Real-world database evaluation of drug-associated vitreous opacities and machine learning for clinical interpretability.

Background: With visual disturbances from vitreous opacities (VOs) and floaters drawing increasing attention, we analyzed real-world data from the U.S. Food and Drug Administration Adverse Event Reporting System (FAERS) to characterize VO-associated drug profiles and inform clinical strategies for reducing VO-related complications. Materials and methods: Disproportionality analysis was performed on FAERS reports (2004-2024) to identify VO-associated drugs. Drugs were then classified to assess the onset time and baseline characteristics. Multivariable logistic regression was used to evaluate confounders. The predictive performance was compared using six machine learning algorithms, with SHapley Additive exPlanations (SHAP) used for feature importance. Results: Among 3,817 VO-related reports, 38 drugs were identified as independent risk factors, and they were mainly ocular, oncologic, hormonal, antimicrobial, and immunologic agents. Antimicrobial drugs had the earliest onset (mean 43.6 days), and hormonal drugs had the latest (mean 409.2 days). In the bootstrapped aggregating (BAG) model, the top predictors of VO were dexamethasone, reporter, time, brolucizumab, and age. The five highest-risk drugs were dexamethasone, brolucizumab, triamcinolone, faricimab, and fingolimod. Conclusion: This first systematic real-world evaluation of VO-related adverse drug reactions identifies high-risk drugs, susceptible populations, and onset patterns, thus offering guidance for preventive medication strategies. The BAG model showed higher sensitivity in real-world analysis, suggesting potential for further research in VO and floater prevention and treatment.