Metaheuristic hyperparameter optimization of deep neural networks for demographic-aware autism spectrum disorder
Mohammed Aly1,2, Naif M Alotaibi3
1Department of Artificial Intelligence, Faculty of Artificial Intelligence, Egyptian Russian University, Badr City, 11829, Egypt. mohammed-alysalem@eru.edu.eg.
Scientific Reports
|June 29, 2026
Summary
This study introduces a novel deep learning framework for Autism Spectrum Disorder (ASD) classification using structural MRI data, optimizing models for demographic subgroups like age and gender. The approach enhances diagnostic accuracy by leveraging metaheuristic optimization for improved neuroimaging analysis.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Computational Neuroscience
Background:
- Autism Spectrum Disorder (ASD) classification from neuroimaging data is challenging due to data heterogeneity and demographic variability.
- Existing deep learning methods often use binary classification and lack systematic optimization, limiting their robustness across diverse populations.
Purpose of the Study:
- To develop a demographic-aware ASD classification framework using structural MRI (sMRI) data.
- To address challenges in ASD classification by formulating it as an optimization-driven learning problem.
- To automatically optimize model architecture and hyperparameters for task-specific adaptation.
Main Methods:
- Proposed a deep learning framework with customized Convolutional Neural Network (CNN) models for gender, age, and joint age-gender classification.
- Employed the Optimized Artificial Bee Colony (OptABC) algorithm for automatic model architecture and hyperparameter optimization.
- Utilized a dedicated preprocessing pipeline with structural localization and data augmentation on the multi-site ABIDE dataset, evaluated via five-fold cross-validation.
Main Results:
- Achieved accuracies of 84.25% (gender), 88.07% (age), and 71.58% (joint age-gender) using the optimization-driven framework.
- Demonstrated competitive performance compared to pre-trained transfer learning models.
- Found age-aware modeling to be more discriminative than gender-based classification, with joint stratification increasing complexity.
Conclusions:
- Metaheuristic optimization effectively enhances deep learning models for complex, demographic-aware neuroimaging classification tasks.
- The proposed framework offers a robust approach to ASD classification, accounting for demographic variability.
- Future work will involve evaluating the framework's robustness on independent datasets.