The use of mathematical models to assess the probability of an autism spectrum disorder diagnosis in children

Abstract

Autism spectrum disorder (ASD) is a complex, multisystem disorder whose symptoms can range in severity, as well as be confounded by symptoms of co-occurring conditions (COCs). 1 in every 44 eight-year-old children in the United States is diagnosed with ASD, and 95% of children diagnosed are also diagnosed with at least one COC. This can lead to a diversity in ASD presentation and difficulties in diagnosis especially with current ASD diagnostic standards limited to a clinical evaluation on physical or behavioral abnormalities. This work explores the use of mathematical models to assess the risk of an ASD diagnosis in children. The similarities between biomarker identification and biological pathway models are discussed, highlighting the shared problem of regularization used to avoid overfitting. Solutions to this problem translates across model types and an example is shown using a candidate biomarker for ASD diagnosis. Fisher discriminant analysis and support vector machines are compared on their biomarker identification using a dataset comprising metabolic measurements from four separate studies. This work also uses risk analysis models to identify prenatal factors that occur during pregnancy and are associated with (1) having a child diagnosed with ASD and (2) belong to the three following pre-identified subgroups of children based on COCs: children with a High-Prevalence of COCs, children with mainly developmental delays and seizures (DD/Seizures COCs), and children with a Low-Prevalence of COCs. These retrospective analyses on maternal medical claims are made up of more than 100,000 mothers with a diverse mixture of ages, ethnicities, and geographical regions across the United States. Logistic regression analysis revealed that having a biological sibling with ASD, maternal use of antidepressants or psychiatric services, as well as non-pregnancy related claims such as hospital visits, surgical procedures, and radiology exposure were associated with an increased risk of having a child diagnosed with ASD. It also found that while some risk factors were shared between all three COC-based subgroups, unique factors were identified distinguishing the three groups. Anti-inflammatories, infections, or other complex medications were associated with the High-Prevalence COCs group; immune deregulatory conditions such as asthma or joint disorders were associated with the DD/Seizures COCs group; and overall pregnancy complications were most influential with the Low-Prevalence COCs group. Thus, previously identified subgroups of children with ASD have distinct associated prenatal risk factors. The use of mathematical models to assess ASD diagnoses in children will further develop both the diagnostic standards and current understanding of ASD pathophysiology.