Language proficiency assessment of autistic children using large language models

Language impairment is a common comorbidity in children with autism spectrum disorder (ASD), and language proficiency assessment is a primary method for identifying such impairments. However, traditional assessment tools are often subjective and inefficient, while existing computer-assisted methods are limited by a narrow focus and insufficient use of natural language samples. To address these issues, this study proposes a framework for assessing children’s language abilities based on large language models (LLMs). We first preprocess the natural language samples from children and design multiple assessment dimensions and workflows. To enhance the stability of the assessment, we introduce a multi-expert voting mechanism and perform a comparative analysis of various large language models’ performance. The experimental results demonstrate a strong correlation between the framework’s assessment results and the Mullen Scales of Early Learning (MSEL) verbal developmental quotients, with a Pearson correlation coefficient of 0.8 (p < 0.001). Furthermore, the results show that the multi-dimensional evaluation can accurately differentiate between ASD and typically developing (TD) children, achieving a classification accuracy of 0.98. These findings suggest that the proposed framework has significant potential for improving the accuracy of ASD identification.