Stress Recognition with EEG Signals Using Explainable Neural Networks and a Genetic Algorithm for Feature Selection

Stress is a natural human response to external conditions which have been studied for a long time. Since prolonged periods of stress can cause health deterioration, it is important for researchers to understand and improve its detection. This paper uses neural network techniques to classify whether an individual is stressed, based on signals from an electroencephalogram (EEG), a popular physiological sensor. We also overcome two prominent limitations of neural networks: low interpretability due to the complex nature of architectures, and hindrance to performance due to high data dimensionality. We resolve the first limitation with sensitivity analysis-based rule extraction, while the second limitation is addressed by feature selection via a genetic algorithm. Using summary statistics from the EEG, a simple Artificial Neural Network (ANN) is able to achieve 93.8% accuracy. The rules extracted are able to explain the ANN’s behaviour to a good degree and thus improve interpretability. Adding feature selection with a genetic algorithm improves average accuracy achieved by the ANN to 95.4%.