Secondary Logo

Journal Logo

Institutional members access full text with Ovid®

A Machine Learning Strategy for Autism Screening in Toddlers

Achenie, Luke E. K. PhD*; Scarpa, Angela PhD†,‡; Factor, Reina S. MS†,‡; Wang, Tao MS§; Robins, Diana L. PhD; McCrickard, D. Scott PhD

Journal of Developmental & Behavioral Pediatrics: June 2019 - Volume 40 - Issue 5 - p 369–376
doi: 10.1097/DBP.0000000000000668
Brief Report

Objective: Autism spectrum disorder (ASD) screening can improve prognosis via early diagnosis and intervention, but lack of time and training can deter pediatric screening. The Modified Checklist for Autism in Toddlers, Revised (M-CHAT-R) is a widely used screener but requires follow-up questions and error-prone human scoring and interpretation. We consider an automated machine learning (ML) method for overcoming barriers to ASD screening, specifically using the feedforward neural network (fNN).

Methods: The fNN technique was applied using archival M-CHAT-R data of 14,995 toddlers (age 16–30 months, 46.51% male). The 20 M-CHAT-R items were inputs, and ASD diagnosis after follow-up and diagnostic evaluation (i.e., ASD or not ASD) was the output. The sample was divided into subgroups by race (i.e., white and black), sex (i.e., boys and girls), and maternal education (i.e., below and above 15 years of education completed) to examine subgroup differences. Each subgroup was evaluated for best-performing fNN models.

Results: For the total sample, best results yielded 99.72% correct classification using 18 items. Best results yielded 99.92% correct classification using 14 items for white toddlers and 99.79% correct classification using 18 items for black toddlers. In boys, best results yielded 99.64% correct classification using 18 items, whereas best results yielded 99.95% correct classification using 18 items in girls. For the case when maternal education is 15 years or less (i.e., associate degree and below), best results were 99.75% correct classification when using 16 items. Results were essentially the same when maternal education was 16 years or more (i.e., above associate degree); that is, 99.70% correct classification was obtained using 16 items.

Conclusion: The ML method was comparable to the M-CHAT-R with follow-up items in accuracy of ASD diagnosis while using fewer items. Therefore, ML may be a beneficial tool in implementing automatic, efficient scoring that negates the need for labor-intensive follow-up and circumvents human error, providing an advantage over previous screening methods.

This article has supplementary material on the web site:

Departments of *Chemical Engineering,


Computer Science, Virginia Polytechnic Institute and State University, Blacksburg, VA;

Virginia Tech Center for Autism Research, Blacksburg, VA;

§Department of Economics, University of California, Riverside, CA;

A.J. Drexel Autism Institute, Drexel University, Philadelphia, PA.

Address for reprints: Luke E. K. Achenie, PhD, Department of Chemical Engineering, Signature Engineering Building, Room 273, Virginia Polytechnic Institute and State University, Blacksburg, VA 24061; e-mail:

This study was funded in part by the Virginia Tech Institute for Society, Culture and Environment (ISCE): Summer Scholars Program and by the Virginia Tech Center for Autism Research. Archival data used in analyses were funded by the Eunice Kennedy Shriver National Institute for Child Health and Human Development, R01 HD 039961.

D.L. Robins receives royalties for commercial products that incorporate the Modified Checklist for Autism in Toddlers, Revised/Follow-Up (M-CHAT-R/F) and is co-owner of M-CHAT, LLC, which licenses commercial use of the M-CHAT-R/F. The remaining authors declare no conflict of interest.

Supplemental digital content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal's Web site (

Received February 02, 2018

Accepted February 04, 2019

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.