MDNCT: a multi-domain neurocognitive transformer architecture approach for early prediction of autism spectrum disorders

Abstract

Intellectual disability (ID) refers to a disorder involving intelligence and adaptive behavior that meets specific criteria involving deviance from the norm in terms of degree. ID is more common in males than females, and the causes can be genetic or environmental. This population has historically been characterized by significantly decreased life expectancy because they have not been diagnosed and treated for such diseases as cardiovascular and respiratory ones. However, medical progress in the last few years has slightly narrowed this gap, highlighting that understanding ID requires its consideration as a comorbidity to neurodevelopmental and cognitive diseases like Autism Spectrum Disorder (ASD), dementia, or learning disability. Thus, this work proposes the multi-domain NeuroCognitive Transformer (MDNCT) suitable for different prediction tasks on different datasets. Therefore, MDNCT obtains high performance based on the adequate preprocessing level according to the domain data's specific characteristics, more advanced feature extraction methods, and the use of Transformer-based neural networks. The structure of the framework incorporates common means to align multiple features across modalities and also other state-of-the-art features like multi-head self-attention and residual connections for learning. The use of the MDNCT includes important domains, including early dementia diagnostics for health purposes, social media comments toward learning disabilities, and effective identification of ASD in toddlers.

Keywords: Artificial intelligence; Autism spectrum disorder (ASD); Multi-domain analysis; NeuroCognitive transformer architecture; Neurodevelopmental disorders; Self-attention mechanism.

Fig. 1

Workflow of proposed framework of MDNCT.

Algorithm 1

Workflow of MDNCT framework.