Autism Spectrum Disorder Detection and Classification Using an Advanced Multimodal Federated Transformer Framework

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition marked by persistent deficits in social communication and restricted, repetitive patterns of behavior. Early and accurate diagnosis remains challenging due to the heterogeneity of symptoms and reliance on subjective behavioral assessments. This paper proposes a novel Multimodal Federated Transformer Network (MFTN) for automated ASD detection and classification by jointly learning from functional MRI (fMRI), structural MRI (sMRI), and eye‑tracking data. Unlike prior review‑oriented or unimodal studies, this work presents a complete end‑to‑end journal‑ready research contribution, including model design, experimental evaluation, and comparative analysis. The proposed framework employs modality‑specific encoders, a cross‑modal transformer fusion layer, and federated learning to address data privacy and inter‑site variability. Experiments conducted on the ABIDE I dataset augmented with a public eye‑tracking cohort demonstrate that the proposed model achieves 91.8% accuracy, 0.94 AUC, and balanced sensitivity–specificity, outperforming state‑of‑the‑art CNN and hybrid models. Attention‑based explainability further highlights clinically relevant brain regions and gaze patterns, supporting the model’s translational potential.