Metaheuristic hyperparameter optimization of deep neural networks for demographic-aware autism spectrum disorder classification.
Authors
Affiliations (3)
Affiliations (3)
- Department of Artificial Intelligence, Faculty of Artificial Intelligence, Egyptian Russian University, Badr City, 11829, Egypt. [email protected].
- Department of Software Engineering, Faculty of Artificial Intelligence, Egyptian Russian University, Badr City, 11829, Egypt. [email protected].
- Department of Computer Science, College of Science and Humanities Dawadmi, Shaqra University, Shaqra, 11961, Saudi Arabia.
Abstract
Autism Spectrum Disorder (ASD) classification from neuroimaging data poses significant challenges due to high data heterogeneity and the complexity of learning robust representations across demographic subgroups. While recent deep learning approaches have shown promise, most existing studies rely on binary classification and limited model tuning strategies, overlooking demographic variability and the role of systematic optimization in model design. To address these challenges, this study formulates demographic-aware ASD classification as an optimization-driven learning problem and proposes a deep learning framework based on structural MRI (sMRI) data. The proposed framework employs three customized Convolutional Neural Network (CNN) models targeting gender-based classification, age-group-based classification, and joint age-gender classification using an octal class structure. Model architecture and training hyperparameters are automatically optimized using the Optimized Artificial Bee Colony (OptABC) algorithm, enabling task-specific adaptation without manual tuning. A dedicated preprocessing pipeline incorporating structural localization and controlled data augmentation is applied to improve robustness across heterogeneous imaging sites. All models are evaluated using five-fold cross-validation on the multi-site ABIDE dataset. Experimental results demonstrate that the proposed optimization-driven framework achieves accuracies of 84.25%, 88.07%, and 71.58% for gender-based, age-based, and joint age-gender classification tasks, respectively, yielding competitive performance relative to widely used pre-trained transfer learning models under comparable experimental settings. The results further indicate that age-aware modeling offers stronger discriminative capability than gender-based classification, while joint demographic stratification introduces increased task complexity due to finer class granularity. Overall, the findings highlight the effectiveness of metaheuristic optimization for enhancing deep learning models in complex, demographic-aware neuroimaging classification tasks. Further evaluation on independent datasets is planned to assess robustness across broader settings.