Early Prediction of Autistic Spectrum Disorder Using Developmental Surveillance Data

• Published in: JAMA network open 2024-01, Vol.7 (1), p.e2351052
• Main Authors: Amit, Guy, Bilu, Yonatan, Sudry, Tamar, Avgil Tsadok, Meytal, Zimmerman, Deena R, Baruch, Ravit, Kasir, Nitsa, Akiva, Pinchas, Sadaka, Yair
• Format: Article
• Language: English
• Subjects: Autism Spectrum Disorder - diagnosis; Autism Spectrum Disorder - epidemiology; Autistic Disorder - diagnosis; Autistic Disorder - epidemiology; Child; Child, Preschool; Children & youth; Childrens health; Cohort analysis; Cohort Studies; Female; Gestational age; Humans; Infant; Male; Retrospective Studies; Surveillance
• ISSN: 2574-3805; 2574-3805
• DOI: 10.1001/jamanetworkopen.2023.51052

With the continuous increase in the prevalence of autistic spectrum disorder (ASD), effective early screening is crucial for initiating timely interventions and improving outcomes. To develop predictive models for ASD using routinely collected developmental surveillance data and to assess their performance in predicting ASD at different ages and in different clinical scenarios. This retrospective cohort study used nationwide data of developmental assessments conducted between January 1, 2014, and January 17, 2023, with minimal follow-up of 4 years and outcome collection in March 2023. Data were from a national program of approximately 1000 maternal child health clinics that perform routine developmental surveillance of children from birth to 6 years of age, serving 70% of children in Israel. The study included all children who were assessed at the maternal child health clinics (N = 1 187 397). Children were excluded if they were born at a gestational age of 33 weeks or earlier, had no record of gestational age, or were followed up for less than 4 years without an ASD outcome. The data set was partitioned at random into a development set (80% of the children) and a holdout evaluation set (20% of the children), both with the same prevalence of ASD outcome. For each child, demographic and birth-related covariates were extracted, as were per-visit growth measurements, quantified developmental milestone assessments, and referral summary covariates. Only information that was available before the prediction age was used for training and evaluating the models. The main outcome was eligibility for a governmental disabled child allowance due to ASD, according to administrative data of the National Insurance Institute of Israel. The performance of the models that predict the outcome was evaluated and compared with previous work on the Modified Checklist for Autism in Toddlers (M-CHAT). The study included 1 187 397 children (610 588 [51.4%] male). The performance of the ASD prediction models improved with prediction age, with fair accuracy already at 12 months of age. A model that combined longitudinal measures of developmental milestone assessments with a minimal set of demographic variables, which was applied at 18 to 24 months of age, achieved an area under the receiver operating characteristic curve of 0.83, with a sensitivity of 45.1% at a specificity of 95.0%. A model using single-visit assessments achieved an area under the receiver operating characteristic cur