Predicting autism from written narratives using deep neural networks.

Despite the heterogeneity of language and communication abilities within the autistic population, challenges associated with the pragmatic (social) use of speech remain consistently observable across the entire spectrum of autism. Therefore, the study of narrative competence is particularly relevant, and there has been a considerable rise in research on narrative skills in autism. Most studies have focused on spoken narratives, with some describing the potential use of automated computational methods. In this study, we analyzed written narratives collected in a standardized manner during a national exam. We gathered 363 essays from students in the final eighth grade of primary school: 193 from autistic students (ASD Group) and 168 non-autistic peers (non-ASD Group). We tested several deep neural models to predict whether an essay was written by an autistic student or a student from the non-ASD Group. Several models achieved promising results, exceeding values of 0.85 for sensitivity, specificity, and accuracy coefficients. In addition to studying narrative competence, the data from national exams and their utility in distinguishing autistic individuals may potentially pave the way for large-scale and cost-effective epidemiological studies on autism in the future.