Autism as an alteration of interpersonal predictions.

• Applicant: Dr. Tommaso Tosato, Ph.D.
• Subject Area: Human Cognitive and Systems Neuroscience
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 542430763

Our research seeks to deepen the understanding of social interaction challenges in individuals with Autism Spectrum Disorder (ASD). We aim to explore how individuals, both with and without ASD, anticipate their conversation partners' words and sentences, shifting the focus from isolated individuals to the critical role of real-time interactions. We will analyze both existing and newly acquired datasets featuring simultaneous brain activity recordings from pairs of participants engaged in natural conversation. Notably, our new dataset will include conversations between autistic children and their caregivers, which adds to the datasets currently available in our lab. Our goal is to use existing artificial intelligence (AI) tools and develop new ones to analyze these complex datasets. In particular we want to link the EEG data to conversational behaviors and quantify the level of interpersonal prediction. Conversations will be transcribed into text, followed by the use of conventional alignment metrics to quantify mutual understanding. Additionally, we will employ Large Language Models to transform the spoken words and phrases into numerical vectors - known as embeddings. These embeddings represent the meaning of words in the specific context in which they occur. The project aims to develop two predictive models based on neural networks. The first model will forecast the word embeddings from one participant's speech, based on the other's brain activity captured through EEG. This will be assessed before and after the words are heard. Successful prediction of word embeddings from brain activity occurring before a word is heard, indicates predictive abilities. The second model will attempt to predict the neurophysiological responses of one participant using the other's EEG data, which will be taken as an indication of having a model of the other person. We hypothesize that ASD may alter these interpersonal predictions, which are fundamental to social interaction. We will evaluate this by comparing the performance of our models when trained with data from the ASD group versus a control group. The software and methodologies developed will be released open-source, contributing valuable resources to the scientific community.

• DFG Programme: WBP Fellowship
• International Connection: Canada
• Host: Dr. Guillaume Dumas, Ph.D.