Download the dissertation (PDF, 3 MB)
Abstract
Visual impairments, including cataracts, age-related macular degeneration, and glaucoma, affect millions globally, often progressing gradually and unnoticed until a significant part of vision is lost. These visual impairments lead to irreversible vision loss if left untreated. At the same time, treatment is available to slow down the progression of these conditions, underscoring the need for early detection and intervention. This dissertation explores how gaze behavior changes in response to cataracts, age-related macular degeneration, and glaucoma and evaluates the feasibility of using these changes for early detection through gaze-based machine learning models. Framed within the broader goals of preventive healthcare, this work introduces a comprehensive framework for building scalable, low-burden, and privacy-preserving diagnostic systems grounded in eye tracking.
To achieve this, the dissertation addresses three core research objectives. First, it investigates how simulated cataracts, glaucoma, and age-related macular degeneration affect gaze patterns during visual tasks in high-fidelity virtual environments. It uses perceptually validated simulations, refined through structured interviews with affected individuals. This work captures nuanced behavioral adaptations, such as altered fixation rates, reduced saccadic amplitudes, and increased task durations. Notably, the early stages of cataract and glaucoma often resulted in measurable gaze changes before participants became consciously aware of any visual degradation, highlighting gaze-based detection as a promising avenue for early screening.
Second, building on the finding that visual impairments induce measurable changes in eye movements, the thesis investigates how these changes can be reliably detected using machine learning. A central challenge in this effort is the presence of blink-related artifacts and missing data, which can distort gaze signals and compromise model performance. We conduct a structured literature review and identify inconsistencies in how such artifacts are handled across existing studies. In response, we introduce and validate a standardized preprocessing pipeline using 11 public datasets. Our empirical evaluation demonstrates that trimming blink-related artifacts up to 70 ms before and 118 ms after a blink with empirically derived thresholds and applying interpolation methods such as linear or cubic splines significantly enhances model reliability, particularly in sequence models such as long short-term memory neural networks. Using this pipeline, we allow for more consistent and reproducible gaze data handling, which is crucial for training robust predictive models.
Third, my dissertation evaluates the application of gaze-based models for real-time detection and monitoring of central field loss. Using an existing dataset with simulated gaze-contingent scotomas, we compare traditional feature-based classifiers (support vector machines and random forests) with deep learning models (long short-term memory neural networks) trained on raw gaze sequences. Both approaches achieved high accuracy and precision in identifying central field loss conditions. Importantly, we introduce a novel cost-benefit framework considering model performance and computational latency. This analysis identifies a 1600 ms gaze window as the optimal balance between accuracy and responsiveness, enabling deployment on consumer-grade devices such as smartphones or tablets.
After this, we integrate the findings across the three chapters to demonstrate that gaze behavior can reveal subtle and condition-specific signs of visual impairment before conscious awareness. Simulations of cataracts, glaucoma, and age-related macular degeneration in virtual reality revealed measurable changes in gaze behavior, such as increased fixation rates and altered saccades, that aligned with clinical symptomatology from the literature. By validating simulation realism through interviews, the studies aim to stay close to ecological fidelity. Further, a standardized preprocessing pipeline for blink-induced gaps improved signal quality and model reliability. Finally, long short-term memory neural networks proved capable of detecting early-stage central field loss from short gaze windows, balancing accuracy and efficiency. Together, these results position gaze-based monitoring as a promising tool for scalable, preventive screening of vision disorders.
In conclusion, this dissertation makes three key contributions: (1) a detailed characterization of gaze behavior under different types and severities of visual impairment, (2) a standardized and validated preprocessing pipeline to handle blink-induced artifacts in gaze data, and (3) the demonstration of gaze-based predictive models for early detection of central field loss, optimized for real-time, low-latency applications. These findings form the foundation for future gaze-based diagnostic systems that integrate seamlessly into daily interactions, enabling scalable and privacy-aware screening tools that support timely intervention and reduce the long-term burden of undiagnosed vision loss.
Cite
@phdthesis{grootjen2025gaze,
title = {Analyzing and Predicting Gaze Behavior of People with Visual Impairments},
author = {Grootjen, Jesse W.},
year = {2025},
school = {Ludwig-Maximilians-Universität München},
doi = {10.5282/edoc.36143}
}