Kato Sumi, Hanawa Kazuaki
Department of Neuropsychiatry, Graduate School of Medicine, Hirosaki University, Hirosaki, Japan.
Faculty of Management and Law, Aomori Chuo Gakuin University, Aomori, Japan.
Front Hum Neurosci. 2025 Aug 1;19:1606701. doi: 10.3389/fnhum.2025.1606701. eCollection 2025.
This study examines whether specific lexicogrammatical features can reliably differentiate individuals with autism spectrum disorder (ASD) from non-ASD individuals. Classification models using logistic regression and deep neural networks (DNN) demonstrated high performance-80% accuracy, 82% precision, 73% sensitivity, and 87% specificity. To clarify which linguistic variables contribute to this differentiation, the analysis focused on identifying key syntactic features associated with ASD-specific patterns of lexicogrammatical choices.
This study used the Tag Linear Model, developed in prior work, which enables identification of specific lexicogrammatical discriminators. Although DNN models achieved higher predictive accuracy, their internal processes were not interpretable. To identify statistically significant features, we applied a logistic regression with 10,000 bootstrap iterations; -values derived from this procedure indicated the statistical significance of each feature. The linear model thus provided transparent evidence of differences in lexicogrammatical features between ASD and non-ASD individuals.
Of the 135 lexicogrammatical items analyzed, 46 were identified asstatistically significant discriminators ( < 0.05) between ASD and non-ASD speakers. From these 46 discriminators, 20 showing variation at the clause and phrase level were selected for detailed analysis. These were grouped into seven cognitive-functional domains implicated in ASD, including working memory, inferencing, joint attention, and mental space construction.
These findings suggest that syntactic variation in ASD reflects underlying domain-specific cognitive constraints. Linking lexicogrammatical features to cognitive-functional domains provides a linguistically grounded perspective on the neurocognitive profiles of ASD and informs future diagnostic and intervention approaches.
本研究探讨特定的词汇语法特征能否可靠地区分自闭症谱系障碍（ASD）个体与非ASD个体。使用逻辑回归和深度神经网络（DNN）的分类模型表现出了较高的性能——准确率达80%、精确率达82%、灵敏度达73%、特异性达87%。为了阐明哪些语言变量促成了这种区分，分析聚焦于识别与ASD特定词汇语法选择模式相关的关键句法特征。
本研究使用了先前工作中开发的标签线性模型，该模型能够识别特定的词汇语法区分因素。尽管DNN模型实现了更高的预测准确率，但其内部过程无法解释。为了识别具有统计学意义的特征，我们应用了带有10000次自助法迭代的逻辑回归；从该过程得出的p值表明了每个特征的统计学意义。因此，线性模型为ASD个体与非ASD个体在词汇语法特征上的差异提供了透明的证据。
在分析的135个词汇语法项目中，有46个被确定为ASD和非ASD说话者之间具有统计学意义的区分因素（p<0.05）。从这46个区分因素中，选择了20个在从句和短语层面表现出差异的因素进行详细分析。这些因素被归为与ASD相关的七个认知功能领域，包括工作记忆、推理、联合注意力和心理空间构建。
这些发现表明，ASD中的句法变异反映了潜在的特定领域认知限制。将词汇语法特征与认知功能领域联系起来，为ASD的神经认知概况提供了一个基于语言的视角，并为未来的诊断和干预方法提供了信息。
