Evolutionary transfer optimization-based approach for automated ictal pattern recognition using brain signals

• Published in: Frontiers in human neuroscience Vol. 18; p. 1386168
• Main Authors: Swami, Piyush, Maheshwari, Jyoti, Kumar, Mohit, Bhatia, Manvir
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 11.07.2024
• Subjects: electroencephalography; epilepsy diagnosis; evolutionary multi-objective optimization; evolutionary transfer optimization; Human Neuroscience; ictal pattern; non-dominated sorting genetic algorithm; electroencephalography; evolutionary multi-objective optimization; non-dominated sorting genetic algorithm; evolutionary transfer optimization; epilepsy diagnosis; ictal pattern
• ISSN: 1662-5161; 1662-5161
• DOI: 10.3389/fnhum.2024.1386168

The visual scrutinization process for detecting epileptic seizures (ictal patterns) is time-consuming and prone to manual errors, which can have serious consequences, including drug abuse and life-threatening situations. To address these challenges, expert systems for automated detection of ictal patterns have been developed, yet feature engineering remains problematic due to variability within and between subjects. Single-objective optimization approaches yield less reliable results. This study proposes a novel expert system using the non-dominated sorting genetic algorithm (NSGA)-II to detect ictal patterns in brain signals. Employing an evolutionary multi-objective optimization (EMO) approach, the classifier minimizes both the number of features and the error rate simultaneously. Input features include statistical features derived from phase space transformations, singular values, and energy values of time–frequency domain wavelet packet transform coefficients. Through evolutionary transfer optimization (ETO), the optimal feature set is determined from training datasets and passed through a generalized regression neural network (GRNN) model for pattern detection of testing datasets. The results demonstrate high accuracy with minimal computation time (<0.5 s), and EMO reduces the feature set matrix by more than half, suggesting reliability for clinical applications. In conclusion, the proposed model offers promising advancements in automating ictal pattern recognition in EEG data, with potential implications for improving epilepsy diagnosis and treatment. Further research is warranted to validate its performance across diverse datasets and investigate potential limitations.