Novel Diversified Echo State Network for Improved Accuracy and Explainability of EEG-based Stroke Prediction

Abstract

Echo State Networks (ESNs) are a powerful machine learning technique that can be used for EEG-based stroke prediction. However, conventional ESNs suffer from two main limitations: they are not always accurate, and they are not always interpretable. This paper presents a novel multi-level framework that addresses these limitations. The framework consists of three main components: optimized feature extraction, ensemble learning, and output refinement for improved interpretability. The optimized feature extraction component uses a novel algorithm to extract features from EEG data that are more relevant to stroke prediction. The ensemble learning component uses a diversified Echo State Networks (D-ESN) to combine the predictions of multiple ESNs, which improves the accuracy of the predictions. The output improvement component uses two Explainability techniques, LIME and ELI5, to gain insight into the decision-making of the D-ESN model. These techniques allow users to see how each feature in the dataset contributed to the model's prediction. The framework was evaluated on a well-known EEG dataset from stroke patients. The experimental results showed that the framework significantly outperformed baseline approaches in terms of both accuracy with 95% and interpretability. These results suggest that the proposed framework has the potential to advance the field of stroke prediction and enable informed decision-making in clinical settings.