Enhancing Trust in Brain-Computer Interfaces: Interpretable Deep Learning for EEG Signal Classification

Abstract

Brain signal is one of the methods in healthcare to identify brain activities. Mathematical modeling to machine learning and modern deep learning architectures are used to analyze these signals. These methods have been tested with EEG signals to capture meaningful insights and uncover patterns. Deep learning often outperforms conventional machine learning. But the black box nature of these signals has become a concern, especially in the healthcare sector. Hence, this study focused on explaining the Deep Learning architecture by using methods such as Permutation Feature Importance, Saliency Maps, and Gradient-weighted Class Activation Mapping. To test this, a model for sleep-stage classification using Electroencephalography (EEG) has been developed, as EEG is easier to interpret. The model was created by fusing different types of deep learning architecture. To capture sleep-related frequency activity, Power Spectral Density (PSD) features were extracted. Further, jitter ing was applied. Accuracy and F1-score used to measure the performance of the model. An average accuracy of 0.85 and the weighted F1-score of 0.87 were given on the test set. The results display that combining deep learning with explanation methods and careful info handling can develop EEG-based systems extra transparent and trustworthy.