IT department seminar

Early detection system for ASD in toddlers, based on eye tracking and machine learning

Constat

Autism spectrum disorders (ASD) are currently diagnosed primarily by clinical observation, as no validated biomarkers are available for routine clinical use.

Although the first signs may appear around the age of 12 months, a formal diagnosis is often not made until between 3 and 6 years of age.

To remedy this delay, we propose an early screening approach for ASD as early as 9 months of age.

This procedure enables pediatricians to objectively identify early indicators of ASD risk, promote early referral to preventive intervention programs and potentially reduce the long-term impact of the disorder.

Proposal

The proposed system presents infants with a series of videos specifically designed to elicit visual behaviors associated with ASD risk.

Eye-tracking data collected during viewing is analyzed using machine-learning classifiers to estimate the likelihood of ASD risk.

Several algorithms have been evaluated, including Random Forest, support vector machines (SVM), multilayer perceptron (MLP), k-nearest neighbors (kNN) and AdaBoost.

When it came to distinguishing typical development (TD) from ASD levels 1, 2 and 3, the best-performing classifier - SVM - achieved an AUC ROC of 0.9005 and a sensitivity of 75.2%.

In a more restrictive comparison between DT and autism levels 2 and 3, the Random Forest model achieved an AUC ROC of up to 0.9508 and a sensitivity of 87.64%.

Machine learning, eye tracking, autism

Daniel Fernández-Lanvin is a lecturer in the Department of Computer Science at the University of Oviedo, Spain, where he is part of the Human-Computer Interaction research group.

His work focuses on human-computer interaction, user experience, eye tracking and the analysis of user behavior based on interaction data and artificial intelligence techniques.

His work explores how interaction patterns and visual attention can be analyzed to better understand user behavior and develop intelligent systems in fields such as digital health, accessibility, e-learning and web ergonomics.

He is the author of numerous publications in international journals and conferences listed in major databases, including in journals such as Internet Research, Virtual Reality, Multimedia Tools and Applications and the International Journal of Human-Computer Interaction.

Dr. Fernández-Lanvin has participated in and led several research and innovation projects and collaborates with interdisciplinary teams in fields related to digital health technologies and early detection systems based on behavioral data.

He also supervises PhD and Master's research and teaches courses related to human-computer interaction and software interface design at the University of Oviedo.