STTR Phase I: Real-time Automatic Analysis of Electroencephalograms in an Intensive Care Environment Using Deep Learning

这项小型企业技术转让第一阶段项目的更广泛影响/商业潜力是使重症监护病房的实时癫痫检测成为可能。特别是在那些没有全天候神经科医生覆盖的医院里，无法及时解释扫描结果。在ICU环境中对关键脑电图事件进行高性能实时检测将增加重症监护中脑监测的使用，从而改善患者预后，提高医疗保健效率，减少护理人员的认知负担。目前的自动检测方法存在高得令人无法接受的误报率，使护理提供者不堪重负，并且在这种环境中使用有限。一项可靠的服务将为美国4000多家社区医院的877,500名神经系统受损的重症患者提供高质量的护理。ICU实时发作检测的市场机会每年约为8000万美元。拟议的项目将开发一种脑电图分析辅助技术，以支持临床医生在ICU环境中评估重要医学事件的脑电图信号。脑电图信号的分析需要训练有素的神经科医生，而且既耗时又昂贵，因为识别罕见的临床事件需要分析长数据流。大多数社区医院没有训练有素的神经科医生提供全天候服务，也不能提供连续的脑电图监测，以检测神经系统受损患者的非惊厥发作。通过自动扫描脑电图信号和标记需要临床医生进一步检查的信号片段，可靠的自动检测改善了患者对长期大脑监测的访问。该工具将需要人工审查的数据量减少了两个数量级，在临床环境中提供了大量的生产力提高。该项目将利用一种创新的方法来集成隐藏马尔可夫模型、深度学习和主动学习，从而从最少量的手动注释数据中快速开发高性能机器学习系统。由此产生的自动分析将达到95%的检测准确率，检出率为每8小时1次误报。

The broader impact / commercial potential of this Small Business Technology Transfer Phase I project is enabling real-time seizure detection in intensive care units ? especially units at hospitals without 24/7 neurologist coverage to interpret scans in a timely manner. High performance real-time detection of critical EEG events in an ICU setting will increase the use of brain monitoring in critical care, thereby improving patient outcomes, increasing the efficiency of healthcare and decreasing the cognitive burden placed on caregivers. Current approaches to automatic detection suffer from unacceptably high false alarm rates that overwhelm care providers, and are of limited use in this environment. A reliable service would expand access to quality care for 877,500 neurologically compromised critical care patients in 4,000+ community hospitals in the United States. The market opportunity for real-time seizure detection in the ICU is approximately $80M per year. The proposed project will develop an assistive technology for EEG analysis to support clinicians in evaluating EEG signals for medically important events in an ICU environment. Analysis of EEG signals requires a highly trained neurologist, and is time consuming and expensive since identifying rare clinical events requires analysis of long data streams. Most community hospitals do not have 24/7 access to trained neurologists and can not provide continuous EEG monitoring to detect non-convulsive seizures in neurologically compromised patients. Reliable automatic detection improves patient access to long-term brain monitoring by auto-scanning EEG signals and flagging sections of the signal that need further review by a clinician. The tool reduces the amount of data needing manual review by two orders of magnitude, offering substantial productivity gains in a clinical setting. The project will leverage an innovative approach for integrating hidden Markov models, deep learning and active learning to allow the rapid development of a high performance machine learning system from minimal amounts of manually annotated data. The resulting automatic analysis will achieve 95% detection accuracy for seizures with a false alarm rate of 1 per 8-hour period.