[Machine learning algorithms for identifying autism spectrum disorder through eye-tracking in different intention videos]

Abstract

Objectives: To investigate the differences in visual perception between children with autism spectrum disorder (ASD) and typically developing (TD) children when watching different intention videos, and to explore the feasibility of machine learning algorithms in objectively distinguishing between ASD children and TD children.

Methods: A total of 58 children with ASD and 50 TD children were enrolled and were asked to watch the videos containing joint intention and non-joint intention, and the gaze duration and frequency in different areas of interest were used as original indicators to construct classifier-based models. The models were evaluated in terms of the indicators such as accuracy, sensitivity, and specificity.

Results: When using eight common classifiers, including support vector machine, linear discriminant analysis, decision tree, random forest, and K-nearest neighbors (with K values of 1, 3, 5, and 7), based on the original feature indicators, the highest classification accuracy achieved was 81.90%. A feature reconstruction approach with a decision tree classifier was used to further improve the accuracy of classification, and then the model showed the accuracy of 91.43%, the specificity of 89.80%, and the sensitivity of 92.86%, with an area under the receiver operating characteristic curve of 0.909 (P<0.001).

Conclusions: The machine learning model based on eye-tracking data can accurately distinguish ASD children from TD children, which provides a scientific basis for developing rapid and objective ASD screening tools.

Keywords: Action understanding; Autism spectrum disorder; Child; Eye tracking; Joint intention; Machine learning.

图1. 试验流程 ［JI］联合意图；［NJI］非联合意图。

图2. 11个AOI划分示意图

图3. ML流程图

图4. DT分类器最高分类准确度对应的受试者操作特征曲线