Interpretable multi-agent paradigm for ASD detection and social interaction training in B2C transactions for autistic person
Prashant Kumar Gupta, Bireshwar Dass Mazumdar, Rama Komaragiri, Ishika Sharma, Shubhangi Kundu, Akansha Rawat · 2025
The neurological disorders, primarily the Autism spectrum disorder (ASD), significantly hinders a person&s;s developmental abilities and their capacity for social interaction. Lately, sub-symbolic models for detecting ASD have become quite popular. Although these models have exhibited exceptional performance by obtaining nearly perfect accuracy in both training and testing stages, individuals remain doubtful of the recommendations provided by these models and perceive them as ‘Black Box’ systems. Further, wherever these algorithms use unimodal data set in the form of facial images, the ML/DL models classify the subject as autistic (or non-autistic), based on the complete facial image as input. This lacks interpretability as one does not know based on which facial features and how based on those features, does the AI model recommend a subject to belong to a particular class. Furthermore, there hasn&s;t been much research done on using AI techniques to create social interaction training programmes for people with ASD. Therefore, in this article, we propose the architecture and design of an ASD care agent that will teach an ASD subject on how to behave a non-autistic human-like in a social interaction situation and classifies the subjects as autistic (or non-autistic) in an interpretable manner. Furthermore, we use a relevant real-world case study to demonstrate how our agent operates.