Skip to Main content Skip to Navigation

Détection automatisée du trouble du spectre de l'autisme via eye-tracking et réseaux de neurones artificiels : conception d'un système d'aide à la décision

Abstract : The Computer Sciences and Psychology fields are very far from each other. However, some needs in Cognitive Psychology (CP) can be satisfied through the use of Artificial Intelligence (AI), in particular its connectionist approaches. There are a few uses of AI principles in CP, but they are quite inconspicuous. In particular, the case of the diagnosis support applied to Autism Spectrum Disorder (ASD) is a blank slate, with a few exceptions. As some work in CP show a great difference between some autistic traits and the ability to focus, notably visually, we have worked on data from an Eye-Tracker to try and detect young children with ASD apart from others without ASD of equivalent age. This data is of two distinct formats, on one hand based on the strict recording of the eyes positions on the screen (2 dimensions) and from the screen (3 dimensions), and on the other hand based on an automatic analysis of the eye sight dynamics. These two data types are digital models of information, generally read via a graphical plot over a video, by CP experts while observing their participants. The work in this thesis have focused on the application of various data modifications (presented as images, with a dimension reduction or statistics). It used various AI models (Artificial, Recurrent and Convolutional Neural Networks) to produce detection support techniques. Then, we have applied a weighted mean over these results to increase the precision of this detection support technique. Through the work conducted in this thesis, various approaches have been tried, keeping as a common element the neural networks as learning tool. Starting with a study about eye focus event and LSTM, we continued with ocular recorded raw data and a image-based modeling, used with a CNN. After applying PCA on these images, this data are also used with a simple ANN, with enhanced results. Then, a statistical study aims for the study of correlation between the values recorded by our Eye-Tracking device, the correlation are then given to an ANN to reach the highest results of this work. Finally, we tried to gather the information of some of these models to improve the obtained results once more.The best of our results allowed to produce a ROC curve with an AUC reaching 95%, which allow to think about an almost perfect support, given we could add some data used in manual diagnosis, and a complete freeing of the experts' time to enable a total focus on the setup of the child's support and his/her closer following. Moreover, the opportunity of an earlier child diagnosis can help better reducing the child's neurodevelopmental delay. None of these models are perfect. Still, there is to be noted that the diagnosis is never made, in practice, with the sole Eye-Tracking related data. Aiming to constitute a diagnosis support system, it essentially serves as a guiding tool for the professional in his/her work and to free a part of his/her time
Document type :
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Thursday, March 31, 2022 - 2:11:09 PM
Last modification on : Friday, August 5, 2022 - 11:22:17 AM
Long-term archiving on: : Friday, July 1, 2022 - 7:10:20 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03626269, version 1



Romuald Carette. Détection automatisée du trouble du spectre de l'autisme via eye-tracking et réseaux de neurones artificiels : conception d'un système d'aide à la décision. Autre [cs.OH]. Université de Picardie Jules Verne, 2020. Français. ⟨NNT : 2020AMIE0025⟩. ⟨tel-03626269⟩



Record views


Files downloads