Early Prediction of Cognitive Deficits in Very Preterm Infants using Machine Learning and Brain Connectome

项目摘要/摘要 神经发育障碍的高风险是父母和临床医生关心的主要问题 早产儿。每年，大约有22,000名极早产儿 (即≤32周胎龄) 美国人会出现认知缺陷。 努力针对预防神经发育的干预措施 后遗症由于我们目前无法在3至5岁之前诊断残疾而受到阻碍。 研究支持这样一种观点，即认知缺陷可能是由于神经连接和 沟通。最近对成年人和年龄较大的儿童的脑连接组研究表明，异常网络 这些特性可作为预测预后的生物标志物。其中许多研究都利用了机器学习模型 基于大脑连接体数据来预测各种神经疾病，然而，这 进展还没有完全扩大到早产人口。我们的初步研究表明，早期和 使用基于以下的机器学习模型可以准确预测个体水平的认知缺陷 足月校正年龄(CA)的脑连接体特征。我们正确地对91.3%的早产儿进行了分类 认知缺陷高危婴儿的特异度为90%，敏感度为92.3%。我们的总体目标是 开发一个健壮的机器学习模型，可以分析整合的结构和功能的大脑 在CA期时获得的连接组数据，以预测极早产儿以后的认知障碍。 我们的中心假设是，机器学习技术分析整合的结构和功能的大脑 出生时的连接体特征可以在早产儿的个体水平上预测2年后认知障碍 婴儿正确率超过90%，超过当前经典多变量的表现 分析。检验中心假设的两个具体目标是：1)开发和实现一台机器 用于提取高级大脑连接体特征的学习模型以及2)开发和验证机器学习 预测认知缺陷的框架。在完成第一个目标后，我们将解释大脑连接体， 并提取最能代表大脑连接体的高维连接体特征。在第二个 目的：将我们提出的机器学习模型应用于两个认知缺陷(即两类)的预测 分类)和连续等级(即回归)的认知分数。要量化 对于模型的判别，我们还将验证其在未用于模型的数据中的性能 发展，并与目前传统的多变量方法进行比较。拟议的研究是 意义重大，因为它将增加关于早产儿大脑连接体发育的科学知识 并有助于及早识别神经发育缺陷的高危婴儿，以便及时 最佳认知结果的临床干预。

Project Summary/Abstract The high risk of neurodevelopmental impairments is a major concern for parents and clinicians caring for premature babies. Annually, approximately 22,000 very preterm infants (i.e. ≤32 weeks gestational age) in the United States develop cognitive deficits. Efforts to target interventions to prevent neurodevelopmental sequelae are hampered by our current inability to diagnose disabilities before the age of 3 to 5 years. Research supports the notion that cognitive deficits may result from a perturbation of neural connection and communication. Recent brain connectome studies in adults and older children show that abnormal network properties are useful as prognostic biomarkers. Many of these studies have exploited machine learning models based on brain connectome data for the prediction of a variety of neurological conditions, however this progress has not been fully extended to the preterm population. Our preliminary studies suggest that early and accurate prediction of cognitive deficits at an individual level is possible using machine learning models based on brain connectome features at term corrected age (CA). We have correctly classified 91.3% of very preterm infants at high risk of cognitive deficits with 90% specificity and 92.3% sensitivity. Our overall objective is to develop a robust machine learning model that can analyze integrated structural and functional brain connectome data obtained at term CA to make a prediction of later cognitive deficits in very preterm infants. Our central hypothesis is that machine learning techniques analyzing integrated structural and functional brain connectome features at birth can predict cognitive deficits at 2 years CA at an individual level in very preterm infants with accuracy of greater than 90%, exceeding the performance of current classical multivariate analyses. The two specific aims to test the central hypothesis are: 1) Develop and implement a machine learning model to extract high-level brain connectome features and 2) Develop and validate a machine learning framework to predict cognitive deficits. On completion of the first aim, we will explicate the brain connectome, and extract high-dimensional connectome features that best represent the brain connectome. In the second aim, the machine learning model we proposed will be applied in predicting both cognitive deficit (i.e. 2-class classification) and cognitive scores on a continuous scale (i.e., regression) at 2 years CA. To quantify the model's discrimination, we will also validate its performance in data that are not used for the model development, and compare with the current conventional multivariate approach. The proposed research is significant because it will increase scientific knowledge about the developing brain connectome in very preterm infants and facilitate earlier identification of babies at high risk of neurodevelopmental deficits, allowing timely clinical interventions for optimal cognitive outcome.