Elucidating Emotional Patterns in Autism Spectrum Disorder: BERT-Based Analysis Reveals Novel Dimensional Structure

Autism spectrum disorder (ASD) is characterized by difficulties in emotion recognition and regulation, posing challenges for effective treatment support. This study leverages Bidirectional Encoder Representations for Transformers (BERT) and its variants to analyze emotional content of ASD clinical records. To estimate emotion profile, we fine-tuned the models using the WRIME dataset, a robust resource annotated for emotional intensity across eight basic emotions in Japanese. We found that the best performing model, based on tohoku-nlp/bert-large-japanese-v2, achieved high accuracy (78.9%) and average cosine similarity (94.1%) in emotion profile tasks. Text samples from a total of 1,239 days were extracted from the clinical records of 14 adolescents with ASD. Our model generated prediction vectors for the emotion profile of these text samples, which were then analyzed using principal component analysis (PCA). The results revealed dominant emotional patterns, such as sadness and unique contrast between disgust and anticipation, diverging from theoretical frameworks, such as Plutchik's wheel emotions. These findings highlight the potential of machine learning models in uncovering nuanced emotional dynamics in ASD, and offer insights for clinicians and researchers to tailor interventions.