Introduction
Age-related Macular Degeneration (AMD) remains a leading cause of irreversible vision loss globally. Despite its prevalence, clinical management is often reactive, relying on static grading systems that categorize disease severity based on a single snapshot in time. These traditional metrics—such as the size of fatty deposits known as drusen—fail to account for the dynamic nature of the disease. Because AMD is a longitudinal process, static biomarkers cannot adequately capture the rate or specific trajectory of progression, which are essential for predicting a patient's transition to sight-threatening late-stage AMD.
To address this, researchers have developed a novel framework that shifts the focus from static snapshots to temporal trajectories. By modeling patient histories as paths within a self-supervised latent space, this method automatically discovers "dynamic biomarkers" that are more predictive of future vision loss and highly interpretable to clinicians.
Methodology
The proposed pipeline utilizes a multi-stage approach to identify and cluster disease progression patterns without the need for exhaustive manual annotation.
1. Self-Supervised Feature Extraction
The framework begins by mapping complex retinal Optical Coherence Tomography (OCT) scans into a high-dimensional feature space. Using Bootstrap Your Own Latent (BYOL), a self-supervised learning technique, a ResNet50 model was trained on over 170,000 OCT scans. This allows the system to group similar biological states together in a 2048-dimensional space based on visual features, entirely independent of human-provided labels.
2. Trajectory Partitioning and Comparison
Since patient histories vary in duration and frequency of visits, the authors partitioned long-term data into "sub-trajectories" representing approximately one year of disease progression. To compare these paths, they developed a custom distance metric, $D_{subtraj}$, which evaluates three critical components:
- Transition Similarity: Whether the paths start and end in similar biological states.
- Path Similarity: Utilizing Dynamic Time Warping (DTW) to determine if the intermediate steps are similar, regardless of the speed of progression.
- Directional Similarity: Whether the trajectories move in the same relative direction within the latent feature space.
3. Discovery of Temporal Biomarkers
Finally, the sub-trajectories were grouped using spectral clustering. Each resulting cluster represents a distinct progression pattern, effectively serving as a candidate temporal biomarker that describes a specific way the disease evolves over time.
Experimental Results
The framework was developed and validated using two large-scale longitudinal datasets: a development set from the Southampton Eye Unit (46,496 scans) and an independent "unseen" dataset from Moorfields Eye Hospital (114,062 scans). The use of an independent test set ensures that the discovered biomarkers are robust across different clinical environments and imaging equipment.
Predictive Performance
The discovered temporal clusters were evaluated on their ability to predict the time remaining until a patient converts to late-stage AMD (specifically CNV or cRORA). The results demonstrated that these automatically discovered clusters outperformed established clinical grading systems. On the development dataset, the temporal clusters achieved a Mean Absolute Error (MAE) of 0.746 years, compared to 0.757 years for traditional grading methods.
Clinical Interpretability
Beyond quantitative performance, the clusters were found to be highly meaningful to ophthalmologists. Expert review confirmed that the algorithm successfully separated distinct biological pathways, such as the specific development of intraretinal fluid versus subretinal fluid, and the growth or regression of pigment epithelial detachments (PEDs). Many of these patterns are recognized by clinicians as vital indicators but are currently absent from formal clinical grading protocols.
Discussion and Future Work
This research represents a significant move toward "precision ophthalmology." By treating disease progression as a continuous path in a high-dimensional space rather than a series of discrete categories, the framework allows AI to assist in the discovery of the very metrics humans should use to track AMD.
However, several areas for improvement remain:
- Dimensionality: The current model utilizes 2D OCT B-scans. Transitioning to 3D volumetric scans could provide a more comprehensive view of retinal changes.
- Cohort Bias: The datasets used were skewed toward patients already in late-stage disease. Future applications could focus on earlier-stage cohorts to identify subtle, pre-symptomatic biomarkers.
- Standardization: While the clusters are predictive, further refinement is necessary before they can be integrated into standardized clinical practice.
By moving beyond the limitations of static snapshots, this trajectory-based approach provides a more nuanced understanding of AMD, offering the potential for earlier intervention and more personalized patient care.
Summary updated