Application of AI Algorithms to Develop ASD Investigation Process for Early Screening of Autism in Toddlers

Abstract

Autism Spectrum Disorder (ASD) is a neurobiological disorder related to human brain develop- ment that usually remains with a person for life. According to Diagnostic and Statistical Manual of

Mental Disorders Edition-5 (DSM-5) criteria, ASD marked by setbacks in communication, indigent social

interactions, restricted interest, and repetitive behavior that are becoming increasingly common in vari- ous parts of the world. In Pakistan, many of the autistic cases left unreported or misreported majorly due

to un-awareness of disease and early signs, insufficient health facilities, unavailability of trained health- care professionals and social stigma. Symptoms of autism can mostly be identified at its initial stages

and screened mostly in 6 months after birth. But due to the lack of knowledge, the disease remained undiagnosed for the longer time. Diagnosing ASD is a challenge because, so far, there is no unique standard medical test available for its detection. Its symptoms are usually recognized by observations and development history of a child. The foremost tailback in early detection of ASD rests in irrational and monotonous nature of prevailing diagnosis procedures. Because of this fact, after initial apprehension, the parents must wait for at least 13 months for the actual diagnosis. A common issue attached with ASD labeling is that it is time consuming and cost impotent. It takes several hours/sessions/days of thorough examinations. Early prediction of ASD is important so, in addition to conventional diagnostics, self-administered assessment methods (also referred as screening tools) have been developed. Among these screening tools Quantitative CHecklist for Autism in Toddlers (Q-CHAT) is one of the most commonly used screening method to assess ASD traits. Literature review indicate that most of the screening tools and applications were developed to screen adults (18+ year old), children (4 to 11 years old), and teenagers (11 to 18 years old). There is no such tool available for screening of toddlers. Furthermore, in DCM-5, 2 categories are defined for autism screening. Category A defines the social interaction and communication while the category B defines the restricted and repetitive behavior. According to DCM-5 criteria if both categories are satisfied, it will refer to autistic patients otherwise not. While observing the literature review of recent studies, reveal that most of the machine learning models and applications were applied and designed on the bases of Q-CHAT-10 questionnaire in which only category A is satisfied. According to DSM-5 criteria if screening tool have questions related to only social and communication interactions then the child is suffering from

xii

social communication disorder. If we want to assess ASD, then the screening tool must have questions for both social and communication interactions and restricted and repetitive behaviors. Following are major issues with respect to the subjective assessments of ASD features in a child. I. Lack of standardization in the screening process II. Reluctance towards adoption of data driven tech solutions to enhance ASD screening.

This research will be a step towards standardization in subjective assessment of ASD. Rational- izing the screening process by designing an efficient, time and cost-effective DSM-5 criteria-based ASD

screening procedure that will help professionals and parents in general for early screening of ASD. To solve the above-mentioned problem, we use secondary data set of 252 responses on 25 features

(termed as Q-CHAT 25) with 10 demographic and other features of the respondents (in total of 35 fea- tures) culminating into a binary target feature (ASD Yes or No). Machine learning algorithms including

logistic regression, decision tree, random forest and support vector machine are applied to predict the target feature. For features selection, point bi-serial correlation, association of attributes, and random forest classifier–recursive feature elimination (RFC–RFE) algorithm has been used. Assessment measures such as precision, accuracy, sensitivity, specificity and F1 score has been implemented to obtain the best suited algorithm. After final selection of subsets of features and ML model, desktop based and mobile

based application is designed on the basis of selected features and model to assess health care profession- als and parents with respect to ASD early screening..

On the basis of performance evaluation parameters random forest with 16 features classify ASD best with 94 percent accuracy, 92 percent precision and 94 percent for both specificity and sensitivity. Out of these 16 features 11 features belongs to core domain named social and communication interactions, 5 features belong to restricted and repetitive behaviors. From this observation it is clear that our selected features fulfilled the pre-defined criteria of DSM-5 for assessment of ASD. In our selected features we have 10 out of 16 features are common with QCHAT-10 features presented by Carrie Allison. It implies that features which are common among our selected features and QCHAT-10 are related to social and communication interactions. This research proposed a new user-friendly desktop based as well as mobile based ASD screening solution which is specifically used for screening of ASD in toddlers from age 18 month to 2 years. This

application contains 16 questions linked to images, so the entire user population can utilize this applica- xiii

tion. This user-friendly and cost-effective solution is necessary for pre-diagnosing ASD since it increases user accessibility in an easy and time-efficient platform as well as being accessible to a wide range of users, including caregivers, parents, and health professionals. These findings will be helpful to professionals and parents in manifolds like cost as well as time effective solution of screening of ASD child based on standard international scientific guidelines, bringing awareness in general public/ society regarding crucial features in the development of a child suspected to ASD in long term, hence breaking social stigmas. This will also indirectly helps the autistic children in getting good quality support and basic emotional, social, and educational needs in a better way