HCC: Small: Toward Computational Modeling of Autism Spectrum Disorder: Multimodal Data Collection, Fusion, and Phenotyping

自闭症谱系障碍（ASD）是一种复杂的神经发育障碍，在美国每54名儿童中就有1名受到影响。ASD可以说是我们这个时代最大的公共卫生挑战之一，它对儿童及其家庭造成了重大影响，更不用说给当前的医疗保健和教育系统带来了负担。尽管经过了几十年的研究，与ASD相关的许多基本问题仍然存在，从早期诊断到个性化干预。自闭症谱系障碍的异质性极大地增加了在遗传或行为上识别与该疾病相关的特定特征（即表型）的难度。除了明显增加的患病率和未知的病因外，ASD表型建模一直是自闭症研究中一个长期存在的开放性问题。提高对ASD表型的认识可以为更准确的ASD诊断和更有效的干预提供新的见解。该项目旨在基于行为测量和传感器收集的数据，包括神经记录、眼动追踪、视频/音频捕获和其他传感器数据，了解ASD生物标志物。通过多学科合作，该项目将在行为科学和数据驱动的计算神经科学方面取得变革性进展。改进和早期诊断可以大大提高ASD个体及其社区的生活质量。该项目将为培养神经科学与计算机科学交叉领域的研究生和本科生提供一个极好的平台。该项目将采用多模态数据驱动的方法，将行为成像数据（眼动追踪、音频/视频）与功能磁共振成像（fMRI）、脑电图(EEG)/脑磁图（MEG）等神经成像数据相结合，解决自闭症谱系障碍建模问题。研究小组将开展多模态数据融合，提取ASD相关的生物标志物，而不需要特征工程和数据驱动建模，以了解ASD的神经基础，特别是行为和传感器导向信号之间的关系。这种基于多模态数据的建模将结合不同模态的突出ASD生物标志物的互补信息，例如动态功能连接。为了避免基于启发式的ASD表型特征工程，研究人员将使用两种基于流的深度学习技术以及XAI可解释的AI（人工智能）。XAI将为深度学习算法做出的决定提供解释，以识别与ASD相关的特征。除了ASD诊断之外，多模态神经成像将导致对ASD丰富性和复杂性的研究，这里称为ASD表型。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder affecting one out of 54 children in the US. ASD is arguably one of the greatest public health challenges of our time, which has imposed a significant impact on children and their families, not to mention the burden on the current healthcare and educational systems. Despite decades of research, many fundamental issues related to ASD remain from early diagnosis to personalized intervention. The heterogeneity of ASD has contributed significantly to the difficulty in identifying the specific traits associated with this disorder (i.e., phenotyping), genetically or behaviorally. In addition to apparently increasing prevalence and unknown etiology, modeling the ASD phenotype has remained a long-standing open problem in autism research. An improved understanding of ASD phenotypes can shed novel insight to both more accurate diagnosis and more effective intervention of ASD. This project aims to understand ASD biomarkers based on behavioral measurement and sensor-gathered data, including neural recording, eye tracking, video/audio capture, and other sensor data. Through multi-disciplinary collaboration, this project will lead to transformative advances in behavioral science and data-driven computational neuroscience for ASD phenotyping. Improved and earlier diagnosis can substantially improve quality of life of ASD individuals and their communities. This project will provide an excellent platform to train both graduate and undergraduate students at the intersection of neuroscience and computer science.This project will address the problem of ASD modeling by taking a multimodal data-driven approach integrating behavior imaging data (eye-tracking, audio/video) with neuroimaging data such as functional magnetic resonance imaging (fMRI), electroencephalography (EEG)/ magnetoencephalography (MEG). The research team will carry out multimodal data fusion to extract ASD-relevant biomarkers without feature engineering, and data-driven modeling to obtain an understanding of the neural underpinnings of ASD, especially in the relationship between behavioral and sensor-oriented signals. This multimodal data-based modeling will combine complementary information about salient ASD biomarkers, such as dynamic functional connectivity, across different modalities. To avoid heuristics-based feature engineering for ASD phenotyping, the researchers will use two stream-based deep learning techniques along with XAI explainable AI (Artificial Intelligence). XAI will provide the interpretations for the decisions made by the deep learning algorithms to identify the traits associated with ASD. In addition to ASD diagnosing, multimodal neuroimaging will lead to investigations into the richness and complexity of ASD, referred to here as ASD phenotyping.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.