An Explainable Multimodal Deep Learning Framework for Personalized Autism Spectrum Disorder Diagnosis

Abstract

Early identification of autism related neurodevelopmental conditions remains difficult due to behavioral heterogeneity and limited clinical transparency. Automated decision support attracts interest because reliable diagnosis supports timely intervention and improved developmental outcomes. The work addresses limitations of unimodal and opaque learning systems by proposing an interpretable multimodal learning design within a single diagnostic pipeline. The scope covers structured behavioral phenotypes and structural brain imaging drawn from a large multi-site public cohort. The proposed method asserts personalized multimodal fusion with built in interpretability improves diagnostic accuracy without sacrificing clinical trust. The methodology integrates explainable additive behavioral modeling neural imaging representation with graph aware encoding latent space fusion and subject adaptive classification. Evidence derives from stratified hold out testing cross validation ablation analysis and explainability assessment using feature attribution and latent visualization. Results report near perfect discrimination with stable precision recall and receiver operating characteristics across evaluation settings. Findings show behavioral explainability and subject adaptive inference contribute strongly to performance gains. The study advances knowledge by unifying interpretability personalization and multimodal learning within one diagnostic design. Practical implications include support for clinician assisted screening in resource limited settings. Future research targets validation across external cohorts longitudinal assessment and inclusion of additional biological modalities.