A new fuzzy-based ensemble framework based on attention-based deep learning architectures for automated detection of abnormal EEG

Biomedical science research encompasses a wide array of fields such as biomedical engineering, gene analysis, biomedical signal and image processing. The significance of detection, classification, and recognition is paramount for disease diagnosis and analysis within this realm. This study specifically delves into the topic of biomedical EEG classification. For this purpose, three different attention-based deep classifier as CNN+Att, LSTM+Att and InstaGATs are considered, and meta data is produced. Moreover, a fuzzy min–max classifier is employed for the purpose of detection to mitigate uncertainty. The results from the individual base models, including probability of matches and their associated classes, are combined and inputted into fuzzy min–max model for the ultimate phase. The fuzzy model receives inputs from the base classifiers in the form of class probabilities and corresponding labels. Within the fuzzy model, the min–max algorithm is applied to ensure accurate decisions. Performance measures including recall, precision, accuracy, and F1-score are then calculated and assessed. The proposed ensemble framework was applied to two different EEG databases: TUH EEG Abnormal and TUH EEG Seizure. Our proposed approach achieved 84.67% accuracy, 85.20% precision, 83.92% recall and 84.55% F1-score for TUH EEG Abnormal database and 99.51% accuracy, 99.72% precision, 99.24% recall and 99.47% F1-score for TUH EEG Seizure database. By demonstrating enhanced performance relative to base classifiers incorporating attention mechanism, the suggested combined network validates the concept of ensemble learning for automated detection of abnormal EEGs. Our findings demonstrate that we have attained leading-edge performance in all classification tasks, despite the substantial dataset variations and the simplistic design of the attention-enhanced models put forth.