Predicting autism from written narratives using deep neural networks
Izabela Chojnicka1, Aleksander Wawer2
1Faculty of Psychology, University of Warsaw, 00-183, Warsaw, Poland. izabela.chojnicka@psych.uw.edu.pl.
Scientific Reports
|July 2, 2025
Summary
Researchers used deep neural models to analyze written narratives from autistic and non-autistic students. The models accurately distinguished between the groups, highlighting potential for large-scale autism studies.
Area of Science:
- Neuroscience
- Developmental Psychology
- Computational Linguistics
Background:
- Autism Spectrum Disorder (ASD) presents diverse language challenges, particularly in pragmatic (social) communication.
- Narrative competence is a key area of study in autism research, with increasing interest in computational analysis.
- Previous research often focused on spoken narratives, with limited exploration of written forms in standardized assessments.
Purpose of the Study:
- To investigate the utility of written narratives from a national exam for distinguishing autistic students.
- To evaluate the effectiveness of deep neural network models in analyzing narrative competence in autism.
- To explore the potential of using standardized written data for future large-scale epidemiological studies on autism.
Main Methods:
- Analysis of 363 written essays from eighth-grade students (193 autistic, 168 non-autistic) collected during a national examination.
- Application and testing of several deep neural network models to classify essays based on authorship (autistic vs. non-autistic).
- Evaluation of model performance using sensitivity, specificity, and accuracy metrics.
Main Results:
- Deep neural models demonstrated high performance in distinguishing between essays written by autistic and non-autistic students.
- Several models achieved promising results, with coefficients for sensitivity, specificity, and accuracy exceeding 0.85.
- The findings indicate that written narratives contain discernible markers of autism.
Conclusions:
- Written narratives from standardized assessments can be effectively analyzed using computational methods to identify characteristics associated with autism.
- The study demonstrates the potential of leveraging national exam data for large-scale, cost-effective autism research and screening.
- Further research in this area could significantly advance epidemiological studies and early identification of autism spectrum disorder.
