Variational Autoencoder for Image-Based Augmentation of Eye-Tracking Data.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Computer Vision

Background:

• Deep learning excels with large datasets but faces challenges in data-scarce domains like healthcare.
• Healthcare data is often limited, imbalanced, and inaccessible due to privacy concerns.
• Generative modeling and data augmentation are crucial for advancing AI in healthcare.

Purpose of the Study:

• To explore a machine learning-based approach for generating synthetic eye-tracking data.
• To investigate the novel application of variational autoencoders (VAEs) for creating synthetic scanpath data.
• To assess the utility of VAE-generated data for augmenting limited datasets and improving classification performance.

Main Methods:

• Developed and trained a variational autoencoder (VAE) model.
• The VAE was trained to generate image-based representations of eye-tracking data, known as scanpaths.
• Evaluated the VAE's ability to produce plausible scanpath data from limited input.

Main Results:

• The VAE model successfully generated plausible synthetic eye-tracking data from a limited dataset.
• The generated data demonstrated the potential for effective data augmentation.
• The approach showed promise in enhancing classification task performance.

Conclusions:

• Variational autoencoders offer a viable method for generating synthetic eye-tracking data.
• This technique can serve as a valuable data augmentation strategy in data-limited fields like healthcare.
• The study validates the use of VAEs for improving AI model performance in classification tasks with scarce data.