TLDR: A study utilized machine learning to predict Primary Angle-Closure Glaucoma (PACG) progression by combining structural data from OCT scans of the optic nerve head (ONH) with functional data from visual field (VF) tests. The Random Forest model achieved the best performance (AUC 0.87), outperforming models using single data types. Key predictors included inferior ONH features like Minimum Rim Width (MRW) and Retinal Nerve Fiber Layer (RNFL) thickness, nasal-temporal Lamina Cribrosa (LC) curvature, and superior nasal VF sensitivity, demonstrating the enhanced accuracy of integrated data for early risk classification.
Primary Angle-Closure Glaucoma (PACG) is a significant cause of irreversible vision loss, particularly prevalent in Asian populations. Unlike its more common counterpart, Primary Open Angle Glaucoma (POAG), PACG often remains undetected until advanced stages, making early and accurate prediction of its progression crucial for preserving vision.
A recent study titled Fusing Structural Phenotypes with Functional Data for Early Prediction of Primary Angle-Closure Glaucoma Progression, conducted by a team of researchers including Swati Sharma, Thanadet Chuangsuwanich, Royston K. Y. Tan, Shimna C. Prasad, Tin A. Tun, Shamira A. Perera, Martin L. Buist, Tin Aung, Monisha E. Nongpiur, and Michaël J. A. Girard, addresses this challenge by integrating structural and functional eye data with machine learning to predict glaucoma progression.
The Challenge of Glaucoma Progression
Monitoring glaucoma progression traditionally relies on visual field (VF) tests, which assess a patient’s visual function. However, these tests can be subjective and variable, often requiring multiple tests over several years to establish a reliable progression rate. This delay can hinder timely treatment decisions, especially for fast-progressing cases. Optical Coherence Tomography (OCT) provides structural information about the optic nerve head (ONH) and retinal layers, which can sometimes show damage before functional impairment. While previous studies have combined these data types, many focused on POAG and lacked clear explanations of which factors were most influential.
A Novel Integrated Approach
This study aimed to overcome these limitations by developing a predictive model specifically for PACG. The researchers combined 31 structural parameters extracted from OCT images of the ONH with 10 functional parameters derived from sector-based VF maps. These parameters were fed into various machine learning (ML) models to classify eyes as either ‘slow progressors’ or ‘fast progressors’. Fast progression was defined as a Visual Field Index (VFI) decline of less than -2.0% per year, while slow progression was a decline of -2.0% or more per year.
The study analyzed 451 eyes from 299 PACG patients. To ensure robust and unbiased results, the dataset was split into training, validation, and testing sets over 2000 iterations using Monte Carlo cross-validation. Due to an imbalance in the dataset (more slow progressors than fast progressors), higher weights were assigned to the fast progression group during training.
Key Findings and Predictive Factors
Among the various ML models tested, the Random Forest algorithm demonstrated the best performance, achieving an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.87. This indicates a high accuracy in distinguishing between slow and fast progressors. Importantly, models that combined both structural and functional data significantly outperformed those using only structural data (AUC 0.82) or only functional data (AUC 0.78).
To understand which factors were most critical for prediction, the researchers used Shapley Additive Explanations (SHAP) analysis. This method identifies the most influential parameters driving the model’s predictions. The top six key predictors identified were:
- Minimum Rim Width (MRW) in the inferior region of the ONH.
- Average Retinal Nerve Fiber Layer (RNFL) thickness in the inferior region.
- Nasal-temporal Lamina Cribrosa (LC) curvature.
- Average VF sensitivity in the superior nasal sector.
- RNFL thickness in the inferior-temporal region.
- Ganglion Cell-Inner Plexiform Layer (GCL+IPL) thickness in the inferior region.
The SHAP analysis also revealed important correlations: decreasing RNFL thickness, GCL+IPL thickness, and MRW were linked to a higher risk of fast progression. Conversely, increased nasal-temporal LC curvature was associated with faster progression. These findings align with anatomical understanding, as structural damage in the inferior ONH often corresponds to superior hemifield visual field loss.
Also Read:
- Uncertainty in AI: The Role of Data Augmentation in Diabetic Retinopathy Prediction
- AI Models Enhance Knee MRI Analysis for Injury Detection
Implications and Future Directions
This study provides compelling evidence that integrating structural and functional data significantly improves the early prediction of PACG progression. The identified key parameters offer valuable insights for clinicians, highlighting specific areas of the eye that are most indicative of disease advancement. This framework has the potential to guide proactive management strategies, helping to preserve visual function and optimize patient outcomes.
While promising, the study acknowledges certain limitations, including its focus on a specific ethnic group and the exclusion of severe PACG cases. Future research could incorporate additional clinical factors like age, gender, intraocular pressure, and treatment history, and validate the model with data from diverse populations and imaging devices to enhance its generalizability and clinical applicability.
