A Practical Model for ASD Screening By Eye Tracking: Combined Fixation on Human Faces and Pupillary Light Responses (PLR)

Background: As a country with huge population and limited experienced pediatric psychiatrists on ASD, China has the need for auxiliary diagnostic technique on ASD to do large scale screening. In the field of early ASD detection, there are some prominent achievements in research, but few of them are turned into practicing, especially in China.

A.Klin and W. Jones who were the first using eye tracking for ASD research found 12-month high-risk ASD babies showed different interests on eyes and mouths. Pupillary Light Response (PLR) was another biomarker that was proved to be efficient in identifying ASD and TD.

Objectives: The goal is to design a practical model for ASD screening on children of 3-6 years old by eye tracking technique, the model features come from two experiments with different paradigms of fixation on human faces and pupillary light response (PLR).

Methods: The experiments use SMI desktop eye tracking system with frequency 250Hz. Experiment I consist of 12 pictures of human faces in same size and luminance with different gender and age. Starting with a 10 second black picture, 12 pictures were showed continuously with 5 seconds on each. Fixation time on Area of Interest (AOI) of eyes, mouths, and faces were collected, as well as pupillary data. Experiment II also consists of 12 pictures, but with 6 black and 6 white pictures. Starting with a 10 second grey picture, one black and one white picture were showed alternately. Pupillary data were collected. Data were analyzed later in SPSS and MATLAB.

53 ASD and 51 TD children of 3-6 years old were recruited. 34 ASD children passed the test on calibration of Eye Tracker and double checked by CARS (Child Autism Rate Scale). After deleting invalid data, 73 subjects with 26 ASD and 47 TD were left. Data from the 73 subjects were split into two sets of train and test. The train set has 36 subjects with 13 ASD and 23 TD, while the test set has 37 subjects with 13 ASD and 24 TD. The train set was used to train the model, and the test set was used to evaluate the performance of the model.

Based on the data of experiment I and experiment II, feature extraction and selection were carried out respectively. Two independent classifiers were constructed, whose outputs were combined for final results. ASD/TD decision was made only if output I and output II both gave ASD/TD judgement, otherwise no decision was made. Please refer to attached flow chart.

Results:

By applying the data of test set to the model, it showed a ratio of 73% on valid prediction for the 37 subjects, while no decision for the remaining 27% of the subjects. Among the valid predictions, the precision rate on ASD is 100%, and the precision rate on TD is 94.73%.

Conclusions: The model can be used for ASD screening on 3-6 years old children.