ASDPred: An End-to-End Autism Screening Framework Using Few-Shot Learning

Abstract

Autism spectrum disorder (ASD) is a neurodevelopmental condi-tion that affects social communication and behavior. Diagnosing ASD as early as possible is desirable because early detection enables timely access to interventions and support. This study aims to develop an innovative and interactive ASD diagnostic tool that incorporates artificial intelligence (AI) technology to empower parents and medical professionals to act on early concerns. Collecting and annotating large-scale ASD data is costly, time-consuming, and labor-intensive. Moreover, significant domain knowledge is required to annotate the collected data. Consequently, there are only a few samples available to train AI models to learn the memorization and generalization from the data. Therefore, we designed a few-shot learning framework that combines a Siamese network with a Wide & Deep network to learn both linear and non-linear relationships from small ASD datasets. The experiment results show that it is effective to apply both Siamese networks and Wide & Deep models to achieve ASD diagnostics using a limited number of samples. Based on the proposed model, we developed a diagnostic tool - ASDPred - embedded within a web-based platform to facilitate ASD diagnosis using the designed model.