DNA methylation in bipolar disorder: mechanistic insights and translational challenges.

Autism Spectrum Disorder (ASD) is a complicated neurodevelopmental disorder that is marked by the enduring difficulties in social communication, in addition to limited and repetitive behavior patterns. According to the recent epidemiological research, the prevalence of ASD in the world is steadily growing, which demonstrates the necessity of the timely and reliable screening system to facilitate the timely clinical intervention and better long-term developmental outcomes of the individuals (). Although clinical practice has been advanced, the traditional diagnostic pathways are still resource intensive and, in most cases, they may take a long period of time to complete through observational assessment by expert clinicians. These limitations translate to late diagnoses, higher healthcare expenses and low accessibility, especially where limited resources and non-urban areas are involved. The existing screening and diagnostic procedures of ASD are largely based on standardized behavioral checklists and clinician-administered tests like Modified Checklist for Autism in Toddlers (M-CHAT) and Autism Diagnostic Observation Schedule (ADOS) (). Although these instruments have been shown to be clinically valid, they are subjective in nature and largely rely on the ability of clinicians, caregivers to report accurately, and to use language clearly. As a result, they can be less effective in multilingual or culturally heterogeneous population groups, where the language barrier and interpretational bias can affect the results of screening. To address these shortcomings, more recent studies have investigated the use of machine learning and deep learning methods to screen ASD by utilizing data. Previous researchers have used neuroimaging technologies, including fMRI and MRI, and neurophysiological measures, such as EEG, to find an ASD-related biomarker (, ). These methods have major practical and methodological limitations, although they are promising. Models that are trained using data of particular clinical locations usually do not generalize well when deployed to new environments because of hardware differences, acquisition practices, and noise in particular locations (). Furthermore, due to the high sensitivity of the neurodevelopmental data of children, there is a significant privacy constraint, which restricts the capacity of sharing data and centralized model training on a large scale (). Lastly, most effective deep learning algorithms are opaque black boxes that give diagnostic scores without a clear insight into the neurophysiological or language characteristics that could justify the predictions, making them less trustworthy and adoptable in clinical settings ().