Maher Institute Inc Iranian Journal of Neurodevelopmental Disorders 3115-848X In Press 2026 11 01 Optimal Feature Selection and CNN Hyperparameter Tuning Using the Grey Wolf Optimizer for Depression Detection Optimal Feature Selection and CNN Hyperparameter Tuning Using the Grey Wolf Optimizer for Depression Detection 1 16 EN Iman Ahanian Department of Electrical Engineering, ST.C., Islamic Azad University, Tehran, Iran Journal Article 2026 04 01 Purpose: This study aimed to develop an optimized electroencephalography-based depression-detection framework by integrating the Grey Wolf Optimizer with a convolutional neural network for simultaneous feature selection and hyperparameter tuning. Methods and Materials: An applied analytical design was used to analyze a standard EEG-derived dataset obtained from the Figshare repository, containing observations from healthy individuals and patients with major depressive disorder. Data preprocessing included the removal of zero-variance variables, moving-average imputation of missing values, Z-score standardization, and class balancing using the Synthetic Minority Over-sampling Technique. A binary Grey Wolf Optimizer was employed to select an informative subset of EEG-derived temporal, frequency-domain, and functional-connectivity features. Candidate feature subsets were evaluated using a radial basis function support vector machine with five-fold cross-validation. The selected features were converted into two-dimensional feature maps and entered into a custom CNN. GWO was additionally used to optimize the learning rate, batch size, convolutional-filter numbers, dropout rate, training epochs, and feature-selection threshold. Findings: The GWO-based optimization process identified a refined feature subset that preserved the majority of diagnostically relevant EEG information while removing variables that contributed limited discriminatory value. The optimized CNN comprised three convolutional layers with progressively increasing filter numbers of 32, 64, and 128. The best-performing configuration included a learning rate of 0.0001, a batch size of 32, a dropout rate of 0.50, 100 training epochs, and a feature-selection threshold of 0.50. The combined optimization procedure indicated that simultaneous feature refinement and hyperparameter tuning produced a more stable and computationally coherent framework than reliance on unoptimized feature inputs and manually selected network parameters. Conclusion: Integrating GWO-based feature selection and CNN hyperparameter optimization provides a systematic and promising approach for EEG-based depression detection by improving input quality, reducing unnecessary computational complexity, and supporting the extraction of complex neurophysiological patterns associated with major depressive disorder.

https://www.maherpub.com/index.php/jndd/article/download/935/716