孤独症(autism spectrum disorder, ASD)是一类广泛存在的神经发育谱系障碍，主要表现为社会互动和社会交往障碍。近年来，世界范围内患有孤独症的人数一直加速递增，中国孤独症儿童数量已经超过1000万，为社会和家庭带来巨大的悲痛和损失。针对孤独症，虽然目前还没有能够彻底根治的疗法与特效药，但尽早确诊和干预可以有效的降低孤独症给患者带来的生活上的影响。因此提高孤独症诊断的准确性有着重要意义。

目前孤独症的临床诊断的主要方法有行为量表和结构化访谈等两种形式，诊断结果在一定程度上依赖于临床医生的经验，因此常常伴随着较高的误诊率。静息态EEG作为一种有效的脑科学研究工具，可以为孤独症的诊断提供生理方面的重要参考，可作为辅助手段提高诊断的准确率。低密度EEG相对于其他脑成像技术具有操作便捷、价格低廉、环境适应力强等优点，更加适合儿童研究及应用。因此，开发一套基于低密度EEG的儿童孤独症辅助诊断系统，具有重大的现实意义。

为满足以上需要，本文主要进行了以下三项工作，其一是低密度EEG伪迹去除算法设计，其二是孤独症儿童EEG数据的特征识别，其三是儿童孤独症辅助诊断系统的开发。

由于EEG数据信噪比较低，因此数据预处理十分重要，本研究采集了受到不同类型伪迹污染的EEG数据以及相应的肌电参考数据。基于这些数据，设计了针对不同生理和非生理伪迹的识别和去除算法，并验证了算法的稳定性和有效性。

去除伪迹后，对所得数据进行特征提取，提取了固定频带功率、小波熵、通道间互信息与相干，发现孤独症儿童与正常儿童在多个特征上存在着显著差异。对这些特征进行选择并基于此训练分类模型，比较不同分类模型的效果，发现采用线性核支持向量机的方法可以在较低算法复杂度下，获得相对较好的分类结果（分类准确率78.7%，ROC曲线下面积0.84）。进一步地，本研究使用SHAP算法对模型特征权重进行了排序，并对各特征对分类结果的影响模式进行了分析，发现分类模型中特征权重分布与统计检验结果相一致，证明了分类模型的有效性和可靠性。最后，开发了脑电数据管理系统和脑电辅助诊断报告系统，以方便脑电数据的管理和辅助诊断结果的呈现。

综上，为满足儿童孤独症临床诊断上的实际需求，本文基于低密度EEG，设计了数据预处理与分类算法，实现了儿童孤独症的辅助诊断，开发了集成上述功能的一整套软件系统，为儿童孤独症的辅助诊断提供了有效的方法和工具。

This thesis mainly carried out the following three tasks. One is the design of low-density EEG artifact removal algorithm, the second is the feature recognition of EEG data for children with autism, and the third is the development of an auxiliary diagnosis system for children with autism. 

The signal-to-noise ratio of EEG data is relatively low, data preprocessing is very important. In this study, EEG data contaminated by different types of artifacts and corresponding EMG reference data were collected simultaneously. Based on these data, the identification and removal algorithms for different physiological and non-physiological artifacts are designed, and the stability and effectiveness of the algorithm are verified.

After removing the artifacts, feature extraction is performed on the obtained data, include the fixed frequency band power, wavelet entropy, mutual information and coherence between channels,and it is found that there are significant differences between autistic children and normal children in many features. These features are selected and the classification model is trained based on this, and the effects of different classification models are compared. It is found that the method of using linear kernel support vector machines can obtain relatively good classification results with lower algorithm complexity (classification accuracy rate is 78.7% , The area under the ROC curve is 0.84). Furthermore, this study uses the SHAP algorithm to rank the weights of features, and analyzes the influence mode of each feature on the classification results. It is found that the feature weight distribution in the classification model is consistent with the statistical test results, which proves the effectiveness of the classification model. Sex and reliability. Finally, an EEG data management system and an EEG-assisted diagnosis report system were developed to facilitate the management of EEG data and the presentation of auxiliary diagnosis results. 

In summary, in order to meet the actual needs for clinical diagnosis of childhood autism, this paper designs data preprocessing and classification algorithms based on low-density EEG to realize the auxiliary diagnosis of childhood autism, and develops a complete set of software systems that integrate the above functions. It provides effective methods and tools for the auxiliary diagnosis of autism in children.