate is too close to zero.Our proposed RNN model, with its reduced number of parameters shows to be an excellent application to be implemented on embedded systems, thanks to its lightweight reducing both training time and memory consumption.