Creating a diagnostic assessment model for autism spectrum disorder by differentiating lexicogrammatical choices through machine learning

Abstract

This study explores the challenge of differentiating autism spectrum (AS) from non-AS conditions in adolescents and adults, particularly considering the heterogeneity of AS and the limitations ofssss diagnostic tools like the ADOS-2. In response, we advocate a multidimensional approach and highlight lexicogrammatical analysis as a key component to improve diagnostic accuracy. From a corpus of spoken language we developed, interviews and story-recounting texts were extracted for 64 individuals diagnosed with AS and 71 non-AS individuals, all aged 14 and above. Utilizing machine learning techniques, we analyzed the lexicogrammatical choices in both interviews and story-recounting tasks. Our approach led to the formulation of two diagnostic models: the first based on annotated linguistic tags, and the second combining these tags with textual analysis. The combined model demonstrated high diagnostic effectiveness, achieving an accuracy of 80%, precision of 82%, sensitivity of 73%, and specificity of 87%. Notably, our analysis revealed that interview-based texts were more diagnostically effective than story-recounting texts. This underscores the altered social language use in individuals with AS, a crucial aspect in distinguishing AS from non-AS conditions. Our findings demonstrate that lexicogrammatical analysis is a promising addition to traditional AS diagnostic methods. This approach suggests the possibility of using natural language processing to detect distinctive linguistic patterns in AS, aiming to enhance diagnostic accuracy for differentiating AS from non-AS in adolescents and adults.

Fig 1. Position of cognition in language activities defined by SFL.

(Adapted from Kato et al. [11]) This figure illustrates the SFL hierarchy enhanced by Kato’s cognition layer addition [11]. It demonstrates how culture and situation shape lexicogrammatical choices via Field, Tenor, and Mode, thereby impacting communication. The diagram underscores the importance of cognition in selecting suitable lexicogrammar for successful social interactions.

Fig 2. Distribution of AS and co-occurring conditions.

Fig 3. Indicative sentence type of the system network.

This figure presents a close-up of the segment highlighted by the red circle in the MOOD system (S1 Fig 1 in S1 File for context). It illustrates the progression of delicacy in mood selection choices, moving from left to right across the network.

Fig 4. Procedure for automatic AS differentiation experiments using leave-one-out cross validation (LOOCV).

Fig 5. Bi-LSTM-based classification model using tags.

The term Middle signifies a type of verb that lacks agency from a perspective, while Usuality indicates how often an event tends to occur. Additional explanation is provided in Table 2. Both these selective resources are embedded in the system network.

Fig 6. Bi-LSTM-based classification model that utilizes tag-and-text.

An example sentence is Oko ttari nanka surukoto aru (There are occasions when I get mad). The assigned tags are the same as in Fig 5.