Automated detection of childhood autism through multimodal fusion remains a cutting-edge and underexplored field. The primary challenges include limited data and the strong heterogeneity among multimodal data. In this paper, we propose a childhood autism spectrum disorder (ASD) detection method based on environmental variables. It dynamically adjusts the weights of each modality using a dominant modality confirmation algorithm guided by environmental variables. The dominant modality serves as a basis for augmenting the other modalities, promoting the complementarity of information through attentional mechanisms, and effectively strengthening intermodal correlations. To analyze children’s behavior, language, and expressive characteristics, we create a multimodal dataset encompassing interaction scenarios involving both children with ASD and typically developing children. Extensive experimental evaluations were conducted on this dataset, and our findings demonstrate that our method effectively adapts to the characteristics of multimodal data in diverse scenarios.