Consortium Of MRI Biomarkers In Neonatal Encephalopathy (COMBINE)

摘要 新生儿缺氧缺血性脑病（HIE）是一种发生率约为5/1000的脑损伤。2005年，NIH 新生儿研究网络（NRN）建立了治疗性低温（TH），在前6个月冷却患者， 在高收入国家，新生儿缺氧缺血性脑病的标准治疗方法是在出生后33-34°C下持续72小时。然而，在这方面， 许多患者仍然经历不良结局（死亡或婴儿发育的认知Bayley量表<85） 18-22个月。因此，从2008年到2015年，NRN测试了更深、更长或更晚的TH是否进一步减少了不利影响。 结果，在21个地点进行了两项试验。不幸的是，结果不确定，进一步的进展缓慢， 这主要是因为在18-22个月之前无法可靠地评估不良结局。为了加快治疗 创新和评估新疗法的影响，以更及时的方式，有一个紧迫的，但未满足 需要建立18-22个月不良结局的新生儿生物标志物。为了弥补这一差距，NRN 使用神经放射学专家对临床获得的新生儿脑损伤进行评分， 脑磁共振成像（MRI），称为NRN MRI评分。在一个数据集中有一个读者， 不良结局的敏感性/特异性为81%/78%。然而，在另一个具有两个读取器的数据集中， 阅片者之间的一致性仅为中等，不良结局的特异性仅为56- 69%。问题 出现这种主观和耗时的评分系统是否可靠或完全表征复杂 新生儿缺氧缺血性脑病损伤模式。此外，在许多国家，没有专家可以进行MRI评分。最后， 重要的临床数据元素，如出生体重、性别、APGAR评分、社会经济地位和方面 临床考试的评分没有完全融入评分系统。我们的总体假设是， 新生儿脑MRI和临床数据元素的智能（AI）算法可以提供更高的灵敏度， 在预测18-22个月的不良HIE结局方面，NRN MRI评分的特异性高于专家评分。我们的R61目标 目的1，从两个已完成的NRN多中心HIE试验中编译大型HIE数据集（N=430）;目的 2、使用新生儿脑MRI开发AI生物标志物，并与Aim 2a的NRN评分进行比较 关注MRI损伤模式和关注全脑MRI信号强度模式的目标2b;以及目标3， 结合临床和MRI数据开发AI生物标志物，并与NRN评分进行比较。通过/否- R33的Go标准是在我们的N=430中，如果至少一种生物标志物（2a、2b或3）优于NRN MRI评分 队列（p<0.05; AUC的DeLong检验）。R33 Aim 4是为了进一步评估新系统的准确性和可靠性， 队列（N=231）。数据：公开发布的数据和AI软件。影响：大脑MRI和临床AI- 强大的新生儿预后生物标志物可以加速未来全球HIE试验中的治疗创新。

Abstract Hypoxic Ischemic Encephalopathy (HIE) is a brain injury occurring in ~5/1000 newborns. In 2005, the NIH Neonatal Research Network (NRN) established therapeutic hypothermia (TH), cooling patients in the first 6 postnatal hours to 33-34°C for 72 hours, as the standard treatment for HIE in high-income countries. However, many patients still experience adverse outcomes (death or cognitive Bayley Scales of Infant Development <85) by 18-22 months. Thus, from 2008 to 2015, the NRN tested if deeper, longer, or later TH further reduced adverse outcomes, with two trials in 21 sites. Unfortunately, results were inconclusive and further progress has been slow, largely because adverse outcomes cannot be reliably assessed until 18-22 months. To expedite therapeutic innovations and assess the impact of novel therapies in a more timely manner, there is an urgent but unmet need to establish a neonatal biomarker of 18-22 month adverse outcomes. To address this gap, the NRN developed such a biomarker using neuroradiological expert scoring of brain injury on clinically acquired neonatal brain magnetic resonance images (MRIs), known as the NRN MRI score. In one dataset with one reader, sensitivity/specificity for adverse outcomes was 81%/78%. However, in another dataset with two readers, the inter-reader agreement was only moderate and specificity for adverse outcomes was only 56-69%. Questions arise for whether this subjective and time-consuming scoring system is reliable or fully characterizes complex HIE injury patterns. Also in many countries, there are no experts available to perform MRI scoring. Finally, important clinical data elements such as birth weight, sex, APGAR scores, socioeconomic status, and aspects of the clinical exam are not fully integrated into the scoring system. Our overall hypothesis is that Artificial Intelligence (AI) algorithms on neonatal brain MRI and clinical data elements can provide higher sensitivity and specificity than the expert NRN MRI scores in predicting adverse HIE outcomes by 18-22 months. Our R61 Aims are as follows: Aim 1, Compile a large HIE dataset (N=430) from two completed NRN multi-site HIE trials; Aim 2, Develop an AI biomarker of outcome using neonatal brain MRI, and compare with NRN scores with Aim 2a focusing on MRI injury patterns and Aim 2b focusing on whole brain MRI signal intensity patterns; and Aim 3, Develop an AI biomarker of outcome combining clinical and MRI data, and compare with NRN scores. Go/No- Go criteria for the R33 is if at least one biomarker (2a, 2b, or 3) outperforms NRN MRI scores in our N=430 cohort (p<0.05; DeLong Test of AUC). The R33 Aim 4 is to further evaluate accuracy and reliability in a new cohort (N=231). Deliverables: Publicly released data and the AI software. Impact: A brain MRI and clinical AI- powered neonatal prognostic biomarker could expedite therapeutic innovations in future HIE trials worldwide.