毕业生知识库

阿尔兹海默症(Alzheimer Disease, AD)是老年期痴呆最常见的类型，约占老年期痴呆的60%-70%，目前临床上尚未找到能直接治疗AD的有效药物。轻度认知功能障碍(Mild Cognitive Impairment, MCI)被认为是介于老年人正常衰老和AD之间的一种中间过渡状态，每年约有10%-15%的MCI患者会转化成AD患者，这一转化率远高于正常老年人群。因此，对MCI患者的早期诊断和及时干预具有极其重要的意义。但是，认知功能障碍需要通过多种检测手段进行临床诊断，这些检测不仅设备成本高昂，而且很大程度上依赖于医生的经验和主观判断。因此，亟需寻找一种客观、有效、便携且经济的诊断手段。随着生理信号采集与分析技术的发展，越来越多的研究发现脑电(Electroencephalogram, EEG)、近红外光谱(functional Near-Infrared Spectroscopy, fNIRS)等生理信号可以有效地反映大脑的认知功能，通过使用机器学习等算法对采集的数据集进行比较和分析，可以对老年认知功能障碍进行分类和自动检测。然而这些研究存在一些未解决的问题：（1）样本数量少，模型容易过拟合，导致实际应用时泛化性差；（2）不同个体之间的生理信号存在显著的差异，难以提取出能有效区分不同认知障碍程度的高敏感特征；（3）使用多模态数据融合诊断老年认知功能障碍的研究较少，未能充分利用不同模态数据的互补性优势。
本文在国家重点研发计划“老年认知障碍多模态评估与智能康复系统研发”(2018YFC2001700)的支持下，针对认知功能障碍临床诊断依赖医生经验、检查成本高的问题，开展了基于多模态生理信号特征优选的认知功能障碍早期智能诊断研究。论文的主要工作内容和创新点如下：
1.针对静息态EEG信号识别MCI准确率不高的问题，本研究提出了一种基于事件相关电位 (Event-Related Potentials, ERP)高敏感特征优选的MCI诊断方法。首先，设计了延迟匹配任务范式，建立了一个包含71名被试者的EEG数据集。然后，对不同组别的EEG信号进行频域分析和时频分析，并通过ERP波形的组间比较，提取出了最能体现MCI组与正常组差异的高敏感特征。最后，提取了功率谱密度、微分熵、连续小波变化系数特征作为对比，讨论了不同特征、不同电极以及不同分类器对识别MCI患者准确率的影响。结果表明，相比于正常老年人，MCI患者在β频带的脑电活动减弱，而在α频带的脑电活动增强，且这种差异主要表现在额叶和顶叶区域。相较于其他特征，ERP特征在区分MCI患者与正常老年人方面展现出更优的性能，具有更低的过拟合风险，并与频域特征有良好的互补性。选取合适的特征组合输入模型后，MCI患者的识别准确率达到了84.74%。
2.针对目前fNIRS信号在诊断认知功能障碍研究中存在的样本量不足的问题，本研究建立了两个fNIRS数据集：一个利用Stroop任务范式采集了140名被试者的fNIRS数据，另一个利用延迟匹配任务范式采集了78名被试者的fNIRS数据。基于上述数据集，提出了一种基于fNIRS空间信息和孪生网络的MCI诊断方法。首先，对MCI组和正常人组的血流动力学响应进行分析，提取差异性较大的特征进行分类，比较了不同的脑区对分类结果的影响。然后，提出了一种特征子图的构建方法，以保留并提取光极帽的空间信息，并通过随机森林对提取的fNIRS特征进行优选。使用孪生网络将分类问题转化成匹配问题，以解决因样本数量少导致训练不充分的问题。结果表明，氧合血红蛋白浓度可以为诊断MCI提供依据，本文提出的方法相较于传统的机器学习算法，在两种范式下MCI患者的识别准确率分别提升了12.54%和5.47%。
3.针对生理信号个体差异大、使用单一模态数据诊断AD患者和MCI患者效果不理想的问题，本研究利用所采集的多模态数据（EEG、眼动和行为学数据），提出了一种基于领域对抗神经网络(Domain-Adversarial Neural Network, DANN)的多模态融合方法来提高分类准确率。首先，基于ERP波形的叠加原理扩充了样本数量，设计了ERPNet特征提取模块，并与其他模态信号在特征层融合，利用DANN减小不同个体之间的特征差异，以更好地识别MCI患者。然后，比较了基于DANN的特征层融合、多核支持向量机和决策层多种融合方法对分类器性能的影响。最后，根据不同模态数据采集的难易程度设计了一套分阶段的痴呆诊断流程。结果表明，基于DANN的特征层融合方法取得了最佳性能，并减轻过拟合的现象，与单模态的EEG信号相比，提升了10.77%的准确率。

Alzheimer's disease (AD) is the most common type of Alzheimer's disease, accounting for approximately 60%-70% of all cases of dementia. Currently, no effective clinical drug has been found to directly treat it. Mild Cognitive Impairment (MCI) is considered an intermediate transitional state between normal aging in the elderly and AD. About 10%-15% of MCI patients will convert to AD patients per year, which is a rate significantly higher than that of the general elderly population. Therefore, early diagnosis and timely treatment of MCI patients are of extremely important significance. However, clinical diagnosis of cognitive impairment requires various examinations, which require expensive equipment and rely on the experience and subjective judgment of doctors. Hence, there is an urgent need to find an objective, effective, portable, and affordable diagnostic method. With the development of physiological signal collection technology, an increasing number of studies have found that physiological signals such as Electroencephalogram (EEG) and functional Near-Infrared Spectroscopy (fNIRS) can effectively reflect the cognitive functions of the human brain. By using computer-aided technologies such as machine learning algorithms, cognitive impairment in the elderly can be classified and automatically detected. However, there are some unresolved issues in the existing studies: (1) The small sample size can lead to overfitting in the diagnostic model, resulting in poor generalization when applied in practice; (2) There is a large individual difference in physiological signals, making it difficult to extract effective features to distinguish different levels of cognitive impairment; (3) There are few studies using multi-modal data fusion to diagnose cognitive impairment in the elderly, and they fail to fully utilize the complementary advantages of different modal data. Supported by the National Key R\&D Program ``Research and Development of Multi-modal Assessment and Intelligent Rehabilitation System for Elderly Cognitive Impairment'' (2018YFC2001700), this thesis aims at the problem that clinical diagnosis of cognitive impairment relies on doctor experience and high examination costs, conducts research on early intelligent diagnosis of cognitive impairment based on multi-modal feature optimization. The main contributions and innovations of this research are as follows:
1. To address the issue of low accuracy in diagnosing MCI patients from resting-state EEG signals, this study proposed an MCI diagnosis method based on the selection of highly sensitive features from Event-Related Potentials (ERP). Firstly, a delayed match-to-sample task paradigm was designed, creating an EEG dataset comprising 71 participants. Then, frequency domain analysis and time-frequency analysis were performed on the EEG signals of different groups, and through inter-group comparison of ERP waveforms, the highly sensitive features that best reflected the differences between the MCI group and the normal group were extracted. Finally, the power spectral density, differential entropy, and continuous wavelet transform coefficient features were extracted for comparison, and the impact of different features, different channels, and different classifiers on the accuracy of detecting MCI patients was explored. The results show that compared with healthy elderly, MCI patients have weakened brain electrical activity in the β band, while enhanced brain electrical activity in the α band, and this difference is mainly manifested in the frontal and parietal areas. Compared to other features, ERP features demonstrated superior performance in distinguishing MCI patients from healthy elderly, offering a lower risk of overfitting and complementing frequency domain features well. The accuracy of MCI recognition reached 84.74% after combining multiple appropriate features into the model.
2. In view of the problem of insufficient sample size in current studies of fNIRS signals in diagnosing cognitive impairment, this study established two fNIRS data sets: one collected fNIRS data from 140 participants through the Stroop task paradigm, and the other collected data from 78 participants via a delayed match-to-sample task paradigm. Based on these datasets, a novel diagnostic method for MCI using fNIRS spatial information and Siamese networks was proposed. Firstly, the hemodynamic response of the MCI group and the normal group was analyzed, features with large differences were extracted for classification, and the effects of different brain regions on the classification results were compared. Then, a method of constructing fNIRS feature map was proposed to preserve and extract the spatial information of the optode cap, and the extracted fNIRS features are selected through random forests. The Siamese network is used to transform the classification problem into a matching problem to solve the problem of insufficient training due to the limited sample size. The results show that oxyhemoglobin concentration can provide a basis for diagnosing MCI. Compared with the traditional machine learning algorithm, the method proposed in this study improved the identification accuracy of MCI patients by 12.54% and 5.47% respectively under the two paradigms.
3. Aiming at the problem of large individual differences in physiological signals and the unsatisfactory effect of using unimodal data to diagnose AD and MCI patients, this study used collected multi-modal data (EEG, eye movement, and behavioral data) to propose a multi-modal fusion method based on Domain-Adversarial Neural Networks (DANN) to improve the classification accuracy. Firstly, the sample size was expanded based on the superposition principle of ERP waveforms, and an ERPNet feature extraction module was designed for feature-level fusion with other modal signals. DANN was used to reduce the feature differences between different individuals to identify MCI patients better. Then, the performances of DANN-based feature layer fusion, multi-kernel support vector machine, and decision-making layer fusion methods were compared. Finally, a phased dementia diagnosis process was designed according to the difficulty of data collection in different modalities. The results show that the feature layer fusion method based on DANN achieves the best performance and reduces the over-fitting phenomenon. Compared with single-modal EEG signals, the use of multi-modal data improves the accuracy by 10.77%.