Big Data and Deep Learning for the Interictal-Ictal-Injury Continuum

项目摘要/摘要：发作期-发作期-损伤连续过程的大数据和深度学习 重症监护中的大脑监测在过去20年中急剧增长，发现大量的 ICU患者有亚临床癫痫和癫痫样电事件的比例，统称为 “发作期-发作期-损伤连续体异常”(IIICAS)，只有脑电(EEG)才能发现。这 增长在重症监护领域制造了一场危机：很明显，IIICA损害大脑并导致永久性 神经性残疾。然而，专家视觉检查对IIICA的检测通常会延迟，这表明我们需要 更好的实时监测工具，以应对海量的ICU脑电数据。在其他情况下，IIICA似乎 是无害的附带现象，许多人担心对IIICA的认识的提高造成了一种流行病 过度激进的抗惊厥药物处方导致可预防的不良事件和成本。这 危机突出了对IIICA自动脑电监测的关键需求未得到满足，并更好地了解 其中哪些类型的IIICA会导致神经损伤，需要干预。 IIICA的原因很广泛，从出血性中风和颅内等原发脑损伤 出血，到全身内科疾病，如败血症和尿毒症。直到最近，这个庞大的临床 异质性一直是了解IIICA对神经学影响的不可逾越的障碍 结果。然而，深度学习方面的最新进展，加上史无前例的 我们团队在过去三年开发的海量数据集，使其首次成为可能 系统研究IIICA与神经系统预后的关系。 为了满足对更好的监测工具和更好的模式理解国际投资协定的需要，我们将采取 深度学习方法，利用海量ICU脑电数据集中尚未开发的信息。我们会 追求三个具体目标：SA1：在一套庞大的 CEEG记录，从而为训练计算机检测IIICA模式准备脑电数据；SA2：开发 有监督的DL算法可以像人类专家一样准确地检测IIICA，从而提供了强大的工具 用于IIICA研究和临床脑监测；SA3：评估IIICA对 神经学结果：我们将开发模型来量化IIICA在控制后对残疾风险的影响 用于煽动疾病和其他临床因素，并预测抑制IIICA的干预措施的效果。 这项工作将为推进精确重症监护神经学领域提供四个关键好处，并通过 扩展，我们有能力在危重疾病期间为患者提供最佳的神经护理。1)改进 对类似IIICA状态的癫痫的临床意义的理解；2)开发强大的工具和 用于重症监护脑遥测的算法；3)独特的、海量的、公开可用的、彻底注释的 数据集，使其他研究人员能够进一步推进该领域；以及4)可测试的模型，预测 哪些类型的cEEG异常需要积极治疗，为介入试验奠定了基础。

Project Summary/Abstract: Big Data and Deep Learning for the Interictal-Ictal-Injury Continuum Brain monitoring in critical care has grown dramatically over the past 20 years with the discovery that a large proportion of ICU patients suffer from subclinical seizures and seizure-like electrical events, collectively called “ictal-interictal-injury continuum abnormalities” (IIICAs), detectable only by electroencephalography (EEG). This growth has created a crisis in critical care: It is clear that IIICAs damage the brain and cause permanent neurologic disability. Yet detection of IIICAs by expert visual review is often delayed suggesting we need better tools for real-time monitoring, to cope with the deluge of ICU EEG data. In other cases, IIICAs appear to be harmless epiphenomena, and many worry that increased awareness of IIICAs has created an epidemic of overly-aggressive prescribing of anticonvulsant drugs leading to preventable adverse events and costs. This crisis highlights critical unmet needs for automated EEG monitoring for IIICAs, and a better understanding of which types of IIICAs cause neural injury and warrant intervention. Causes of IIICAs range widely, from primary brain injuries like hemorrhagic stroke and intracranial hemorrhage, to systemic medical illnesses like sepsis and uremia. Until recently, this massive clinical heterogeneity has been an insurmountable barrier to understanding the impact of IIICAs on neurologic outcome. However, recent advances in deep learning, coupled with the unprecedented availability of a massive dataset developed by our team over the last three years, makes it feasible for the first time to systematically study the relationship between IIICAs and neurologic outcomes. To meet the need for better monitoring tools and better models for understanding IIICAs, we will take a deep learning approach to leverage the as-yet untapped information in a massive ICU EEG dataset. We will pursue three Specific Aims: SA1: Comprehensively label all occurrences of IIICAs in a massive set of cEEG recordings, thus preparing the EEG data for training computers to detect IIICA patterns; SA2: Develop supervised DL algorithms to detect IIICAs as accurately as human experts, thus providing powerful tools for both research on IIICAs and for clinical brain monitoring; SA3: Estimate the effect of IIICAs on neurologic outcome: we will develop models to quantify effects of IIICAs on risk for disability after controlling for inciting illness and other clinical factors, and to predict effects of interventions to suppress IIICAs. This work will provide four crucial benefits to advance the field of precision critical care neurology, and by extension, our ability to provide optimal neurologic care for patients during critical illness. 1) Improved understanding of the clinical significance of seizure like IIICA states; 2) development of robust tools and algorithms for critical care brain telemetry; 3) a unique, massive, publicly available, thoroughly annotated dataset that will enable other researchers to further advance the field; and 4) a testable model that predicts which types of cEEG abnormalities warrant aggressive treatment, setting the stage for interventional trials.