高复杂性脑电数据的数据挖掘及其在癫痫性发作自动检测中的应用

Continuous Electroencephalography monitoring in the Intensive Care Unit environment (ICU-EEG) has become an essential component of top-tier medical care for patients with acute neurological injuries. ICU-EEG data presents a special challenge for computational analysis because of its high complexity, large scale, massive variability, imbalance, and the large number and varied character of both external and physiological background noise that can contaminate signals acquired in the ICU environment. Despite the urgent medical demand for automated analysis of ICU-EEG data, very little research has been done to bring the existing data mining to ICU-EEG data, and to develop the novel theoretical advances that will be needed to cope with the unique challenges posed by mining this data. This project focuses on developing fundamental theory and methods of data mining for such highly complicated EEG data systematically, with a focus on automated detection of epileptic seizures. Firstly, a multi-element feature extraction method will be developed, where 2-D features in frequency domain, linear and nonlinear features in time domain, and the features based on the location of seizure origin and the spatial pattern of seizure spread throughout the brain over time are all extracted and combined. Secondly, a semi-supervised domain transferred (SSDT) learning model will be constructed on the basis of L_0.5 sparse regularization and manifold regularization. Fast algorithms for solving the SSDT model will be designed and the convergence of the algorithms will be analyzed. Thirdly, an ensemble classification system based on boosting and cascaded classifier methods from computer vision will be adapted to epileptic seizure detection in ICU-EEG data. Finally, the application of the automatic epileptic seizure detection will be tested, refined, and validated on a large set of ICU-EEG recordings to verify the feasibility and efficiency of the above constructed system. The project results will provide significant theoretical advances, methods and a system for extracting the essential data features, clarifying the appropriate learning framework for highly complicated biomedical data, and will yield a much needed practical solution to the problem of epileptic seizure detection which can be used to assist physicians in caring for critically ill patients.

面向重症监护癫痫病人的脑电图是一类具有超大规模、多变、非平衡、含复杂噪声的高复杂性脑电数据，有关此类数据的数据挖掘是医学临床诊断的急需,但目前几乎还是空白。本项目拟以癫痫性发作自动检测为背景，聚焦对高复杂性脑电数据数据挖掘的基本理论与方法展开系统研究。拟提出包含二维表示的频域特征、融合线性与非线性的时域特征、以及刻画发作起始位置及其演变过程的空间模式特征的特征提取方法；建立基于L_0.5稀疏正则化理论与流行正则化理论的异域半监督转移学习模型，设计快速求解算法并对其收敛性进行系统分析；构造基于boosting思想及计算机视觉的完整、可解释的综合分类集成系统；以癫痫性发作的自动检测为应用来验证所研制系统的可用性与有效性。项目成果将为高复杂性脑电数据的本质特征提取、高复杂性脑电数据的学习模式挖掘、实用癫痫性发作自动检测系统研发等提供重要的理论、方法与系统基础,亦可直接用于相关医疗诊断的辅助工具。

面向重症监护病人的脑电图是一类具有超大规模、多变、非平衡、含复杂噪声的高复杂性脑电数据，有关此类数据的数据挖掘是医学临床诊断的急需,但目前儿乎还是空白。源于这一现状，本课题以癫痫性发作自动检测为背景，研究了高复杂性脑电数据的综合特征提取方法。分别从时频域分析、非线性动力学分析等不同角度设计了多种新的脑电特征提取方法，并提出了融合癫痫脑电特征，对脑电信号在癫痫发作时的本质变化进行了多方位刻画。研究了混合结构的统一学习模型、算法与理论。构建了相关的神经计算模型并分析其动力学性质，展开了复杂脑电的数据平衡方法以及神经计算模型参数辨识算法的研究，并提出了一种新的神经生理系统复杂性分析框架。研究了综合分类集成系统的实现与关键技术。通过对已有分类技术的综合比较，提出了一种基于超限学习机的分类系统；并结合logistic回归，超限学习机，支撑向量机、决策树，提出了一种基于多混合分类器的综合分类集成系统。研究了所提技术在重症监护病人癫痫性发作自动检测中的应用。建立了典型异常放电模式的脑电数据库以及高复杂性脑电数据库，提出了复杂脑电的伪迹去除方法，并在基于模型和数据共同驱动的癫痫性发作早期检测及发作追踪方面展开了具体的应用研究，取得令人满意的结果。项目成果将为高复杂性脑电数据的本质特征提取与学习模式挖掘、以及脑功能障碍临床辅助诊断系统的实现提供理论依据与技术基础。