Introduction
Neovascular Age-Related Macular Degeneration (nAMD) remains a primary cause of irreversible vision loss globally. While anti-VEGF therapies have significantly improved outcomes, many patients eventually develop macular atrophy or subretinal fibrosis, leading to permanent central vision impairment. Predicting these long-term complications remains a clinical challenge. This study evaluates the capacity of machine learning (ML) models to forecast these outcomes using clinical and genetic data from the first year of treatment.
Methodology
The researchers framed the prediction of atrophy and fibrosis at 36 months as a supervised binary classification task. Three algorithms were evaluated: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machines (SVM). To ensure robustness in a relatively small dataset, Nested Cross-Validation (NCV) was employed. Feature importance was assessed using relative AUC (rAUC), which measures the contribution of specific feature groups—such as visual acuity (ETDRS), retinal fluid markers, and genetic variants (SNPs)—to the model's predictive accuracy.
Dataset and Features
The study utilized the PI15/01374 dataset, comprising 335 eyes from 354 patients across 17 Spanish clinical sites. Unlike controlled trials, this "real-world" dataset reflects routine clinical practice. The feature set included 20 clinical variables (e.g., foveal thickness, presence of retinal fluid) and 14 Single Nucleotide Polymorphisms (SNPs) associated with AMD risk pathways.
Results and Feature Importance
The XGBoost model achieved the highest performance, particularly in predicting the combined outcome of "atrophy and/or fibrosis," with a balanced accuracy of 63% and an AUC of 0.72.
Analysis of feature importance revealed that visual acuity (ETDRS score) at 12 months was the most significant predictor across all tasks. For fibrosis prediction, retinal fluid presence was also highly influential, whereas foveal thickness played a larger role in the combined prediction task.
The "Genetics Gap"
A notable finding was the discrepancy between univariate statistical significance and ML predictive power for genetic markers. While certain SNPs, such as those in the CFI gene, showed significant associations in isolation, they did not materially improve the performance of the integrated ML models.
This "genetics gap" suggests that clinical symptoms at 12 months may already capture much of the prognostic information provided by these specific genetic variants, or that the signal-to-noise ratio of these markers is too low for current models to exploit effectively in a real-world setting.
Discussion and Limitations
The study highlights several hurdles for ML in ophthalmology. The models were likely constrained by the categorical nature of the data (e.g., binary "yes/no" for retinal fluid) rather than quantitative volumetric measurements. Furthermore, predicting atrophy proved more complex than fibrosis, likely due to the multifaceted metabolic and environmental drivers of tissue wasting that are not fully captured in standard clinical charts.
Conclusion
While machine learning offers a promising path toward personalized medicine in nAMD, this study serves as a "reality check." Achieving higher predictive accuracy will likely require larger datasets and higher-fidelity, quantitative imaging data to overcome the inherent noise and complexity of real-world clinical progression.
Summary updated