Problem Behavior Prediction for Severe Autism Spectrum Disorder using Patient and Environmental Data

Introduction: Autism Spectrum Disorder (ASD) now affects 1 in 44 children by 8 years of age [1]. While ASD is typically characterized by social and communication impairments, such as atypical behaviors [2, 3], it also has a high incidence of co-occurring conditions [4]. These comorbidities include sleep disorders [5], gastrointestinal (GI) problems [6], and immune disorders [7], among others. Severe behaviors that may cause harm to the person with ASD, peers, or caregivers, including aggression and self-injurious behavior, may be influenced by discomfort from co-occurring conditions or from environmental factors. Sleep patterns have been shown to be predictive of behavioral episodes [8]. In this study, we examine sleep patterns and additional predictors, including daily GI, atmospheric, and allergen data, to determine their predictive relationships with severe behavioral episodes.

Materials and Methods: Patient data, including GI, sleep, and behavior data, were obtained by The Center for Discovery (TCFD), a non-profit provider of educational, health, clinical and residential services for children and adults with ASD and other complex disabilities, between 2015 and 2021. Parental consent and individual assent were required for data collection and the study was approved by the TCFD Institutional Review Board. Patients were required to be 19 years of age or younger and have a proportion of 40% or less missing data, with at least 20 examples of the behavior and a maximum of 18 months of data [8] (n = 19, with 3933 total observations). All patients were diagnosed with ASD. Atmospheric and allergen data were obtained from the National Oceanic and Atmospheric Administration (NOAA) [9] and MoonCalc.org [10], and the American Academy of Allergy, Asthma and Immunology (AAAAI) [11], respectively, from the stations located nearest to TCFD. Variables were lagged such that prior days were matched with subsequent days’ behavior, and a logistic regression model was fit to account for both categorical and continuous features.

Results and Discussion: Balanced accuracies of over 65% for predicting severe behavioral episodes were obtained for the population of patients included in the study. Balanced accuracy was used to mitigate class imbalance [8], and logistic regression models were assessed using leave-one-out cross-validation. Prior behavior was a predictor of note for future behavior. Conclusions: Prior data collected for patients with severe ASD allows severe behavior to be predicted with a reasonable accuracy. Caregivers and facility staff can use models based on patient data already collected to predict whether an individual will need extra attention to prevent harmful outcomes of severe behavioral episodes, or to potentially mitigate the behavior altogether.

Acknowledgements: The authors are grateful to TCFD, the NOAA, the AAAAI, and MoonCalc.org for providing the data used in this study.

References: 1. CDC, 2018. 2. Faras, H., Ann Saudi Med, 2010. 30: p. 295-300. 3. Dominick, K.C., Res Dev Disabil, 2007. 28: p. 145-62. 4. Tye, C., Front Psychiatry, 2018. 9: p. 751. 5. Singh, K., Semin Pediatr Neurol, 2015. 22: p. 113-25. 6. Chakraborty, P., Autism, 2021. 25: p. 405-415. 7. Hughes, H.K., Front Cell Neurosci, 2018. 12: p. 405. 8. Cohen, S., Autism Res, 2018. 11: p. 391-403. 9. NOAA, 2015-2021. 10. Hoffmann, T., Mooncalc.org, 2014-2022. 11. AAAAI, 2015-2021.