• Author / Creator
    Reid, Kyle Burke
  • Background: The Autism Observation Scale for Infants (AOSI) has been used to investigate early features of children with an increased likelihood (IL) of a later diagnosis of autism spectrum disorder (ASD). Though independent research groups have evaluated its use in IL infant siblings (younger siblings of the children with ASD), recent studies have examined the AOSI’s use in other IL populations. Since first published in 2005, an assessment of the AOSI’s discriminatory and predictive utility in infant siblings and other IL populations is warranted. In addition, growing access to sophisticated computational technology has facilitated increased use of powerful computing techniques such as machine learning in research and healthcare spaces. While common in clinical and research fields, ASD research has yet to fully leverage this technology. Currently, Bussu et al. (2018) is the only study to have generated supervised machine learning classifiers using early AOSI data.

    Objectives: (1) examination of research assessing the individual classification properties and group differences of the AOSI across different IL groups from 6-18 months, (2) generation of supervised machine learning classifiers using 12-month AOSI and Mullen Scales of Early Learning (MSEL) data in a cohort of infant siblings (n=373), and (3) assessment of classifier performance at predicting 36-month ASD diagnosis in infant siblings from two Canadian longitudinal studies (n=92; n=90).

    Methods: A systematic search for relevant articles was conducted across six databases: CINAHL, EMBASE-OVID, ERIC, JSTOR, PubMed, and Web of Science, with articles independently reviewed for inclusion and exclusion criteria by two reviewers. Supervised machine learning classifiers using logistic regression (with and without regularization) and support vector machines using linear, polynomial, and radial basis function kernels were generated in R/RStudio using combinations of participant biological sex, 12-month MSEL standard scores (Visual Reception, Receptive Language, Expressive Language, Fine Motor, and Early Learning Composite), and 12-month AOSI item-level and Total Score data. Factor analysis (informed by principal axis parallel analysis) was used as a means of reducing item-level AOSI data dimensionality during modelling to mitigate model overfitting. Classifiers were assessed by their ability to predict 36-month ASD diagnosis in subsets of infant siblings (n=92; n=90) from two Canadian longitudinal cohorts.

    Results: The systematic search identified 354 articles with 17 meeting inclusion criteria. Four IL infant populations were assessed: younger siblings of children diagnosed with ASD, and infants with Fragile X Syndrome (FXS), Tuberous Sclerosis Complex (TSC), and Down Syndrome (DS). The systematic review had three main findings. First, five studies reported individual classification properties, although they did not use a consistent approach. Second, stable group differences emerged between IL non-ASD, IL-ASD, and infants at low likelihood of ASD (i.e., no family history) beginning at 12 months. Third, meta-analyses resulted in a large effect size for comparisons between low likelihood and IL-ASD samples and a moderate effect size for comparisons of IL non-ASD and IL-ASD samples. For supervised machine learning classifiers built with 12-month data, best-performing classifiers across all algorithm types were between 76-77% accurate and had areas under the curve (AUC) between 0.73 and 0.76. Though their specificity was excellent (0.94-1.0), they were characterized by extremely poor sensitivity (0-0.19). Relative to the performance of a 12-month AOSI Total Score cut point of 7 at predicting 36-month ASD diagnosis (informed by Youden index assessment; AUC = 0.66, sensitivity = 0.52, specificity = 0.74), machine learning classifiers had enhanced AUC and specificity, but significantly decreased sensitivity. The best-performing classifiers in this study yielded higher accuracy, AUC, and specificity (but not sensitivity) relative to the best performing classifier generated by Bussu et al. (2018) using 14-month data (accuracy = 64%, AUC = 0.71, sensitivity = 0.61, specificity = 0.67) using similar machine-learning methodology.

    Conclusion: Utility of the AOSI to identify early signs of ASD in IL populations, including infant siblings of children diagnosed with ASD, FXS, TSC, and DS was demonstrated. Though the best-performing supervised learning classifiers performed below levels recommended for early screening, accuracy, AUC, and specificity were moderately improved relative to those generated by Bussu et al. (2018) using 14-month AOSI data. Further exploration into feature selection, extraction, or inclusion of 12-month AOSI and MSEL data may allow continued refinement of machine learning models built using 12-month clinical data and capable of predicting ASD at 36-months.

  • Subjects / Keywords
  • Graduation date
    Spring 2023
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.