An Adaptive Transfer Learning Framework for Multimodal Autism Spectrum Disorder Diagnosis.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition with diverse behavioral, genetic, and structural characteristics. Due to its heterogeneous nature, early diagnosis of ASD is challenging, and conventional unimodal approaches often fail to capture cross-modal dependencies. To address this, this study introduces an adaptive multimodal fusion framework that integrates behavioral, genetic, and structural MRI (sMRI) data, addressing the limitations of unimodal approaches. Each modality undergoes a dedicated preprocessing and feature optimization phase. For behavioral data, an ensemble of classifiers using a stacking technique and attention mechanism is applied for feature extraction, achieving an accuracy of 95.5%. The genetic data is analyzed using Gradient Boosting, which attained a classification accuracy of 86.6%. For the sMRI data, a Hybrid Convolutional Neural Network-Graph Neural Network (Hybrid-CNN-GNN) architecture is proposed, demonstrating a strong performance with an accuracy of 96.32%, surpassing existing methods. To unify these modalities, fused using an adaptive late fusion strategy implemented with a Multilayer Perceptron (MLP), where adaptive weighting adjusts each modality's contribution based on validation performance. The integrated framework addresses the limitations of unimodal approaches by creating a unified diagnostic model. The transfer learning framework achieves superior diagnostic accuracy (98.7%) compared to unimodal baselines, demonstrating strong generalization across heterogeneous datasets and offering a promising step toward reliable, multimodal ASD diagnosis.