A Comprehensive Database Based on Multiple Data Sources to Facilitate Diagnosis of ASD

Autism spectrum disorder (ASD) is a neurodevelopmental disorder which has an increasing prevalence in children. ASD is clinically highly heterogeneous and lacks objective diagnostic criteria. In recent years, magnetic resonance imaging and genomics have been widely used in the diagnosis of ASD, and some valuable biomarkers have been found, which has improved people’s understanding of the neural and molecular development mechanism of ASD. However, most studies focus on limited data sources with lack of integration research, thus leading to inconsistent or biased results. In this paper, we design a compute-aided diagnosis framework based on multiple ASD-relevant data sources for purpose of distinguishing the ASD patients more accurately. We first establish a multiple data collection procedure from initial diagnosis to regular follow-up visits. Various medical big data including structured and unstructured forms are collected from different devices or protocols and then they are deposited and accessed based on Hadoop platform. Furthermore, we design a classification framework to identify ASD patients by integrating the complementary information from multiple data sources. Deep learning is used to extract features from each data source automatically, and then all extracted features are integrated by the multiple kernel learning method for improving the diagnostic accuracy of ASD.