随着自闭症发病率的逐年提高，全国已有超过千万自闭症患者。自闭症的早期诊断有助于早期干预治疗，帮助患者学习生存技能，更好融入社会。眼动追踪技术由于其良好的安全性、便捷性、可拓展性为自闭症相关研究提供了强大有效的帮助。基于此，本文分析了多场景下自闭症儿童的眼动指标，结合机器学习算法实现了多场景的模型融合。本文的工作内容包括以下几点：本文基于瞳孔适应性响应与静态正脸图片，分析了自闭症儿童异常的瞳孔响应和注视行为，结合多种机器学习算法对比了多种模型融合方式，构建了融合瞳孔响应与注视行为的自闭症分类系统。使用了无偏的嵌套式交叉验证，在融合模型上得到了平均84.9%、最高93.3%的分类正确率和平均65.15%、最高84.9%的马修斯相关系数（Matthews-correlation-coefficient，MCC）。本文从分类角度分析对比了瞳孔适应性响应与瞳孔对光反射范式在眼动指标上的异同性。结果显示，瞳孔适应性响应相对于瞳孔对光反射在特征的数目和显著性上具有一定优势，但是从分类的角度来看并没有提供互补性的信息。瞳孔对光反射范式在一定程度上具有更高的被试收集率。本文提出了具有更高分类性能的动态干扰-自然活动范式，从分类角度分析对比了活动-社交场景、自我社交场景、动态干扰-自然活动场景三种注视行为范式的异同点。使用注视率特征，摆脱了服从性测试和注意力缺陷的干扰，构建了基于瞳孔响应与视频范式的自闭症分类系统。结果显示，自闭症在多场景下均具有对人的活动、社交性信息、人脸信息的显著不敏感性，其指标上存在显著的冗余性与相关性，但也存在一定的互补性。在分类上，动态干扰-自然活动场景得到了平均86.6%、最高93.6%的分类正确率和平均73.3%、最高87.7%的MCC。最后在融合模型中，本文得到了平均88.5%、最高91.5%的分类正确率和平均76.4%、最高83.1%的MCC。本文结合量表指标对模型进行一定解释。结果显示，本次研究选择的眼动指标在自闭症组合和正常组中具有显著性相关，将注视行为与自闭症核心症状损伤的量化建立联系。同时，组内分析的结果表明自闭症组较差的生活能力、学习能力、团体适应能力等量表指标与其对社交场景的排斥和对非人物体的过分关注息息相关。

As the incidence of autism has been increasing year by year, there are now over tens of millions of autistic patients nationwide. Early diagnosis of autism can help with early intervention and treatment, helping patients learn survival skills and better integrate into society. Eye-tracking technology provides powerful assistance for autism-related research due to its good safety, convenience, and scalability. Based on this, this paper analyzes the eye movement indicators of autistic children in multiple scenarios and combines machine learning algorithms to achieve model fusion in multiple scenarios. The work of this paper includes the following points: This paper is based on pupillary adaptation response and static face images to analyze the abnormal pupil response and gaze behavior of autistic children. It compares various model fusion methods with multiple machine learning algorithms and constructs an autistic classification system that fuses pupillary response and gaze behavior. Unbiased nested cross-validation is used to obtain an average classification accuracy rate of 84.9%, a highest accuracy rate of 93.3%, an average Matthews-correlation-coefficient (MCC) of 65.15%, and a highest MCC of 84.9% on the fused model. This paper analyzes and compares the differences and similarities of eye movement indicators between pupillary adaptation response and pupillary light reflex paradigms from the classification perspective. The results show that pupillary adaptation response has certain advantages over pupillary light reflex in terms of the number and significance of features but does not provide complementary information from the classification perspective. Pupillary light reflex paradigm has a higher subject collection rate to some extent. This paper proposes a dynamic interference-natural activity paradigm with higher classification performance and analyzes the differences and similarities of three types of gaze behavior patterns: activity-social setting, self-social setting, and dynamic interference-natural activity setting. By using gaze rate features, interference from compliance testing and attention deficit is eliminated, and an autistic classification system based on pupillary response and video paradigm is constructed. The results show that autism is significantly insensitive to human activity, social information, and facial Abstract III information in multiple scenarios, and there is significant redundancy and correlation between its indicators, but also some complementarity. In terms of classification, the dynamic interference-natural activity scenario obtained an average classification accuracy rate of 86.6%, a highest accuracy rate of 93.6%, an average MCC of 73.3%, and a highest MCC of 87.7%. On the fused model, this paper obtained an average classification accuracy rate of 88.5%, a highest accuracy rate of 91.5%, an average MCC of 76.4%, and a highest MCC of 83.1%. Finally, this paper combines scale indicators to explain the model to some extent. The results show that the eye movement indicators selected in this study have a significant correlation between the autism group and the normal group, establishing a connection between gaze behavior and the quantification of core symptoms of autism damage. At the same time, the results of within-group analysis show that poor life skills, learning ability, and group adaptation ability in the autism group are closely related to their rejection of social settings and excessive attention to non-human objects.