ISSN 1671-3710
CN 11-4766/R

心理科学进展 ›› 2022, Vol. 30 ›› Issue (10): 2321-2337.doi: 10.3724/SP.J.1042.2022.02321

• 研究前沿 • 上一篇    下一篇


侯婷婷1, 陈潇2, 孔德彭2, 邵秀筠3, 林丰勋1, 李开云1()   

  1. 1济南大学教育与心理科学学院, 济南 250022
    2浙江工业大学教育科学与技术学院, 杭州 310023
    3青岛市晨星实验学校, 青岛 266000
  • 收稿日期:2021-11-02 出版日期:2022-10-15 发布日期:2022-08-24
  • 通讯作者: 李开云
  • 基金资助:

Application of machine learning in early identification and diagnosis of autistic children

HOU Tingting1, CHEN Xiao2, KONG Depeng2, SHAO Xiujun3, LIN Fengxun1, LI Kaiyun1()   

  1. 1School of Education and Psychology, University of Jinan, Jinan 250022, China
    2School of Educational Technology and Science, Zhejiang University of Technology, Hangzhou 310023, China
    3Qingdao Chenxing Experimental School, Qingdao 266000, China
  • Received:2021-11-02 Online:2022-10-15 Published:2022-08-24
  • Contact: LI Kaiyun


早发现、早诊断、早干预是开展自闭症儿童教育康复工作的共识, 但传统识别和诊断方法局限及专业人员缺乏常导致自闭症儿童错失最佳干预期。为改善现状, 近年来机器学习凭借其客观准确、简便灵活等方面的优势, 逐渐被应用到自闭症的早期预测、筛查、诊断和评估过程管理中, 积累了较为丰富的成果。但是机器学习也在研究对象选取、分类数据采集和理论模型应用等方面存在局限性。未来研究应推动构建孕产期和新生儿病理生理信息追踪数据库和标准化模型分类指标体系, 同时继续优化算法, 加快智能化自闭症识别和诊断理论成果向实践转化。

关键词: 机器学习, 自闭症, 早期识别与诊断, 系统综述


In recent years, autism spectrum disorder has gradually become a major global public health problem due to complex etiology and large group, and its prevalence rate has been rising all over the world. A series of explorations and discussions, which focused on effective intervention, have been conducted in this field. Early detection, early diagnosis and early intervention have been determined as the consensus of autistic children education and rehabilitation. However, both the limitations of traditional identification and diagnosis methods and the lack of professionals mostly caused the miss of the well-timed intervention in autistic children. However, traditional identification and diagnosis methods often have some limitations (e.g., strong subjectivity, time-consuming) and face the situation that lack of professionals and higher requirements for professional clinical ability, which seriously affect the early intervention of autistic children and even make them miss the best opportunity of education and rehabilitation; Furthermore, which brings huge psychological and economic burden to autistic children and their families. In order to improve this situation, machine learning has made great achievement in recent years. With its advantages of objectivity, accuracy, simplicity and flexibility, machine learning has been gradually applied in the early prediction, screening, diagnosis and evaluation of autism. Previous studies were mostly conducted on the basis of traditional machine learning algorithms (e.g., support vector machines, naive bayes and random forest) or deep learning algorithms (e.g.,, convolutional neural network, artificial neural network and multilayer perceptions), and the analysis of EEG signals, behavioral data and genetic data could support for establishing predictive or classification models which contribute to predict, identify and diagnose autistic children. Compared with the traditional classification, this method could not only improve the efficiency and reduce the pressure of professionals, but also automatically extract and sort out subtle and potentially critical information from a large amount of data. Moreover, it could gain insight into the inherent laws of complex problems, build a multimodal classification model, complement each other's advantages, and improve the classification accuracy. However, machine learning has limitations in the selection of research objects, the extraction and collection of classified data, and the application of theoretical models. To solve the above problems, first of all, future research should clarify the pathology, etiology and course factors of autism in the early stage of development, and promote the construction of a tracking database of pathophysiological information of pregnant women and newborn. The data collection of the health information of pregnant women and their family could know about related genetic and environmental teratogenic factors and try to transfer "passive waiting" to "active defense" and avoid obstacles as soon as possible. Secondly, it is very important to sort out the criteria and basis for selecting classification indicators and improve the fineness of indicators. For example, the appropriate classification indicators, which adapt to autistic children come from different age groups and different disability degree groups, should be determined. Following studies should gradually detectestable, widely adaptable and higher accurate indicators. The build of a more scientific and stable standardized classification index system could promote the internal order of classification indexes. Finally, the actual needs and worries of different audiences should be carefully investigated. The problems such as sample representativeness, model applicability, algorithm accuracy and stability should be solved as well. The extension of propaganda about theoretical research achievements is needed; Meanwhile, more efforts should be taken to accelerate the transformation from theoretical achievements of intelligent autism identification and diagnosis to practice.

Key words: machine learning, autism, early identification and diagnosis, systematic review