Electrooculography signal generation with conditional diffusion models for eye movement classification

Abstract

Electrooculography (EOG) is a biosignal that encodes the directional information of eye movements and is widely used in eye-tracking applications and human-computer interaction. These applications provide intuitive, accessible interfaces, making EOG valuable as a communication aid. Among these applications, eye writing—an approach in which users draw characters using eye movements—is distinguished for its ability to convey significantly more information than traditional methods. This technique has potential applications for individuals who rely on eye movements for communication, including those with amyotrophic lateral sclerosis. However, achieving high accuracy in eye writing typically requires deep learning models constrained by ethical and legal challenges in collecting large EOG datasets. In this study, we developed a diffusion-model-based generative framework for eye-written character recognition under limited-data conditions, addressing challenges in both accuracy and data availability. The model was implemented using conditional vectorized class information to further increase the diversity of the generated data by indicating the class. The effectiveness of the proposed method was assessed through visual representations of generated signals and comparison of classification accuracies. The model outperformed the generative adversarial network in terms of the visual quality of generated data. It also achieved classification accuracies of 94.63% for Arabic numerals and 80.36% for Japanese Katakana strokes when trained with ninefold data at intermediate steps. Despite having significantly more parameters, it achieved shorter inference time than TimeGAN, further demonstrating computational efficiency and feasibility. Additionally, adapting the proposed method to other bioelectric signals demonstrates significant potential, suggesting a flexible framework for future biosignal research and applications.