Multivariable Prognostic Models for Age-Related Macular Degeneration: A Systematic Review

Prognostic prediction models for age-related macular degeneration (AMD) are proliferating, driven by machine learning applied to retinal imaging. Although the methodological limitations of clinical prediction models are well recognized in general, AMD-specific multivariable prognostic models, as distinct from diagnostic or image-classification models, have not been appraised against contemporary prediction-model standards, with integrated assessment of outcome definitions, prediction horizons, calibration, validation design, and risk of bias.

To systematically identify and critically appraise multivariable prognostic models predicting progression from non-late to late AMD and to determine whether quantitative synthesis was appropriate, with emphasis on model design, predictors, performance reporting, validation, and risk of bias.

We searched PubMed/MEDLINE, and Embase from inception to April 2022, updated to December 2024, for studies developing, validating, or evaluating multivariable prognostic models for progression to late AMD. We extracted study population, outcomes, prediction horizons, predictors, modeling methods, performance measures, and validation approach and assessed risk of bias with PROBAST. Given substantial heterogeneity, findings were synthesized narratively in accordance with a prespecified conditional analysis plan (PROSPERO CRD42022323522).

Twenty studies were included. Models were derived from a small number of recurring cohorts, notably AREDS and HARBOR, of predominantly European ancestry; this limited population diversity was one of several constraints, alongside heterogeneous outcome definitions, infrequent calibration, and uncommon external validation. Approaches ranged from regression-based risk scores to deep learning applied to fundus photography, optical coherence tomography, or multimodal data. Because outcome definitions varied (composite late AMD, neovascular or exudative conversion, geographic atrophy or other atrophic endpoints, and treatment-initiation proxies), the observed AUC/c-statistic range (∼0.65-0.97) reflects differing prediction tasks rather than directly comparable performance. Calibration was addressed in six studies (30%), and fewer still met our definition of quantitative calibration. Imaging features and age were the most consistently retained predictors, and measures beyond discrimination were rare. Overfitting could not be excluded in several studies, as event counts, candidate-predictor numbers, use of shrinkage, and internal-validation procedures were incompletely reported. Ten of 20 studies were at low overall risk of bias, but calibration and external validation were uncommon across the evidence base as a whole.

Prognostic models for AMD progression are methodologically diverse but are drawn from few, homogeneous cohorts, with limited calibration and external validation, features that constrain interpretation of reported performance and transportability to routine, more diverse practice. Progress will depend less on new algorithms than on staged, feasible validation: prespecified outcomes and horizons, robust internal validation, routine calibration, and temporal or geographic validation with recalibration before clinical use. Until then, existing models should be regarded as exploratory rather than clinically actionable.