Dimensional Structure of BERT-Estimated Emotion Representations in Adolescents With Autism Spectrum Disorder: An Exploratory Analysis of Clinical Narratives

Background: Autism spectrum disorder (ASD) is characterized by impairments in emotion recognition and regulation. While natural language processing (NLP) offers a systematic approach to estimating emotions from clinical narratives, the latent dimensional structure of these estimated emotions in ASD remains insufficiently characterized using methodologically rigorous frameworks. Objective: This exploratory study aimed to characterize the dimensional structure of model-estimated emotions in adolescents with ASD using a fine-tuned Japanese BERT model, while explicitly accounting for domain shift and compositional data constraints. Methods: A Japanese BERT model (tohoku-nlp/bert-large-japanese-v2) was fine-tuned on the WRIME dataset (social media posts). The model was applied to 1,239 clinical sessions from 14 adolescents with ASD to generate eight-dimensional emotion probability vectors. To address the compositional nature of softmax outputs and the non-independence of repeated sessions, we performed a centered log-ratio (CLR) transformation followed by within-patient centering. Principal component analysis (PCA), with patient-level bootstrapping, was conducted to identify robust dimensions. In-domain validation was performed using 150 manually annotated clinical snippets. Results: The model achieved 81.3% accuracy on the out-of-domain test set, though in-domain validation revealed performance variability across emotion categories (e.g., higher reliability for sadness than for trust). PCA identified two primary dimensions: PC1 (57.8% variance), characterized by a dominant sadness-driven internalizing axis, and PC2 (14.2% variance), representing a contrast between anticipation and aversive emotions (disgust/anger). These dimensions remained stable across bootstrap resampling. Conclusions: This study demonstrates that transformer-based NLP, when combined with rigorous compositional data analysis, can elucidate latent emotional structures within clinical narratives. However, these patterns reflect a model-mediated representation space influenced by clinical documentation practices and domain-specific model characteristics. While providing a novel quantitative framework for psychiatric NLP, our findings emphasize the necessity of in-domain validation and cautious clinical interpretation.