Abstract

Autism Spectrum Disorder is a complex neurological disorder characterized by developmental disabilities that cause significant social and behavioral challenges. In 2018, CDC reported 1 of 54 children in the US were identified with ASD. Symptoms vary and exacerbate with age, requiring life-long support. Individuals with ASD are frequently subject to stigma, discrimination, and human rights violations. 

​

Although early signs of ASD usually appear around 12-24 months, diagnosis at such an early age is an arduous process since young children lack the cognitive ability required to take the complex standardized tests currently used. This leads to a delay in diagnosis and treatment, increasing the average age of diagnosis to 4-6 years. Studies indicate that children who receive early intervention and support at key developmental stages are more likely to gain essential social skills and interact better in society. 

​

Our product is a deep learning, multimodal algorithm that enables prompt and accurate diagnosis of ASD with 95.8% accuracy. We utilized functional MRI imaging data and phenotypic history which can be obtained at the earliest sign of symptoms [ABIDE II Multisite Repository]. First, we developed a 53-layer convolutional neural network for imaging data using MobileNetV2 and a phenotypic classification algorithm for statistical data using KNN. Then, we programmed three ensembles to integrate these models, ultimately creating a sophisticated, multimodal algorithm achieving 95.8% accuracy. Our model, when combined with a clinical exam and medical expertise, can facilitate early detection and intervention of ASD, improving outcomes for children across the globe.