Reduced False Positives in Autism Screening Via Digital Bio-markers Inferred from Deep Co-morbidity Patterns

Abstract Autism spectrum disorder (ASD) is a developmental disability associated with significant social and behavioral challenges. There is a need for tools that help identify children with ASD as early as possible. Our current incomplete understanding of ASD pathogenesis, and the lack of reliable biomarkers hampers early detection, intervention, and developmental trajectories. In this study we develop and validate machine inferred digital biomarkers for autism using individual diagnostic codes already recorded during medical encounters. Our risk estimator identifies children at high risk with a corresponding area under the receiver operating characteristic curve (AUC) exceeding 80% from shortly after two years of age for either sex, and across two independent databases of patient records. Thus, we systematically leverage ASD co-morbidities - with no requirement of additional blood work, tests or procedures - to compute the Autism Co-morbid Risk Score (ACoR) which predicts elevated risk during the earliest childhood years, when interventions are the most effective. By itself, ACoR has superior performance to common questionnaires-based screenings such as the M-CHAT/F, and has the potential to reduce socio-economic, ethnic and demographic biases. In addition to superior standalone performance, independence from questionnaire based screening allows us to further boost performance by conditioning on the individual M-CHAT/F scores - we can either halve the false positive rate of current screening protocols or boost sensitivity to over 60%, while maintaining specificity above 95%. Adopted in practice, ACoR could significantly reduce the median diagnostic age for ASD, and reduce long post-screen wait-times experienced by families for confirmatory diagnoses and access to evidence based interventions.