Cognitive Fatigue Detection from EEG Signals using Topological Signal Processing

Topological signal processing has attracted substantial interest lately owing to its attribute of multi-scale tracking of simplicial complexes. This distinctive aspect is exploited to study the topological evolution of time series/signals. Specifically, EEG signals analysis is considered here for the challenging task of cognitive fatigue detection. This work utilizes the topological attributes like Betti numbers, and persistent homology of dimension 0 and 1 extracted from EEG signals to study the cognitive state of an individual. Using the CogBeacon dataset, a comparison of the topological features with the conventional time and frequency domain features is presented. Random forest classifier is used to classify the fatigue state. Results show that the performance of topological features is at par with the conventional features even when significantly less number of topological features are used. Also, enhancement in classification accuracy is observed by appropriately combining both conventional and topological features which outperforms the state-of-the-art method for fatigue detection. Additionally, recursive feature elimination is applied on combined features to reduce redundancy by selecting a subset consisting of prominent features. Analysis indicates that all topological features derived from EEG signals contribute to the best performing subset, which also increases the overall accuracy.