Data-driven approaches to identify biomarkers from multimodal imaging big data

1.项目总结/摘要 翻译生物标志物在脑疾病中的研究是一个非常具有挑战性和富有成效的方法， 将有助于更好地了解健康和患病的大脑。该项目将促进 先进的工程解决方案和数学工具，以新颖的神经影像学在精神科的应用 包括重性抑郁症（MDD）、双相情感障碍（BD）和精神分裂症（SZ）在内的疾病， 对高度复杂的数据集进行复杂而强大的分析。迄今为止，统一的综合征分类 （ICD-9/10;DSM-IV/5）模糊了我们对潜在病理生理学的认识， 无法指导最佳治疗。例如，没有生物标志物能够精确地预测反应， 一些治疗方法。其中一个原因是，迄今为止，大多数神经成像预测研究都使用 单一成像测量或报告的简单相关关系，而不考虑多模态交叉- 信息、非线性关系或多站点交叉验证。因此，开发新的数据挖掘 深度学习、参考融合和稀疏回归等技术可以补充和利用 丰富的神经影像学数据，为识别客观生物标志物提供了有前途的途径， 描述性使用脑成像传统上用于研究脑部疾病的个性化预测。 我们将通过开发3种新的数据驱动方法来促进翻译生物标志物的鉴定：1）A 监督融合模型，可以提供认知障碍如何影响共变脑功能的见解 和结构的精神障碍，通过使用不同的临床测量作为参考，以指导多模态MRI 融合; 2）具有聚合特征选择技术的前沿预测框架，能够 更精确地估计临床结果，例如，个体MDD患者的缓解/复发状态 电休克治疗（ECT）使用基线脑成像和人口统计学措施3）我们将绘制 深度学习与分层相关传播（LRP）或注意力相结合的进步和想法 模块，通过纳入动态功能措施对多组精神疾病进行分类。的 提出的（深度/递归/卷积神经网络，DNN/RNN/CNN）模型将增强 可解释性，能够追溯并从输入中发现最具预测性的功能网络。所有 上述提出的方法将应用于包含多模态成像和行为成像的大数据。 从现有研究中汇集的信息（n~5000），以及我们开发的开源工具箱将被共享 公开地这项开创性的研究可能为治疗和诊断 精神疾病，从而指导个性化的临床护理。该项目的完成具有很大的 发现现有方法遗漏的神经影像学生物标志物的潜力，导致更早的 更有效的干预措施，并为重大的转化影响奠定基础。

1. PROJECT SUMMARY/ABSTRACT The study of translational biomarkers in brain disorders is a very challenging and fruitful approach, which will empower a better understanding of healthy and diseased brains. This project will promote the translation of advanced engineering solutions and mathematic tools to novel neuroimaging applications in psychiatric disorders including major depression disorder (MDD), bipolar disorder (BD) and schizophrenia (SZ), allowing sophisticated and powerful analyses on highly complex datasets. To date, the unifying syndrome classification (ICD-9/10;DSM-IV/5) for these mental disorders obscures our knowledge of underlying pathophysiology and cannot guide optimal treatments. For example, there is no biomarker that is able to precisely predict response of MDD to some treatments. One reason for this is that most neuroimaging prediction studies to date have used a single imaging measure or reported simple correlation relationships, without considering multimodal cross- information, nonlinear relationships, or multi-site cross-validation. Hence, developing novel data mining techniques such as deep learning, fusion with reference, and sparse regression can complement and exploit the richness of neuroimaging data, providing promising avenues to identify objective biomarkers and going beyond a descriptive use of brain imaging as traditionally used in studies of brain disease to individualized prediction. We will facilitate the translational biomarker identification by developing 3 novel data-driven methods: 1) A supervised fusion model that can provide insight on how cognitive impairment may affect covarying brain function and structure in mental disorder, by using different clinical measures as a reference to guide multimodal MRI fusion; 2) A cutting-edge prediction framework with aggregated feature selection techniques that is able to estimate clinical outcome more precisely, e.g., remission/relapse status of individual MDD patient after electroconvulsive treatment(ECT) using baseline brain imaging and demographic measures of 3) We will draw on advances and ideas from deep learning combined with layer-wise relevance propagation (LRP) or attention modules, to classify multiple groups of psychiatric disorders by incorporating dynamic functional measures. The proposed (Deep/Recurrent/Convolutional Neural Network, DNN/RNN/CNN) models will have enhanced interpretability that is able to trace back and discover the most predictive functional networks from input. All above proposed methods will be applied to big data containing both multimodal imaging and behavioral information (n~5000) pooled from existing studies, and our developed open-source toolboxes will be shared publicly. This pioneering study may provide an urgently-needed paradigm shift in the treatment and diagnosis of psychiatric disorders, thereby guiding personalized clinical care. Accomplishment of this project has great potential to discover neuroimaging biomarkers that have been missed by existing approaches, leading to earlier and more effective interventions, and laying the groundwork for a significant translational impact.