Towards Explainable Multi-Modal Fusion Strategies for ASD Detection: A Review

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder marked by complexities that has attracted considerable attention recently, leading to extensive research on automated diagnosis through artificial intelligence. As largescale neuroimaging data becomes more accessible datasets, researchers have begun to build systems incorporating multi-model fusion strategies and Explainable Artificial Intelligence (XAI) techniques to enhance classification performance and model interpretability. This review presents a comprehensive overview with extensive analysis of recent advancements in multi-model fusion approaches for ASD detection including early (featurelevel), intermediate (model-level), and late (decision-level) fusion techniques applied to structural MRI, functional MRI, and phenotypic data, focusing on studies utilizing the ABIDE dataset. In parallel, the growing importance of XAI is addressed, highlighting methods such as “GMDH”, “SHAP”, and “LIME” that provide transparency and insight into model decisions, enabling clinicians to understand these decisions. The paper discusses common challenges in multi-site datasets, model generalizability, and interpretability, and offers guidance for selecting appropriate fusion strategies and explanation tools. Finally, future directions are proposed to improve the clinical relevance, robustness, and ethical deployment of AI systems for ASD diagnosis.