TurkerPat: EEG-Based detection of hunger, thirst, and nicotine withdrawal

Abstract

Background: One of the primary objectives of neuroscience is to gather information from the brain. Therefore, brain data are crucial for understanding its secrets, and one of the most affordable methods for collecting such data is electroencephalography (EEG). To capture meaningful information, machine learning models have been applied to EEG signals. In this research, our main goal is to investigate an innovative feature-extraction method on a new EEG dataset to obtain both accurate classification and interpretable results. Material and Methods: First, we curated a novel EEG signal dataset comprising four classes: (i) hungry, (ii) thirsty, (iii) cigarette-addicted, and (iv) control. Using this dataset, we defined four cases: (1) hunger detection, (2) thirst detection, (3) nicotine-withdrawal detection, and (4) abnormality (hunger + thirst + nicotine-withdrawal) detection. To automatically detect these cases, we introduced a specialized transformer-based feature-extraction method. This transformer, called the Moon Star Transformer (MST), was deployed alongside a Transition Table Feature Extractor (TTFE) to form the Turker Pattern (TurkerPat). Feature selection, ensemble and iterative classification, and an interpretable results generator were then integrated into the TurkerPat-centric XFE framework to achieve both classification accuracy and interpretability. Results: The proposed TurkerPat-centric XFE framework attained over 85 % classification accuracy using leave-one-subject-out cross-validation (LOSO CV). By applying Directed Lobish (DLob) for interpretable result generation, we obtained connectome diagrams for each defined case. Conclusion: The classification and explainable results clearly demonstrate that the TurkerPat-centric XFE framework makes a significant contribution to both neuroscience and feature engineering.