Examining the Influence of Race and Gender on Autism Spectrum Disorder Screening Using the M-CHAT-R: A Self-Organizing Map Approach

Background: Early screening of Autism Spectrum Disorder (ASD) has been found to facilitate early intervention. The Modified Checklist for Autism in Toddlers-Revised (M-CHAT-R; Robins, Fein, & Barton, 2009) is a 20 item revised version of an evidence-based parent-report instrument recommended by the American Academy of Pediatrics (Johnson & Myers, 2007). One survey found only 60% of pediatricians reported using formal ASD screening at 18 months and only 50% at 24 months (Arunyanart et al., 2012). 

Further, differences in diagnosis based on demographic characteristics (i.e., race and gender) can impact the age of diagnosis. One study indicated White children received a diagnosis earlier than African-American children (Mandell et al., 2002). In addition, there may be gender differences in ASD presentation, as ASD is diagnosed later and less frequently in females—especially high functioning females (Head, et al., 2014; Hiller, Young, & Weber, 2014). Since current screening practices rely on clinician training and subjective judgment, machine learning (ML) may be a powerful complementary scoring tool to examine race and gender group differences. Self Organizing Maps (SOM) is a type of ML that clusters data according to similar responses (Bock, 2003). 

Objectives:   This study examined the use of ML to 1) simplify the administration and scoring of the M-CHAT-R, and 2) improve performance across demographic groups, while examining item-level differences. 

Methods:  This study used SOM to classify groups based on archival data of 14,995 toddlers (46.51% male; 15 years average maternal education), ranging from 16-30 months, collected during their 18- or 24- month well visit. Children at risk were referred for diagnostic assessment. Comparisons of group assignment and item analysis were conducted based on race (Whites versus African-Americans) and gender. MATLAB script was created to interface with the SOM Toolbox (SOM Toolbox Team, Helsinki University of Technology, Finland) and run for different data sets. All 20 M-CHAT-R items were included as inputs and SOM selected the most informative questions to create clusters based on similar traits. 

Results: SOM was able to separate toddlers into risk status groups based on fewer items (12/20 items) and identify symptom presentation differences. Three risk status groups emerged: Typically Developing (TD), Low Risk (LR), and High Risk (HR). Risk status indicates that a toddler may be at risk for any developmental disability, including ASD. In all groups, 100% of ASD cases (determined by follow-up testing) fell into HR clusters. Similar cluster results emerged for Whites and African-Americans, and item analysis indicated deficits in joint attention (JA; i.e., pointing, coordinated gaze) differentiated HR from other groups. While SOM also resulted in three risk status levels in males, SOM failed to provide these differentiations for females, resulting in only LR and HR clusters. This indicates inconsistencies in capturing ASD symptoms in females, such that both the M-CHAT-R and SOM might not as acutely capture at-risk females.  

Conclusions:   This illustrates the potential utility in using ML in ASD screening and the need to target JA in screeners, although different items or algorithms may be needed to improve screening in females.