Explainable Eye Tracking Analysis Using a Deep Learning Approach for Autism Detection

Overview

Eye-tracking data can provide useful insights into cognitive processing and attention mechanics, which is particularly helpful for detecting autism. However, most existing models fail to account for the spatial structure of scan routes, and as a result, their predictions lack proper interpretability. This project uses an Explainable Deep Learning (DL) based model for classifying eye-tracking scanpath images. This framework integrates deep feature extraction, dimensionality reduction, as well as a hybrid classification pipeline that utilizes machine learning and deep learning models. Using said models, it is possible to extract distinguishing features from preprocessed scanpath images using pre-trained DL models, such as DenseNet121 and EfficientNetB0. In this project, a series of experiments were conducted, primarily focusing on three approaches. The first approach was to extract the features and use them to train various machine learning (ML) and deep learning (DL) algorithms. The second approach was to dimensionally reduce the extracted features, followed by training on various ML and DL algorithms. The final approach was to train the preprocessed scanpath images directly using the pre-trained models. The highest accuracy was achieved by DenseNet121 at 99.39% which was trained end-to-end on preprocessed images. Visual explanations were generated by Grad-CAM, which were integrated into the framework to enable interpretable predictions as verified by a licensed neurologist. This allows for interpretable predictions, making it useful for high-stakes applications in neuroscience due to transparency and performance. The proposed approach is engineered for extensibility in distributed and parallel learning environments, which makes it appropriate for large-scale eye-tracking datasets.

Competition history

• CSEF 2026 Behavioral & Social Sciences (Senior Division) · Entry S-03-05

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Behavioural and Social Sciences projects
• CSEF Behavioural and Social Sciences projects
• Every CSEF 2026 project