ISSN 1671-3710
CN 11-4766/R

心理科学进展 ›› 2022, Vol. 30 ›› Issue (10): 2303-2320.doi: 10.3724/SP.J.1042.2022.02303

• 研究前沿 • 上一篇    下一篇


袁玉琢1, 骆方2()   

  1. 1中国基础教育质量监测协同创新中心
    2北京师范大学心理学部, 北京 100875
  • 收稿日期:2021-07-26 出版日期:2022-10-15 发布日期:2022-08-24
  • 通讯作者: 骆方
  • 基金资助:

Early screening and diagnosis of autism spectrum disorder assisted by artificial intelligence

YUAN Yuzhuo1, LUO Fang2()   

  1. 1Collaborative Innovation Center of Assessment toward Basic Education Quality, Beijing Normal University, Beijing 100875, China
    2Faculty of Psychology, Beijing Normal University, Beijing 100875, China
  • Received:2021-07-26 Online:2022-10-15 Published:2022-08-24
  • Contact: LUO Fang


自闭症谱系障碍(Autistic Spectrum Disorders, ASD)的症状早在婴幼儿期就会显现, 越早发现, 越早干预, 治疗效果越好。传统自闭症早期筛查与诊断在评估方法、流程上存在局限, 无法满足大规模筛查和诊断需求。随着人工智能技术的快速发展, 使用智能化方法进行自闭症早期大规模无感筛查与诊断逐渐成为可能。近10年间, 国内外对自闭症智能化识别方法的探索在经典任务行为、面部表情和情绪、眼动、脑影像、运动控制和运动模式、多模态6个领域积累了丰富的研究成果。未来研究应围绕构建国内自闭症早期智能医学筛查与诊断体系, 开发针对婴幼儿患者的筛查工具, 构建融合多模态数据的自闭症婴幼儿智能化识别模型, 建立结合脑影像技术的自闭症精细化诊断方法等方面来开展。

关键词: 自闭症谱系障碍, 自闭症早期筛查与诊断, 自闭症智能化识别, 人工智能, 多模态数据


Symptoms of Autistic Spectrum Disorders (ASD) can manifest as early as infancy, and the earlier detection and intervention can lead to better therapeutic results. The traditional screening and diagnosis of autism comes from professionals, which is highly subjective and time-consuming, leading to misdiagnosis or missing the optimal intervention time. In recent years, with the rapid development of artificial intelligence and the accumulation of clinical data on autism, intelligent recognition methods for autism and its early symptoms have developed rapidly. This paper summarizes the research on intelligent recognition of autistic infants in the past decade, and divides the research into six sub-areas based on the data types used in the research: 1) recognition based on classical task behavior data; 2) recognition based on facial expression and emotional data; 3) recognition based on eye gaze data; 4) recognition based on brain image data; 5) recognition based on motor control and movement pattern data; 6) recognition based on multimodal data. The current technology is to use non-contact vision system and sensory devices to collect infants’ behavioral data, such as facial expression, head and limb movement, eye movement, brain image. Researchers usually develop risk behavior detection algorithms and build machine learning or deep learning models for automatic recognition according to task objectives and data characteristics. At present, it can reach the precision of scale tools and manual evaluation. The current modeling trend is to use a multimodal fusion framework to build prediction models based on the complementary relationship between multimodal information, feature transformation and representation patterns of autistic infants, which is expected to further improve the recognition accuracy. In the future, researchers should focus on building an intelligent medical screening and diagnosis system for early autism, developing screening tools for infants and young children, establishing a refined diagnosis method for autism combined with brain imaging technology, and building an intelligent recognition model for autistic infants by integrating multimodal data. In addition, to carry out high-quality intelligent recognition research, a large-scale database of autism and high-risk infants and corresponding behavioral characteristic database should be established as soon as possible, and risk behaviors were marked in coarse and fine granularity according to the early behavioral diagnostic criteria of autism. Currently, in the face of the contradiction between the large demand for model training data and the lack of autistic samples, researchers can first try to use small sample learning methods, such as model fine-tuning, data augmentation, transfer learning.

Key words: autism spectrum disorder, early screening and diagnosis of autism, intelligent screening of autism, artificial intelligence, multimodal data