孤独症是一种通常发病于发育早期并影响认知、社会情感、感觉运动、社会交往的神经系统疾病，其三大核心症状包括社会交往障碍、言语发展障碍、兴趣狭窄及刻板行为。病因可能包括遗传因素、环境影响和基因与环境的交互作用。目前，美国疾病控制与预防中心指出，孤独症发病率为68:1，其中男孩女孩发病比率为4:1。然而，临床上关于孤独症的诊断主要依靠行为诊断和量表评估，因此寻找潜在的客观生理指标，对于孤独症的精确诊断和有效干预具有重要意义。 本论文基于静息态脑电和眼动追踪技术，纳入较大样本的儿童数据，提取孤独症神经电生理指标和眼动注意多模态特征，为孤独症儿童脑发育评估的临床诊断和干预提供辅助评价指标。主要研究发现如下： (1) 正常儿童脑发育的静息态脑电研究：研究纳入年龄范围为3到9岁的253名正常发育儿童，探索正常儿童脑电神经振荡发育电生理指标。采用功率谱和多尺度熵分析等方法寻找正常儿童脑电神经振荡发育电生理指标。研究发现慢波频段(delta和theta频段)相对功率随年龄增长而显著下降，在快波频段（beta1和beta2）频段随年龄增长而增长，并且在alpha频段峰值随年龄增长而增长；在脑电信号复杂度的分析中，发现多个尺度脑电熵值随年龄增加。研究表明脑电相对功率和多尺度熵值有望成为辅助评估正常儿童脑发育水平的电生理指标，也为下一步孤独症儿童的发育提供正常发育参考。 (2) 孤独症儿童脑发育的静息态脑电研究：研究纳入186名3到11岁儿童，其中包括80名孤独症儿童和性别、年龄匹配的正常儿童106名，对比分析孤独症儿童的神经发育异常电生理指标。研究发现：孤独症儿童alpha频段相对功率显著低于正常组，其中差异分布主要在中央区；慢波频段(delta和theta频段)显著高于正常组儿童。在脑电信号复杂度分析中，发现学龄前期儿童脑电信号复杂度在多个尺度和多个脑区差异显著高于同龄正常儿童；在脑网络分析中，孤独症儿童聚类系数降低和平均最短路径长度升高，表明孤独症儿童大脑网络的信息传输能力异常。研究表明，功率谱、复杂度和脑网络分析能从不同的角度量化孤独症与正常儿童的差异，可以作为孤独症症状评估的潜在电生理指标。 (3) 孤独症儿童社会注意的眼动追踪研究：研究采用眼动追踪技术探索孤独症儿童在静态的面孔加工和动态的社交场景中的视觉注意加工特性；实验1中纳入177名儿童参与眼动数据采集，其中包括孤独症儿童82名和正常儿童95名。研究孤独症患者观看陌生面孔与熟悉面孔注视模式的差异，也证明了孤独症儿童在社会情感沟通方面的差异性，研究发现孤独症儿童总体注视时间无显著差异，而对于兴趣区域的访问次数要高于正常组。实验2中纳入293名儿童参与眼动数据采集，其中包括孤独症儿童104名和正常儿童189名。在动态互动社会性场景中，发现孤独症组儿童在社交注意场景中对共同注意存在缺失。通过量化孤独症注视行为异常的客观指标，为揭示孤独症儿童社交障碍提供数据支撑，并为孤独症儿童的诊断提供潜在的客观行为指标。 (4) 融合多模态特征的孤独症儿童辅助诊断研究：融合静息态脑电和眼动追踪技术注意模式的多模态特征，构建了基于深度学习的孤独症诊断模型，包括脑电特征学习网络、眼动特征学习网络和多模态特征融合网络。首先通过特征学习网络分别对脑电和眼动单一模态特征进行非线性处理和变换得到抽象的高阶脑电特征和高阶眼动特征，再通过深度网络实现对多模态特征的融合，获取区分能力更强的低维、高阶、抽象特征，用于识别孤独症儿童和正常儿童。通过对90名被试样本(包括40名孤独症儿童和50名正常儿童)进行训练建模和性能测试，相比于其他诊断模型，其总体分类正确率达到95.56%，表现出较好的诊断性能。 综上所述，本研究从脑电神经振荡节律、振荡同步、图论复杂网络分析和眼动注视特征等多层次多模态出发，寻找潜在的客观生理指标和行为指标，并基于深度学习构建孤独症辅助诊断模型，对于孤独症的临床诊断和早期有效干预提供重要参考。

Autism spectrum disorders (ASD) are neurodevelopmental disorders clinically defined by three core symptoms including impaired social interaction and communication, restricted interest and behaviors, and repetitive stereotypical behaviors. ASD usually begins at an early stage, and affects cognitive ability, social emotion, sensory and motor functioning and social interaction. The causes of autism mainly include genetic factors, environmental impact and the interaction of genes and environment. The prevalence of ASD was 68:1 in United States according to Disease Control and Prevention’s National Center, and the rate ratio of boys and girls reaches to 4:1. However, clinical diagnosis of ASD mainly relies on abnormal behavior diagnosis and scales assessment. Therefore, it is of great importance to find potential objective biomarkers for accurate diagnosis and effective intervention for autism. In this study, we collected a larger number of sample data of children from early childhood to late childhood. The main aim is to extract multimodal biomarkers based on EEG and eye-tracking data to provide the aided evaluation indices for clinical diagnosis and intervention of autism. The main findings of this study are as follows: (1) Study on resting EEG data of typical developing children from early childhood to late childhood.The study was to explore EEG biomarkers of typical development based on EEG data from 253 normal children aged ranging from 3 to 9 years. Power spectrum and multiscale entropy analysis was applied to explore the neuro-oscillatory and complexity of resting-state EEG signals. The results demonstrate that the relative power of slow wave bands (delta and theta bands) decreases significantly with age, and the fast wave bands (beta1 and beta2 bands) band grows with age and a shifting pattern of the peak frequency of the alpha band towards higher frequency range with age. In terms of EEG complexity analysis, the entropy values at multiple scales increase with age. (2) Study on resting EEG data of autism children from early childhood to late childhood. The study was to explore the abnormal EEG biomarkers based on resting EEG data of 186 children aged 3 to 11 years, including 80 children with autism and 106 gender- and age-matched typically developing (TD) children, which is a part of the sample of brain development study. The results demonstrated that the relative power of children with autism in the alpha band was significantly lower than TD children. In the slow-wave frequency bands (delta and theta frequency bands), ASD children were significantly higher than TD children. In the analysis of complexity, we found that the complexity of EEG signals in preschool autism children was significantly higher at multiple scales and at multiple brain regions than TD children. In the brain network analysis, a decrease in the clustering coefficient of autistic children and an increase in the mean shortest path length both indicate an abnormality in the information transmission ability of the autism brain network. (3) Atypical social attention changes of autism children based on eye tracking data. The study was to examine the visual attention processing characteristics of autistic children in static facial processing and dynamic social scenes using eye-tracking technology. In the first experiment, eye-tracking data were collected from 177 children, including 82 autistic children and 95 TD children to study the abnormal gaze pattern of watching strange faces and familiar face. The results demonstrated that only the visit count among different areas of interest in autistic children was significantly higher than TD children. In the second experiment, data were collected from 293 children, including 104 autistic children and 189 TD children. The results showed that there is a lack of common attention in the autism group children in the social attention scene. (4) Multimodal fusion of EEG and eye tracking data for diagnosis of autism. A deep learning-based diagnosis model for ASD was built on fusion of multimodal features including resting-state EEG features and eye-tracking fixation features. The proposed diagnosis model consists of three networks: an EEG feature learning network, an eye-traking feature learning network and a multimodal feature fusion network. The first two networks are used to learn high-level and abstract representations through multilayer nonlinear transformations from raw EEG and eye-traking features, respectively. Then the learned abstract representations are combined and fed to the third fusion network to produce the low-dimensional, high-level and abstract features with stronger discriminativity ability. Totally 90 subjects (including 40 ASD children and 50 TD ones) are used for model training and performance evaluation. The results demonstrate that compared with other considered diagnosis models, our proposed diagnosis model achieved higher classification accuracy of 95.56%. To sum up, this study aims to explore the potential objective biomarkers and behavior markers from multimodal and multilevel perspectives including neural oscillation rhythm and synchronization and graph-based complex network analysis of EEG and statistical features of eye-traking fixation patterns. Futhermore, these useful features are used to build an autism aided diagnosis system based on newly emergering deep learning methods, which can provide important references for clinical diagnosis and ealy intervention of ASD.