Transformer-based deep learning ensemble framework predicts autism spectrum disorder using health administrative and...

Area of Science:

• Pediatric neurology
• Artificial intelligence in healthcare
• Public health informatics

Background:

• Early diagnosis of autism spectrum disorder (ASD) is crucial for improving long-term outcomes in children.
• Current ASD detection methods rely on symptom-based case finding, often leading to delayed or missed diagnoses.
• Existing population-based screening tools for ASD have limitations in accuracy.

Purpose of the Study:

• To evaluate the efficacy of machine learning models in identifying young children at increased likelihood of developing ASD.
• To utilize health administrative and birth registry data for early ASD risk prediction.
• To explore the feasibility of population-wide ASD screening using advanced computational methods.

Main Methods:

• A cohort of 707,274 mother-offspring pairs was assembled using linked maternal-newborn data from BORN Ontario.
• Extreme Gradient Boosting and ensembled Transformer deep learning models were developed to predict ASD diagnosis.
• Explainable artificial intelligence methods were employed to identify key predictive factors for ASD.

Main Results:

• The best-performing ensemble of Transformer models achieved an AUC of 69.6% for ASD prediction.
• The model demonstrated a sensitivity of 70.9% and a specificity of 56.9% in identifying children at high likelihood of ASD.
• Key factors contributing to ASD likelihood were identified at individual and population levels.

Conclusions:

• Machine learning models applied to routinely collected health data can effectively identify young children at high risk for ASD.
• Ensemble Transformer models offer a promising approach for universal ASD screening, facilitating timely diagnosis.
• This data-driven strategy supports early intervention and access to essential resources for children with ASD.