Presented at 2022 15th International Conference on Health Informatics (HEALTHINF) Mahmoud Elbattah, Jean-Luc Guérin, Romuald Carette, Federica Cilia, Gilles Dequen Abstract The potentials of Transfer Learning (TL) have been well-researched in areas such as Computer Vision and Natural Language Processing. This study aims to explore a novel application of TL to detect Autism Spectrum Disorder. We seek to develop an approach that combines TL and eye-tracking, which is commonly used for analyzing autistic features. The key idea is to transform eye-tracking scanpaths into a visual representation, which could facilitate using pretrained vision models. Our experiments implemented a set of widely used models including VGG-16, ResNet, and DenseNet. Our results showed that the TL approach could realize a promising accuracy of classification (ROC-AUC up to 0.78). The proposed approach is not claimed to provide superior performance compared to earlier work. However, the study is primarily thought to convey an interesting aspect regarding the use of (synthetic) visual representations of eye-tracking output as a means to transfer representations from models pretrained on large-scale datasets such as ImageNet.

1. HEALTHINF 2022
Vision-BasedApproach forAutism DiagnosisUsing
TransferLearning and Eye-Tracking
Mahmoud Elbattah, Jean-Luc Guérin, Romuald Carette, Federica Cilia, Gilles Dequen
Laboratoire MIS
Université de Picardie Jules Verne (UPJV), France
mahmoud.elbattah@u-picardie.fr
https://www.researchgate.net/publication/358842561_Vision-based_Approach_for_Autism_Diagnosis_using_Transfer_Learning_and_Eye-tracking

2. HEALTHINF 2022
Background:Autism SpectrumDisorder
• Autism Spectrum Disorder (ASD) is a pervasive developmental disorder
characterised by a set of impairments including social communication problems.1
• ASD has been considered to affect about 1% of the world’s population (US Dep. of
Health, 2018). 2
• The hallmark of autism is an impairment of the ability to make and maintain eye
contact. 3
2
1 L. Wing, and J. Gould, “Severe Impairments of Social Interaction and AssociatedAbnormalitiesin Children: Epidemiologyand
Classification”.Journal of Autism and DevelopmentalDisorders,9(1), pp.11-29,1979.
2
U.S. Departmentof Health & Human Services.Data and statistics | autism spectrum disorder(asd) | ncbddd | cdc, 2018.URL:
https://www.cdc.gov/ncbddd/autism/data.html.
3
Coonrod,E. E. and Stone, W. L. (2004).Early concerns of parents of children with autistic and nonautistic disorders.Infants & Young
Children, 17(3),258–268.

3. HEALTHINF 2022
Background:Eye-TrackingTechnology
3
Image Source: https://imotions.com/blog/eye-tracking/
J.H. Goldberg, and J.I. Helfman, “Visual scanpath representation”, In Proceedings of the 2010 Symposium on Eye-Tracking Research &
Applications, ACM, 2010, pp. 203-210.
Scan-path

4. HEALTHINF 2022
Motivation
Our Goal:
• Detecting ASD-diagnosed individual in eye-tracking data.
Key Idea:
• Compactly render eye movements into an image-based format while maintaining the
dynamic characteristics of eye motion.
• As such, the classification task could be approached as image classification.
4

5. HEALTHINF 2022
Our EarlierWork (HEALTHINF 2019)
5
Carette, R., Elbattah, M., Cilia, F.,Dequen, G., Guérin, J, & Bosche, J. (2019). Learning to predict autism spectrumdisorder based on the visualpatterns of
eye-tracking scanpaths. In Proceedingsof the 12th InternationalConference on Health Informatics(HEALTHINF).

6. HEALTHINF 2022
Data Description
6
Number of Participants(ASD, TD) 59 (29, 30)
Gender Distribution (M, F) 38 (≈ 64%), 21 (≈ 36%)
Age (Mean, Median) years 7.88, 8.1
CARS Score (Mean, Median) 32.97, 34.50

7. HEALTHINF 2022
Image DatasetDescription
• 547 images: 328 (Non-ASD), 219 (ASD)
• Image dimensions: 640x480
7
ASD Non-ASD
Carette, R., Elbattah, M., Dequen, G., Guérin, J. L., & Cilia, F. (2018, September). Visualization of eye-tracking patterns in autism spectrum disorder: method and
dataset. In 2018 Thirteenth International Conference on Digital Information Management (ICDIM) (pp. 248-253). IEEE.

8. HEALTHINF 2022
TransferLearning Models(FeatureExtraction)
• VGG-16: A deep CNN architecture developed by a group of researchers from
the University of Oxford (Simonyan, and Zisserman, 2014).
• ResNet: Residual Networks (ResNet) (He et al., 2016).
• DenseNet: Densely Connected Convolutional Network (DenseNet)
architecture (Huang et al., 2017).
8

9. HEALTHINF 2022
DatasetSplitting(3-FoldCross-Validation)
The dataset was split using the following stepwise procedures:
1. Split Participants: Initially, the group of 59 participants was randomly split
into two independent sets (i.e. train and test).
2. Match Images: Based on the IDs of participants, the images were matched
and loaded into the train and test sets.
3. Repeat: Step #1 and Step #2 would be repeated for each round of the
cross-validation process.
9

10. HEALTHINF 2022
ExperimentalResults
10

11. HEALTHINF 2022
ExperimentalResults(cont’d)
11
AVERAGE PRECISION-RECALL OF MODELS.
Model Recall (~) Precision (~)
VGG-16 0.56 0.67
ResNet 0.54 0.65
DenseNet 0.55 0.65

12. HEALTHINF 2022
Conclusionsand Limitations
• Popular vision models such as VGG-16, ResNet, and DenseNet could achieve a
quite promising performance.
• The TL approach was largely applicable, even though the source dataset (i.e.,
ImageNet) is assumed to have included quite different types of images.
• It is not claimed that the TL approach could provide superior performance.
• However, it is conceived that the scarcity or imbalance of datasets could make such
TL approaches attractive for further investigation.
12

13. HEALTHINF 2022
Thank You!