Predicting Autism Spectrum Disorder: Transformer-Based Deep Learning Ensemble Framework Using Health Administrative & Birth Registry Data

Background Early diagnosis and access to resources, support and therapy are critical for improving long-term outcomes for children with autism spectrum disorder (ASD). ASD is typically detected using a case-finding approach based on symptoms and family history, resulting in many delayed or missed diagnoses. While population-based screening would be ideal for early identification, available screening tools have limited accuracy. This study aims to determine whether machine learning models applied to health administrative and birth registry data can identify young children (aged 18 months to 5 years) who are at increased likelihood of developing ASD. Methods We assembled the study cohort using individually linked maternal-newborn data from the Better Outcomes Registry and Network (BORN) Ontario database. The cohort included all live births in Ontario, Canada between April 1st, 2006, and March 31st, 2018, linked to datasets from Newborn Screening Ontario (NSO), Prenatal Screening Ontario (PSO), and Canadian Institute for Health Information (CIHI) (Discharge Abstract Database (DAD) and National Ambulatory Care Reporting System (NACRS)). The NSO and PSO datasets provided screening biomarker values and outcomes, while DAD and NACRS contained diagnosis codes and intervention codes for mothers and offspring. Extreme Gradient Boosting models and large-scale ensembled Transformer deep learning models were developed to predict ASD diagnosis between 18 and 60 months of age. Leveraging explainable artificial intelligence methods, we determined the impactful factors that contribute to increased likelihood of ASD at both an individual- and population-level. Results The final study cohort included 703,894 mother-offspring pairs, with 10,964 identified cases of ASD. The best-performing ensemble of Transformer models achieved an area under the receiver operating characteristic curve of 69.6% for predicting ASD diagnosis, a sensitivity of 70.9%, a specificity of 56.9%. We determine that our model can be used to identify an enriched pool of children with the greatest likelihood of developing ASD, demonstrating the feasibility of this approach. Conclusions This study highlights the feasibility of employing machine learning models and routinely collected health data to systematically identify young children at high likelihood of developing ASD. Ensemble transformer models applied to health administrative and birth registry data offer a promising avenue for universal ASD screening. Such early detection enables targeted and formal assessment for timely diagnosis and early access to resources, support, or therapy.