Hybrid optimized focal high-order attention network with descriptor computations for autism spectrum disorder detection in federated learning.

Autism spectrum disorder is a neurodevelopmental condition that affects the social interaction, and communication ability of persons. Accurate diagnosis can significantly improve quality of life. However, current detection methods often perform sub optimally due to class imbalance, heterogeneous data, and privacy concerns. For solving such issues, a Federated Learning (FL)-based framework is proposed, integrating a Groupers and Moray Orangutan Optimization Algorithm with a Focal High-order Attention Network (GMOA_Focal-HANet). The autism detection is done in a local model, where autism brain image and autism data are fed to a pre-preparation process. The GMOA performs the functional connectivity-based pivotal region extraction, and descriptor computation is done by Regional Gradient Pattern (RGP). Moreover, the preprepared input data is subjected to attribute screening, where the Chi-Square Test-enabled feature selection is employed. The Focal-HANet detects autism using the outcome of descriptor computation and attribute screening. The GMOA trains the Focal-HANet, and local update and aggregation is done by average method. Experimental results demonstrate that the proposed framework achieves a classification accuracy of 96.83%, with sensitivity and specificity of 95.92% and 96.82%, respectively. These results confirm the effectiveness and robustness of the proposed approach for reliable Autism spectrum disorder detection in a privacy-preserving FL environment.