Prepared: 2026-10-12 Style: Vancouver numbered citations. Every PMID cited in the reference list was mechanically verified via NCBI E-utilities esummary on the date of preparation; the machine-generated verification record accompanies this review. Anchor source: Aljied R, Saggu S, Chaudhary V, Pigeyre M, Griffith LE, Raina P. Multivariable Prognostic Models for Age-Related Macular Degeneration: A Systematic Review. J Ophthalmol. 2026;2026:9231744 [1]. This review treats that article as a seed and sweeps far beyond it.
How to read the evidence labels in this review. Statements drawn directly from a cited source’s abstract or full text are presented as findings. Statements that combine multiple sources into a claim the sources do not state in that form are marked (inference). Where a widely repeated number could not be traced to a verifiable primary source, it is flagged as unverified rather than repeated.
Abstract
Age-related macular degeneration (AMD) is a leading cause of visual impairment and severe vision loss worldwide [2], and progression from early or intermediate disease to its sight-threatening late forms — neovascular AMD and geographic atrophy — is the outcome that matters clinically. A large and accelerating literature has produced multivariable prognostic models intended to estimate an individual’s future risk of that progression, driven in recent years by machine learning applied to retinal imaging. This review synthesizes that literature, anchored on a 2026 systematic review of 20 multivariable prognostic models for progression from non-late to late AMD [1] and extended by PubMed/EuropePMC sweeps, citation snowballing of landmark AMD prediction papers, and ClinicalTrials.gov registry review. Three findings dominate the evidence base. First, progression is prognostically tractable: discrimination across models is commonly in the 0.80–0.92 range and extends from roughly 0.65 to 0.97, but these figures span heterogeneous prediction tasks — different outcomes (composite late AMD, neovascular conversion, geographic atrophy, treatment initiation as a proxy), different horizons (three months to fifteen years), and different analytic units (eye, patient, or scan) — and are therefore not directly comparable [1]. Second, the evidence base is narrow in its foundations: it is derived from a small number of recurring cohorts, above all AREDS and HARBOR, drawn predominantly from European-ancestry populations, which limits confidence in transportability [1]. Third, methodological maturity lags technical output: calibration was addressed in only 6 of 20 studies (30%), external validation in 7 of 20 (35%), model updating/recalibration was essentially absent, and risk-of-bias concerns clustered in the analysis domain, most often reflecting limited event counts relative to candidate predictors and incomplete reporting of model development [1]. Conventional Cox and logistic models based on clinical phenotype, age, smoking, and genetic risk scores frequently performed comparably to deep learning for medium- and long-term prediction, belying the implicit claim that deep learning has displaced traditional prognostic modelling [1]. The most recurrently retained predictors — age, drusen burden and morphology, pigmentary abnormalities, reticular pseudodrusen, outer-retinal thinning, and OCT features such as hyperreflective foci — overlap heavily with established markers of current disease severity, raising the possibility that some imaging models are automating severity staging rather than discovering independent prognostic signal (inference). Genetic risk scores (notably CFH and ARMS2/HTRA1) improve risk stratification but add only modest incremental discrimination over phenotype, and their performance is ancestry-dependent. The field’s principal deficit is not algorithmic novelty but staged validation: prespecified and standardized outcomes, robust internal validation, routine quantitative calibration, temporal or geographic external validation, and recalibration before clinical use. Until then, existing AMD prognostic models should be regarded as exploratory rather than clinically actionable. This review also maps adjacent prognostic tasks (anti-VEGF response prediction, home-monitoring-based early detection), the treatment and regulatory context (complement inhibitors and FDA-recognized retinal AI), the data and equity constraints, and a phased pathway toward clinically useful, well-calibrated, and transportable risk estimation.
Keywords: age-related macular degeneration; prognostic models; risk prediction; machine learning; deep learning; calibration; external validation; geographic atrophy; neovascular AMD; polygenic risk score; TRIPOD; PROBAST.
1. Introduction and Scope
1.1 AMD as a clinical and public-health problem
AMD is a common, age-related degeneration of the central retina and a leading cause of visual impairment and severe vision loss worldwide [2]. Clinically it is described in stages: early disease (medium-sized drusen and retinal pigmentary changes) and late disease, with late disease comprising the neovascular (exudative) form and the atrophic form, geographic atrophy (GA) [2]. Dry (non-neovascular) AMD accounts for roughly 85%–90% of AMD cases and is the substrate from which both atrophic and neovascular late complications arise [3].
The burden is large and growing. Using Global Burden of Disease 2021 data across 204 countries, the number of prevalent cases of low vision and blindness due to AMD rose from approximately 3.64 million in 1990 to approximately 8.06 million in 2021 [4]. Age-standardised prevalence and disability-adjusted life-year rates have shown a modest decreasing trend over the same period, with a global net drift in prevalence of −0.35% per year, indicating that the absolute increase is driven primarily by population ageing and growth rather than by rising age-specific risk [5]. Projections commonly cited in the modelling literature place the number of people affected by AMD at approximately 288 million by 2040 [6], a figure that has become a standard framing device in studies motivating risk prediction [7].
In Europe, meta-analysis of 22 prevalence and four incidence studies estimated pooled prevalence among people aged 60 years and older of 25.3% for early or intermediate AMD and 2.4% for any late AMD, with the number of European Union inhabitants affected by any AMD projected to rise from roughly 67 million to 77 million by mid-century; pooled annual incidence of any late AMD was 1.4 per 1,000 individuals [8]. In American whites aged 50 years and older, the annual incidence of late AMD was estimated at 3.5 per 1,000, equivalent to roughly 293,000 new cases per year [9]. A global meta-analysis of 24 population-based studies found pooled annual incidences of 1.59% for early AMD and 0.19% for late AMD, with people of European descent showing the highest annual incidences of both early (2.73%) and late (0.36%) disease [10]. These incidence figures matter for prognostic modelling because they set the baseline risks against which calibration is judged and shape the feasibility of adequately powered external validation.
1.2 Why prognosis, not just diagnosis, is the operative question
The distinction between diagnostic and prognostic models is central to this review. Diagnostic models identify disease that is present; prognostic models estimate the probability that a defined event will occur within a defined future time horizon [11]. A model that classifies a fundus photograph as showing intermediate AMD is a diagnostic/classification model; a model that estimates the probability that the same eye will develop neovascular AMD within two years is a prognostic model. Much of the AMD artificial-intelligence (AI) literature addresses detection and staging rather than prognosis, and systematic reviews of AMD machine-learning classifiers have overwhelmingly evaluated diagnostic accuracy rather than future-risk estimation [12,13,14,15]. The 2026 anchor systematic review deliberately excluded diagnostic tasks and appraised only multivariable models intended to estimate future risk over a defined horizon [1]. This review follows that scoping logic.
Prognosis matters because the clinical and economic stakes of progression are high and because treatment options are stage-dependent and imperfect. Anti-VEGF therapy is highly effective at treating neovascular AMD and has markedly reduced prevalent visual impairment, while for atrophic disease treatment options remain limited, having only recently expanded with complement-inhibitor therapies [2,16,17]. The need is not merely to identify disease but to triage intensity of surveillance, to counsel patients, and to enrich clinical trials with rapid progressors — tasks that require calibrated absolute risk rather than a binary label [1,18].
1.3 Questions this review answers
This review is organized around eight questions: (1) What is the natural history and what are the outcomes and severity scales that prognostic models target? (2) What methodological standards govern the appraisal of prognostic models, and how do they apply to AMD? (3) What regression- and risk-score-based models exist, and how do they perform? (4) What imaging and machine-learning/deep-learning models exist, and how do they perform? (5) What does the 2026 anchor systematic review conclude about the state of the evidence? (6) What adjacent prognostic tasks (anti-VEGF response, home monitoring, trial enrichment) are relevant? (7) What data, cohort, equity, and regulatory constraints bound the field? and (8) What would move the field from exploratory models to clinically actionable risk estimation?
1.4 Scope and exclusions
Consistent with the anchor review, this review focuses on multivariable prognostic models for progression from non-late to late AMD, and on the methodological, genetic, imaging, and contextual literatures needed to interpret them. Diagnostic and image-classification studies are cited where they inform the discussion (for example, automated AMD severity grading) but are not treated as prognostic evidence. Pure biomarker-discovery studies without a multivariable risk model are cited as mechanistic context. Economic modelling and cost-utility analyses are outside scope except where they bear on implementation. Statistics are cited only when they could be traced to a source’s abstract or full text; no statistic has been reconstructed or estimated by this review.
2. Methods of This Review
2.1 Databases and search strategy
PubMed/MEDLINE was searched via the NCBI E-utilities esearch application programming interface using twenty documented keyword/MeSH queries spanning AMD prognostic and prediction models, AMD progression, machine learning and deep learning in AMD, MeSH terms for Macular Degeneration combined with Prognosis, Risk Assessment, Models, Statistical, and Machine Learning, AMD genetic and polygenic risk scores, geographic atrophy progression, neovascular conversion, anti-VEGF response prediction, OCT biomarkers, AREDS/AREDS2, nomograms, prediction-model methodology (external validation, calibration, PROBAST, TRIPOD), AI in ophthalmology, AMD global burden, complement factor H, AI screening, multimodal models, survival modelling, and treatment landmarks. These queries returned 1,873 unique candidate records, for which full metadata were retrieved via E-utilities esummary. EuropePMC was used for reference and citation endpoints, and ClinicalTrials.gov was queried through the v2 API.
2.2 Snowballing
The seed article’s reference list (52 references) was retrieved via the EuropePMC references endpoint, and reference-and-citation snowballing was performed for sixteen landmark AMD prediction and prediction-methodology papers, yielding an additional 1,424 unique candidate records. Snowballing was essential for surfacing the methodological literature (TRIPOD, PROBAST, calibration, decision-curve analysis, transportability) that is cited only obliquely by primary modelling studies.
2.3 Retrieval of abstracts and full text
Structured abstracts were retrieved for 217 records via E-utilities efetch (PubMed XML), and the anchor systematic review’s full text (including its results text and reference list) was retrieved from PubMed Central as open-access XML. No structured statistics were extracted from figures alone; where a number appears in this review it is traceable to an abstract or to the anchor review’s text.
2.4 Verification of identifiers
Every PMID cited in this review was mechanically verified against NCBI E-utilities esummary by a purpose-written script that extracts all cited PMIDs, batches them to the API, and confirms that each resolves to a live PubMed record with a matching title. Review articles, guideline statements, and registry entries are cited without PMIDs where no indexed record exists, and are labelled as such. Machine-verification results are reported in the companion provenance sidecar.
2.5 Analytical approach
Because of the heterogeneity documented in the anchor review — heterogeneous outcomes, horizons, analytic units, and performance metrics — this review is a structured narrative synthesis rather than a quantitative meta-analysis. Where sources disagree or where a reported statistic is likely to be optimistic, the disagreement is stated explicitly rather than averaged away.
3. Natural History, Outcomes, and Severity Scales
3.1 The AREDS severity scale as the field’s reference standard
The single most influential instrument in AMD prognosis is the Age-Related Eye Disease Study (AREDS) severity scale, developed in AREDS Report No. 17. Using stereoscopic colour fundus photographs graded for drusen characteristics (size, type, area) and pigmentary abnormalities (increased pigment, depigmentation, geographic atrophy), it combined a six-step drusen area scale with a five-step pigmentary abnormality scale into a nine-step severity scale on which the five-year risk of progression to advanced AMD increased progressively [19]. Advanced AMD was defined as neovascular AMD (retinal pigment epithelial detachment, serous or haemorrhagic sensory retinal detachment, subretinal or sub-retinal-pigment-epithelial haemorrhage, or subretinal fibrous tissue), photocoagulation for AMD, or central geographic atrophy [19]. This scale, and the simplified versions derived from it, is the outcome scaffold on which most subsequent prognostic models are built or against which they are compared.
AREDS Report No. 19 characterised the demographic, behavioural, medical, and non-retinal ocular risk factors for incident neovascular AMD and central geographic atrophy in the AREDS cohort followed for a median of 6.3 years, providing the epidemiological foundation for multivariable risk scoring [20]. The AREDS formulation was later tested and refined in AREDS2, a 2×2 factorial randomised trial of lutein plus zeaxanthin and omega-3 fatty acids that also evaluated eliminating beta-carotene and lowering zinc; lutein/zeaxanthin and omega-3 did not significantly reduce progression to advanced AMD overall, but the trial reshaped the recommended supplement formulation and generated long-term follow-up data [21,22,23]. A 2025 post hoc analysis of AREDS and AREDS2 reported that oral antioxidant and lutein/zeaxanthin supplementation slowed progression of geographic atrophy toward the fovea, a finding relevant to the modifiable-risk discussion in Section 3.4 [24]. A 2023 Cochrane review continues to appraise antioxidant vitamin and mineral supplementation for slowing AMD progression [25].
3.2 Late-stage phenotypes: neovascular AMD and geographic atrophy
Progression to late AMD takes two clinically and biologically distinguishable forms. Neovascular AMD is characterised by choroidal neovascularisation and exudation and is treated with anti-VEGF injections; geographic atrophy is characterised by progressive atrophy of the retinal pigment epithelium and outer retina and, until recently, had no approved therapy. The two pathways differ in their prognostic features: models developed for neovascular conversion frequently emphasise drusen-centric features, whereas models for atrophy emphasise outer-retinal thinning and age [26]. The anchor review found that this outcome heterogeneity — composite late AMD, neovascular or exudative conversion, geographic atrophy or other atrophic endpoints, and even initiation of anti-VEGF therapy used as a proxy — is one reason why reported discrimination values are not directly comparable across studies [1].
3.3 Non-imaging severity and functional markers
Beyond imaging, functional and structural surrogates have been evaluated as prognostic candidates. Rod-mediated dark adaptation has been studied as a suitable functional outcome and as a predictor of AMD incidence and progression; in the Alabama Study on Early AMD 2, delayed rod-mediated dark adaptation was associated with AMD incidence and early progression over three years, and dark adaptation measures have been repeatedly evaluated as outcome or predictor variables in early and intermediate AMD [27,28,29]. Microperimetry and low-luminance deficit have been examined for added value in predicting progression beyond conventional features [30]. These functional markers are relevant to the broader predictor space but were not among the most consistently retained predictors in the anchor review, where ocular and imaging variables dominated [1].
3.4 Modifiable risk factors and the prevention argument
Prognosis is clinically actionable only if some risk is modifiable or if surveillance can change outcomes. Smoking and low dietary intake of antioxidants are the principal non-genetic risk factors for AMD, and progression can be slowed with high-dose zinc and antioxidant vitamin supplements [2]. Systematic and meta-analytic evidence supports antioxidant and multivitamin supplementation and smoking cessation as the modifiable factors with the most evidence for slowing progression [31]. Adherence to a Mediterranean-style diet has been associated with reduced AMD development and progression in observational meta-analysis [32], and reviews of modifiable lifestyle factors summarise the current preventive landscape [33]. Importantly for the genetics-versus-behaviour debate, a 2026 prospective analysis reported that adopting healthy lifestyle behaviours offset a substantial portion of high genetic risk for progression to advanced AMD, quantifying the interaction between genetic risk score and lifestyle profile [34]. Metabolic and inflammation-related markers — the metabolic vulnerability index and a metabolomics-based MetaboHealth score — have also been associated with incident AMD in a large prospective cohort, and interacted with polygenic risk on additive and multiplicative scales [35]. These findings supply the clinical rationale for individualised risk: if high-genetic-risk individuals can reduce risk through behaviour, then calibrated individual risk estimates have direct counselling value.
3.5 Genetics: CFH, ARMS2/HTRA1, and polygenic architecture
AMD genetics is foundational to risk prediction because genetic variants are fixed at birth and therefore represent the earliest available prognostic signal. The first robust genetic association was the complement factor H (CFH) polymorphism, identified in a 2005 genome-wide screen: homozygosity for the risk allele increased the likelihood of AMD by a factor of 7.4 (95% CI 2.9–19), with a tyrosine-to-histidine change at amino acid 402 (Y402H) in linkage disequilibrium with the risk allele [36]. More than 50 genetic susceptibility loci have since been identified, with the most important in CFH and ARMS2 [2]. In the AREDS cohort, the strongest single-variant associations with prevalent advanced AMD were with CFH Y402H (odds ratio 3.5), CFH rs1410996 (3.7), LOC387715/ARMS2 A69S (25.4), C2 E318D (0.3, protective), CFB (0.3, protective), and C3 R102G (3.6) [37].
Because single variants explain little risk individually, composite genetic risk scores (GRS) and polygenic risk scores (PRS) were developed. Reviews published in 2014 and 2016 summarised the numerous genetic-risk models for AMD and their strengths and limitations [38,39]. A 2024 multi-ancestry genome-wide association study — the first to include substantial African and Hispanic/Latino ancestry data, combining the Million Veteran Program with five other cohorts — discovered 63 loci (30 novel) and observed marked cross-ancestry heterogeneity: African-ancestry populations showed a primary signal in an MHC class II haplotype and reduced risk at the established CFH and ARMS2/HTRA1 loci, with significantly smaller marginal effect sizes for CFH risk alleles on African-ancestry haplotypes [40]. This ancestry dependence is directly relevant to the equity and transportability concerns of prediction models built on European-ancestry cohorts.
Genetic architecture varies by outcome: a 2024 genome-wide meta-analysis identified risk loci and improved disease prediction, and a 2024 report evaluated the predictive performance of an updated PRS for AMD [41,42]. Population- and clinic-specific studies have validated GRS in diverse settings, including Taiwan [43], Sardinia [44], Poland [45], Ireland [46], the Coimbra eye study [47], and Israel [48], with further analyses of GRS within AMD classification systems [49] and of rare-variant burden [50]. Reviews have cautioned that predictive genetics for AMD has been over-hyped relative to demonstrated clinical utility [51].
Two biological nuances matter for prognosis. First, genotype and phenotype are not redundant: including genotype with fundus phenotype improved the accuracy of predicting conversion to both choroidal neovascularisation and geographic atrophy in AREDS [52]. Second, genetic variants can modify treatment response: CFH and ARMS2 polymorphisms have been reported to predict response to antioxidants and zinc in AREDS participants [53], and a polygenic risk score has been associated with one-year outcomes after as-needed aflibercept for exudative AMD [54]. These gene–treatment interactions hint at a future in which prognosis informs therapy selection (inference).
3.6 Current treatment landscape and its prognostic implications
The treatment context shapes what prognostic models are for. Intravitreal anti-VEGF therapy is highly effective for neovascular AMD and has substantially reduced AMD-related visual impairment [2]. For geographic atrophy, the therapeutic landscape changed with complement inhibition. Pegcetacoplan, a complement C3 inhibitor, was evaluated in a phase 2 trial [55] and then in the phase 3 OAKS and DERBY trials, which randomised patients with GA to intravitreal pegcetacoplan monthly or every other month versus pooled sham; the primary endpoint was change from baseline to month 12 in total GA lesion area [16]. Avacincaptad pegol, a C5 inhibitor, was tested in the phase 2/3 GATHER1 trial, where the reduction in the mean rate of GA growth over 12 months was 27.4% (P=0.0072) for the 2 mg cohort and 27.8% (P=0.0051) for the 4 mg cohort versus sham [56], and in the phase 3 GATHER2 trial, which evaluated monthly avacincaptad pegol 2 mg against sham over 24 months [17]. By contrast, the complement factor D inhibitor lampalizumab failed in the phase 3 CHROMA and SPECTRI trials despite a supportive phase 2 signal, illustrating that even mechanism-based therapies can fail late [57]. Systematic reviews and living reviews of complement inhibition for AMD and GA synthesis these results [58,59], and a Cochrane review appraises complement inhibitors for AMD [60]. Deep-learning analysis of OCT has been used to identify disease activity and therapeutic response to pegcetacoplan, and GA progression in clinical practice has been studied before and after treatment [61,62]. The clinical relevance of prognostic modelling is sharpened by the arrival of treatments that modify progression: risk stratification can identify who to treat, when, and with what intensity (inference).
3.7 A synthetic view of natural history
Taken together, the natural-history evidence supports three propositions that frame the rest of this review. (1) Progression from early/intermediate to late AMD is common and clinically consequential, with annual incidences on the order of 0.19%–1.4% for late AMD depending on population and ascertainment, and substantially higher in enriched cohorts [10,8]. (2) Progression risk is heterogeneous and is driven by a combination of age, baseline severity, drusen burden and morphology, pigmentary change, genetic load, and modifiable exposures [20,19,63]. (3) Because both eyes and repeated visits contribute data, the statistical dependence structure (inter-eye and longitudinal correlation) must be handled explicitly, a methodological point emphasised in biostatistical guidance on correlated eye data [64]. A prognostic model that ignores these dependencies risks both biased performance estimates and over-confident predictions (inference).
4. Appraising Prognostic Models: The Methodological Frame
4.1 Discrimination and calibration are different things
The methodological literature is unambiguous that a useful prognostic model must be characterised by both adequate discrimination — the ability to separate patients who will have an event from those who will not — and adequate calibration — agreement between predicted and observed absolute risks [65]. Discrimination is commonly quantified by the area under the receiver-operating-characteristic curve (AUC) or the concordance (c-) statistic. Calibration is quantified by calibration plots, calibration slope, calibration-in-the-large, observed-to-expected ratios, or the Brier score. A model can discriminate well while being poorly calibrated, and a poorly calibrated model can mislead clinical decisions even when its AUC is high [66]. Because decisions about surveillance intensity, referral, or trial eligibility depend on calibrated absolute risk, calibration is arguably the more decision-relevant property — a point made pointedly in the argument that calibration is “the Achilles heel of predictive analytics” [66]. Practical guides for clinicians on interpreting discrimination and calibration, and on assessing performance of risk models, have been published for general audiences [65,67].
4.2 Internal versus external validation, and the limits of “validated”
All modelling studies require internal evaluation to guard against overfitting; the recommended approaches are bootstrapping and cross-validation. External validation in a genuinely independent setting — different health system, device, or demographic composition — is the stronger test, and is especially important for imaging models trained on a single camera or grading pipeline. The methodological literature shows that external validation of new risk models is infrequent and often reveals worse discrimination than at development [68]. One influential commentary argues that there is “no such thing as a validated prediction model,” because patient populations, measurement procedures, and populations themselves change over time; the implication is that validation should be an ongoing, quantifiable process and that recalibration and updating are the norm rather than the exception [69]. This framing is directly applicable to AMD: models trained on AREDS fundus photographs may not transfer to different OCT devices or to multi-ancestry populations without recalibration [1].
4.3 Reporting and risk-of-bias instruments: TRIPOD, PROBAST, and their AI extensions
Two instruments anchor contemporary appraisal. TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) is a reporting checklist developed because the quality of reporting of prediction-model studies was generally poor [70]. PROBAST (Prediction model Risk Of Bias ASsessment Tool) assesses risk of bias and applicability across four domains — participants, predictors, outcome, and analysis — via 20 signalling questions, with risk of bias defined as shortcomings in design, conduct, or analysis that systematically distort estimates of predictive performance [71,72]. Because machine-learning models raise distinct concerns (for example, feature leakage, opaque preprocessing, and data-splitting that violates patient-level independence), extensions to both instruments were developed: TRIPOD-AI and PROBAST-AI [73]. Related reporting guidelines for AI clinical trials and decision-support systems include CONSORT-AI and SPIRIT-AI for trials and protocols, and DECIDE-AI for early-stage clinical evaluation [74,75,76,77]. For AMD specifically, the anchor review used PROBAST and reported risk-of-bias findings that are summarised in Section 7 [1].
4.4 Clinical utility, transportability, and overfitting
Discrimination and calibration describe statistical performance; they do not by themselves establish clinical value. Decision-curve analysis expresses clinical utility as net benefit across a range of threshold probabilities, explicitly trading off benefits and harms of decisions taken on the basis of a prediction [78]. Model transparency and validation — including the need to state assumptions, structure, and data sources — is a recognised good-practice requirement for model-based analyses [79]. Transportability of estimated effects between populations is a recognised methodological problem with its own literature [80], and the anchor review draws on it explicitly when discussing whether AMD models can move between populations [1]. Overfitting is a pervasive threat when event counts are small relative to candidate predictors; the analysis of events per variable, use of shrinkage, and clarity about candidate-predictor numbers and internal-validation procedures are essential, and their incomplete reporting was a recurring risk-of-bias concern in AMD models [1].
| Instrument / method | Purpose | Relevance to AMD prognosis | Ref |
|---|---|---|---|
| TRIPOD | Reporting checklist for prediction-model studies | Standard against which reporting completeness is judged | [70] |
| TRIPOD-AI | AI extension of TRIPOD | Applies to imaging/ML AMD models | [73] |
| PROBAST | Risk-of-bias and applicability tool (4 domains, 20 questions) | Used by the 2026 anchor review to grade 20 AMD models | [71,72] |
| PROBAST-AI | AI extension of PROBAST | Applies to ML/imaging AMD models | [73] |
| Discrimination (AUC/c-statistic) | Separating events from non-events | Most-reported metric in AMD models | [65] |
| Calibration (slope, in-the-large, plots, Brier) | Agreement of predicted vs observed absolute risk | Reported in only 6/20 anchor studies | [66,65] |
| External validation | Independent-setting performance | Reported in 7/20 anchor studies | [68,69] |
| Decision-curve / net benefit | Clinical utility across thresholds | Rarely reported in AMD models | [78] |
| Correlated-eye/longitudinal handling | Correct variance and dependence | Essential for eye-level AMD data | [64] |
| CONSORT-AI / SPIRIT-AI / DECIDE-AI | Reporting for AI trials and decision support | Relevant to prospective AMD AI trials | [74,77] |
5. Core Evidence I: Regression- and Risk-Score-Based Models
Before machine learning, AMD prognosis was the domain of multivariable regression and points-based risk scores derived from large cohort studies. This body of work remains the methodological backbone of the field and, as the anchor review observed, remains competitive in performance [1].
5.1 AREDS-derived genetic and clinical models (Seddon and colleagues)
The most sustained programme of AMD risk modelling has been built on AREDS. In 2009, a prediction model based on genetic, demographic, and environmental variables in 1,446 AREDS participants (279 progressors over 6.3 years) reported that all evaluated genetic polymorphisms were independently related to prevalent advanced AMD and that, for incident AMD, all variants except CFB were associated with progression after adjustment; the reported multivariable odds ratios included CFH Y402H 3.5, CFH rs1410996 3.7, ARMS2 A69S 25.4, C2 0.3, CFB 0.3, and C3 3.6 [37]. In 2011, expanded time-varying survival models in 2,937 AREDS participants (819 progressors over 12 years) added age, smoking, body mass index, single-nucleotide polymorphisms in CFH, ARMS2/HTRA1, C3, C2, and CFB, presence of advanced AMD in one eye, and drusen size in both eyes, and estimated the sample sizes needed to test new treatments with and without genetic variables [81]. In 2013, the model was validated for progression to advanced AMD subtypes, achieving C-statistics of 0.858 in the derivation sample and 0.750 in an independent validation cohort at five years — a notable drop that presaged the field’s broader external-validation concerns [82]. In 2015, risk prediction incorporating ten common and rare genetic variants — including the newly identified rare variant C3 K155Q (hazard ratio 1.7) — reported an AUC of 0.911 for progression to advanced AMD over ten years, with split-sample validation at 0.907, and attributable-risk calculations suggesting that 80% of incident AMD was attributable to genetic factors after adjusting for demographic covariates and baseline macular phenotype [63]. A companion 2019 study validated prediction models in an independent cohort, reporting an age-adjusted AUC of 0.900 over 12 years for a composite 13-locus model (0.896 in validation), and demonstrated that eyes with the same baseline macular grade had a wide range of estimated progression probabilities [83]. A 2026 analysis from the same programme quantified how healthy lifestyle behaviours offset high genetic risk of progression to advanced AMD [34].
5.2 The macular risk scoring system and other points-based scores
Chiu and colleagues developed an eye-specific clinical prediction model — the macular risk scoring system (MRSS) — using eight baseline predictors (age, sex, education, race, smoking, pigment abnormality, soft drusen, and maximum drusen size) trained in 4,507 AREDS participants and validated in 2,169 participants of the Blue Mountains Eye Study. The internally validated c-index was 0.88 and the externally validated c-index was 0.91, with 87.6% sensitivity and 73.6% specificity at the optimal cut-off [84]. The MRSS is notable as one of the few AMD models in this literature with an external validation showing improved rather than degraded discrimination. A separate risk score was developed for geographic atrophy within the Complications of AMD Prevention Trial (CAPT): among 942 participants followed five years, 5-year incidence of endpoint GA ranged from 0.6% for a score below 7 points to 15% for a score of 12 or more, with an AUC of 0.76 [85].
5.3 Population-based models: the Three Continent AMD Consortium
A recurring criticism of AMD risk models is that case-control designs tend to overestimate risk. The Three Continent AMD Consortium addressed this by developing a prediction model in three population-based studies (Rotterdam Study, Beaver Dam Eye Study, and Blue Mountains Eye Study) comprising 10,106 people without late AMD at baseline. The best model included age, sex, 26 single-nucleotide polymorphisms in AMD risk genes, smoking, body mass index, and baseline AMD phenotype, achieving an AUC of 0.88 in the Rotterdam Study and 0.85 in the validation cohorts, with a risk score derived from the pooled beta coefficients [86]. This remains among the strongest population-based evidence that genetic and phenotypic information can be combined into a calibrated, externally validated risk score.
5.4 Machine learning on structured and clinical data
Machine learning is not confined to imaging. Ajana and colleagues developed a comprehensive prediction model in the Rotterdam Study (3,838 participants, 108 incident advanced AMD cases, median 10.8 years) using bootstrap LASSO survival analysis, with external testing in the ALIENOR study. The model retained age, phenotypic predictors (intermediate drusen, hyperpigmentation in one or both eyes, and the AREDS simplified score), a genetic risk score based on 49 single-nucleotide polymorphisms, smoking, diet quality, education, and pulse pressure. Cross-validated AUC was 0.92 at five years, 0.92 at ten years, and 0.91 at fifteen years in the training cohort, and 0.92 at five years in the external cohort; the model tended to underestimate cumulative incidence (a calibration limitation) [87]. This study is important because it demonstrates that a broad structured-data model can achieve externally validated discrimination of 0.92 in a population-based setting, while still exposing a calibration shortfall — precisely the discrimination-versus-calibration tension of Section 4.1.
Other structured-data models have extended the approach to specific populations and data types. A Korean risk prediction model based on systemic and environmental factors in 10,890 individuals reported a C-statistic of 0.84 [88]. The STARS questionnaire, a simple self-administered tool derived in 12,639 Italian patients and validated in 6,897 French patients, achieved AUCs of 0.78 and 0.72 respectively [89]. Regression models of choroidal neovascularisation and geographic atrophy incidence in AREDS used clinical, environmental, demographic, and genetic factors to reach AUCs of 0.87 (CNV) and 0.89 (GA), but found that adding genetic markers provided only marginal improvement over simple clinical criteria for screening patients into prevention trials — a finding with direct implications for the incremental value of genotyping [90]. A 2025 proposal for a simpler eye-level model incorporating reticular pseudodrusen reported that the updated simplified AREDS model could be halved without compromising prognostic performance, achieving AUCs up to approximately 84.5% at one to three years and producing an interpretable one-point-per-biomarker score with example three-year risks of roughly 4%, 8%, 16%, 32%, and 64% across scores 0–4 [91].
5.5 Geographic-atrophy-specific and cohort-specific models
Progression to geographic atrophy has attracted dedicated modelling. Retrospective work on fundus autofluorescence patterns and GA progression has been reviewed and pooled, and progression-rate patterns have been characterised [92,93]. Deep-learning-based growth prediction models have been developed for GA, including automated growth prediction using quantitative SD-OCT biomarkers and topographic analyses of the contribution of imaging features to GA growth prediction [94,95]. A deep-learning model has been used to identify disease activity and therapeutic response to pegcetacoplan, connecting prognostic modelling to treatment [61]. Mathematical and machine-learning models of GA progression have also been critically examined for epistemic uncertainty, cautioning that different model families can yield different forecasts of the same progression process [96].
| Model (first author, year) | Cohort / n | Outcome & horizon | Discrimination | Validation | Calibration |
|---|---|---|---|---|---|
| Seddon 2009 [37] | AREDS, n=1,446 | Incident advanced AMD, 6.3 y | OR-based risk score | Internal | Not reported quantitatively |
| Seddon 2011 [81] | AREDS, n=2,937 | Advanced AMD, 12 y | AUC computed (derivation/test) | Random split | Reported descriptively |
| Seddon 2013 [82] | Derivation 2,914; validation 980 | Advanced AMD subtypes, 5 y | C=0.858/0.750 | Independent cohort | Reported |
| Seddon 2015 [63] | AREDS, n=2,951 | Advanced AMD, 10 y | AUC 0.911 (0.907 split) | Split-sample | Online calculator |
| Seddon 2019 [83] | AREDS + external | Advanced AMD, 12 y | AUC 0.900 (0.896 ext) | External | Calibrated |
| Chiu 2014 (MRSS) [84] | AREDS 4,507; BMES 2,169 | Advanced AMD, 8–10 y | c=0.88 (int)/0.91 (ext) | External | Cut-off sensitivity/specificity |
| Buitendijk 2013 (3CC) [86] | 10,106 (3 cohorts) | Late AMD, ~11 y | AUC 0.88/0.85 | External | Risk score |
| Ying 2011 (CAPT GA) [85] | CAPT n=942 | GA, 5 y | AUC 0.76 | Internal | Risk gradient 0.6%–15% |
| Ajana 2021 (ML) [87] | RS-I 3,838; ALIENOR 362 | Advanced AMD, 5–15 y | AUC 0.92 (ext 0.92) | External | Underestimates absolute risk |
| Sitnilska 2020 [97] | n=232 | Late AMD, 5.9 y | AUC 0.969 | Internal | Hosmer–Lemeshow p=0.797 |
| McCarthy 2012 (CNV/GA) [90] | AREDS | CNV/GA, 3 y | AUC 0.87/0.89 | Independent | — |
| Trinh 2025 eye-level [91] | n=269 | Late AMD, 1–7 y | AUC up to ~0.845 | Internal | Risk gradient 4%–64% |
| Korean model [88] | n=157 followed | AMD progression, 4.4 y | C=0.84 | Internal | — |
| STARS questionnaire [89] | 12,639 (IT) / 6,897 (FR) | AMD, cross-sectional | AUC 0.78/0.72 | External | Stratum reclassification |
(Discrimination figures are as reported by each source and span different outcomes, horizons, and analytic units; they are not directly comparable — see Section 7.)
5.6 What the regression literature established
Four conclusions can be drawn from the pre-deep-learning evidence base. First, progression to late AMD is prognostically tractable using routine clinical and phenotypic data, with good discrimination achievable without imaging AI [84,86]. Second, genetic information improves risk stratification and reclassification but generally adds only modest incremental discrimination over a strong phenotypic model, and its incremental value depends on the clinical decision and trial economics [90,63]. Third, external validation frequently reveals degradation relative to development, as dramatically illustrated by the 2013 Seddon model’s drop from C=0.858 to 0.750 [82], and calibration is often weak or unreported [87]. Fourth, the retained predictors — age, baseline severity, drusen size/area, pigmentary abnormalities, and smoking — are stable across studies, suggesting a common biological core [1].
6. Core Evidence II: Imaging and Machine-Learning/Deep-Learning Models
The most visible growth in AMD prognosis has come from machine learning applied to retinal images. The anchor review found that 11 of 20 included models used machine-learning or deep-learning approaches applied to retinal imaging — fundus photography, OCT volumes, or derived imaging biomarkers — while seven used traditional regression or survival models and two applied machine learning to structured, non-imaging data [1]. This section surveys the principal model families.
6.1 Fundus-photograph deep learning and AMD severity
The first wave of fundus deep learning addressed staging and severity, because the AREDS severity scale is complex and depends on trained graders. Burlina and colleagues trained deep convolutional neural networks to reproduce the AREDS 9-step detailed severity scale and to estimate five-year risk of progression from 67,401 colour fundus images of 4,613 AREDS participants, reporting weighted kappa statistics and mean unsigned errors for five-year risk estimation [98]. A related model classified 120,656 graded images from 3,654 AREDS participants into 13 classes (nine AREDS steps, three late stages, and a non-gradable category), achieving a quadratic weighted kappa of 92% in the AREDS test set and illustrating the difficulty of generalisation when tested on population-based images [99]. DeepSeeNet took a patient-based approach, simulating the human grading process by detecting drusen size and pigmentary abnormalities in each eye and then computing the AREDS Simplified Severity Scale; trained on 58,402 and tested on 900 images from 4,549 AREDS participants, it achieved patient-based accuracy of 0.671 (kappa 0.558) versus 0.599 (kappa 0.467) for retinal specialists, with AUCs of 0.94 for large drusen, 0.93 for pigmentary abnormalities, and 0.97 for late AMD [100]. These models are diagnostic/severity models rather than prognostic models in the strict sense, but they establish the technical foundation and, in the case of the Burlina model, directly estimate progression risk [98,1].
6.2 OCT-based prediction of conversion
The richest prognostic vein has been OCT-based prediction of conversion to neovascular AMD or atrophy. Schmidt-Erfurth and colleagues, in a study of 495 eyes with intermediate AMD (159 converted within two years), developed a machine-learning model using automatically segmented outer-retinal layers, drusen, and hyperreflective foci; the model achieved an AUC of 0.68 for neovascular conversion and 0.80 for geographic atrophy, and identified outer-retinal thickness, hyperreflective foci, and drusen area as the most critical quantitative features — with distinct feature patterns for the neovascular and atrophic pathways [26]. Russakoff and colleagues trained two convolutional networks on baseline OCT from 71 eyes (40 non-converters, 31 converters over two years), with a segmentation-based preprocessing step that improved performance; their AMDnet achieved an AUC of 0.89 at the B-scan level and 0.91 at the volume level [101]. de Sisternes and colleagues used quantitative SD-OCT drusen features (number, morphology, reflectivity) from 2,146 scans of 330 eyes to predict conversion to exudative AMD, reporting a mean AUC of 0.74 for progression at follow-up (mean 6.16 months) [102].
The field’s most cited single result is arguably Yim and colleagues’ deep-learning system for predicting conversion to exudative (wet) AMD in the second eye of patients already diagnosed in one eye, combining three-dimensional OCT images with automatic tissue maps to predict conversion within a six-month window. Reported performance was 80% sensitivity at 55% specificity and 34% sensitivity at 90% specificity, corresponding to true positives in 78% and 41% of eyes and false positives in 56% and 17% of eyes at those operating points; the system outperformed five of six expert observers and overcame substantial interobserver variability [103]. Yan and colleagues jointly used genotypes and fundus images — 31,262 images and 52 AMD-associated genetic variants from 1,351 AREDS subjects over 12 years — in a modified deep convolutional neural network, reporting an average AUC of 0.85 for predicting progression to late AMD with images plus genotypes versus 0.81 with fundus images alone [104]. Hallak and colleagues, in a post hoc secondary analysis of the HARBOR trial, evaluated imaging, genetic, and demographic factors associated with conversion to neovascular AMD in 686 fellow eyes, finding that 22.4% converted and that automatically extracted drusen features plus genetic and demographic variables were associated with conversion [105]. Nassisi and colleagues evaluated OCT features in 501 HARBOR fellow eyes and found intraretinal hyperreflective foci, hyporeflective foci within drusenoid lesions, subretinal drusenoid deposits, and drusen volume of at least 0.03 mm³ to be associated with progression to late AMD [106].
6.3 Sequential and longitudinal models
Because AMD progression unfolds over visits, the most recent work explicitly models longitudinal sequences. Banerjee and colleagues proposed “Deep Sequence,” integrating radiomics-engineered OCT features, demographics, and visual factors in a recurrent neural network to predict exudation risk in non-exudative AMD fellow eyes, trained on 671 eyes and 13,954 observations. It achieved 0.96 ± 0.02 AUC by cross-validation for exudation within three months and 0.97 ± 0.02 within 21 months, but when trained on the trial dataset and tested on an external real-world clinical dataset, performance fell to 0.82 AUC at three months and 0.68 AUC at 21 months — a vivid demonstration of the transportability problem [107]. Lee and colleagues evaluated CNN-LSTM and CNN-Transformer networks on longitudinal fundus photographs from AREDS, reporting AUCs of 0.879 versus 0.868 for two-year prediction and 0.879 versus 0.862 for five-year prediction compared with single-visit baselines [108]. A temporal-correlation-structure-guided model (LONGL-Net) reported average AUC of 0.905 and accuracy 0.762 on simultaneous grading and late-AMD prediction in about 30,000 AREDS images, with validation on UK Biobank images [7]. CAT-LSTM, combining contrastive attention with a time-aware LSTM and genotype/demographic data, reported 0.925 AUROC for five-year late-AMD prediction, outperforming prior methods by more than 3% [109]. Morph-SSL, a self-supervised longitudinal method, achieved an AUC of 0.779 for predicting conversion to neovascular AMD within six months [110]. A 2020 model using longitudinal fundus images with uneven time intervals and InceptionV3 feature extraction plus interval scaling and a recurrent network reported test sensitivity 0.878, specificity 0.887, and AUC 0.950 for progression to late AMD in 4,903 AREDS eyes [111].
6.4 Geographic-atrophy and atrophy-progression models
GA prediction has developed its own modelling literature. DeepGAze, a position-aware convolutional network, predicted progression from intermediate AMD to GA within 13 months from volumetric SD-OCT scans, trained on AREDS2 ancillary SD-OCT data and validated on two external clinical datasets [112]. Deep-learning quantification of hyperreflective foci — two-thirds localised within 1 mm of the GA border — was positively correlated with GA progression in unifocal and multifocal GA and with de novo GA development, while global progression speed did not correlate with HRF concentration [113]. A deep-learning model segmented GA to study long-term natural history [114], and deep survival modelling of longitudinal OCT volumes predicted onset of atrophy in intermediate AMD, combining time-to-event and deep learning advantages and improving risk estimation over standard classification models across internal (231 eyes) and external (280 eyes) datasets [115]. Automated prediction of GA growth using quantitative SD-OCT biomarkers [94] and topographic ablation analyses identifying the GA rim as the most informative region for deep-learning growth prediction [95] characterise the technical state of the art. Complementary work has examined photoreceptor degeneration in GA, fellow-eye status as a biomarker of GA progression rate, and FAF pattern reclassification based on progression rate [116,93,92].
6.5 Multimodal and multi-source models
The newest models combine modalities and data sources. A 2025 study compared classical machine learning and deep learning to predict imminent exudative conversion within six months using 33,189 OCT volumes from 2,084 patients plus clinical data, with a multimodal perceptron framework integrating a ResNet-50 feature encoder with tabular clinical variables [117]. A 2026 study used deep learning to quantify drusen and reticular pseudodrusen volumes on OCT in 51 eyes with reticular pseudodrusen and contralateral neovascular AMD, finding that each 0.01 mm³ increase in baseline reticular pseudodrusen volume (hazard ratio 1.082) and total drusen volume (hazard ratio 1.080) was associated with progression to late AMD over about 24 months [118]. A 2026 deep-learning study predicted genotype from retinal fundus images in eyes with AMD [119], and a 2024 study explored associations between retinal morphology and AMD polygenic risk score [120]. An “optimized prediction models” study combined fundus imaging and genetics for late AMD [121], and a machine-learning model incorporating both GWAS-derived features and clinical variables has been developed for AMD risk [122]. Genome-wide association study–based deep learning has also been applied to survival prediction in AMD [123] (context).
The I-SCREEN project illustrates the move toward real-world, community-based AI infrastructure: a pan-European programme across six countries using OCT for early AMD detection and progression-risk assessment, spanning optometry practices and ophthalmology clinics with linked longitudinal studies [124]. For anti-VEGF prognosis, the 2026 KongMing study developed and validated a transformer-based multitask deep-learning model to predict visual and anatomical prognosis after anti-VEGF therapy in neovascular AMD across 18 tertiary hospitals in China [125]. Generative models have also been compared for predicting 12-month neovascular AMD progression, with Pixel-to-Pixel HD achieving the highest image-quality metrics but expert discrimination of synthetic images only marginally above chance [126]. A 2021 systematic review evaluated the predictive capabilities of AI-based OCT analysis for AMD progression [127], and a 2022 review summarised SD-OCT biomarkers and AI for predicting progression from intermediate to advanced AMD [128].
| AI/DL model (first author, year) | Data / n | Outcome & horizon | Reported performance | External validation |
|---|---|---|---|---|
| Burlina 2018 [98] | 67,401 CFP, 4,613 AREDS | AREDS 9-step severity; 5-y risk | Weighted κ; 5-y risk error | No |
| Peng 2019 (DeepSeeNet) [100] | 58,402 train / 900 test | AREDS simplified scale | Accuracy 0.671 (vs 0.599 specialists) | No |
| Schmidt-Erfurth 2018 [26] | 495 eyes (159 convert) | CNV/GA conversion, 2 y | AUC 0.68 (CNV), 0.80 (GA) | No |
| Russakoff 2019 [101] | 71 eyes | Conversion, 2 y | AUC 0.89 (B-scan), 0.91 (vol) | No |
| Yim 2020 [103] | Fellow-eye OCT | Exudative conversion, 6 mo | 80% sens@55% spec; 34%@90% | Related cohort |
| Yan 2020 [104] | 31,262 images, 1,351 subjects | Late AMD, 12 y | AUC 0.85 (img+geno), 0.81 (img) | No |
| Peng 2020 [6] | AREDS/AREDS2, 3,298 | Late AMD, 5 y | C=0.864 vs 0.813/0.820 specialists | AREDS2 external |
| Banerjee 2020 [107] | 671 eyes, 13,954 obs | Exudation 3/21 mo | 0.96/0.97 CV; 0.82/0.68 external | Yes (degradation) |
| Bhuiyan 2020 [129] | 116,875 CFP; NAT-2 | Late AMD, 1–2 y | 86.36% acc (2 y); 84% NAT-2 | External |
| Wu 2021 [130] | 280 eyes | Late/atrophic AMD | AUC 0.80 CFP / 0.82 OCT / 0.85 combined | Internal |
| DeepGAze 2023 [112] | AREDS2 + 2 clinical sets | GA within 13 mo | AUROC/AUPRC (position-aware CNN) | 2 external |
| Rivail 2023 [115] | 231 + 280 eyes | Onset of atrophy | Improved over classification | External |
| Lee 2022 [108] | Longitudinal AREDS CFP | Late AMD 2/5 y | AUC 0.879 | Internal |
| LONGL-Net 2022 [7] | ~30,000 AREDS CFP | Late AMD | AUC 0.905 | UK Biobank |
| CAT-LSTM 2022 [109] | AREDS longitudinal | Late AMD, 5 y | AUROC 0.925 | Internal |
| Morph-SSL 2024 [110] | 399 eyes SSL / 343 eval | nAMD conversion, 6 mo | AUC 0.779 | Internal |
| Bridge 2020 [111] | 4,903 AREDS eyes | Late AMD | Sens 0.878 / Spec 0.887 / AUC 0.950 | No |
| Multimodal 2025 [117] | 33,189 OCT, 2,084 patients | Exudative conversion, 6 mo | AUC (MLP multimodal) | Internal |
| RPD/drusen volume 2026 [118] | 51 eyes | Late AMD, 24 mo | HR 1.082/1.080 per 0.01 mm³ | No |
| KongMing 2026 [125] | 18 hospitals, China | Anti-VEGF visual/anatomical | Transformer multitask model | Prospective multicentre |
6.6 Interpreting the deep-learning results
Several interpretive cautions follow from this evidence. First, reported performances span an enormous range (roughly 0.65–0.97) and are measured on different tasks; a model predicting six-month exudative conversion in an enriched fellow-eye cohort is not comparable with a model predicting ten-year incident advanced AMD in a population cohort [1]. Second, the highest AUCs (above ~0.92) tend to occur in selected clinical or imaging cohorts with narrow case mix, and can reflect restricted case mix or leakage rather than genuine prognostic superiority [1]. Third, external transfer frequently degrades performance, as seen in the Deep Sequence model’s fall from 0.96–0.97 by cross-validation to 0.82/0.68 in a real-world dataset [107]. Fourth, because the dominant retained imaging features (drusen, pigmentary change, outer-retinal thinning, hyperreflective foci) overlap heavily with markers of current disease severity, some imaging models may be automating severity assessment rather than learning independent prognostic information — an inference the anchor review explicitly raises and which the field has not yet resolved [1]. Fifth, self-supervised and longitudinal methods are promising for exploiting unlabelled longitudinal scans, but their reported gains remain modest and internally validated [110].
7. Core Evidence III: What the 2026 Systematic Review Established
The anchor review [1] is the single most important document for this topic because it is the first to appraise AMD-specific multivariable prognostic models against contemporary prediction-model standards, integrating outcome definitions, prediction horizons, calibration, validation design, and risk of bias. Its findings define the state of the evidence.
7.1 Design and included studies
The review 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. From 2,239 identified records, 1,742 unique titles/abstracts, and 632 full-text assessments, 20 studies met eligibility and were included: 19 from the initial search and one from the updated search. The protocol was registered (PROSPERO CRD42022323522), data were extracted on population, outcomes, horizons, predictors, methods, performance measures, and validation approach, and risk of bias was assessed with PROBAST; because of substantial heterogeneity, findings were synthesised narratively under a prespecified conditional analysis plan [1].
7.2 Cohorts and population diversity
Evidence was concentrated in a small number of recurring datasets: eight studies used AREDS, three used the HARBOR clinical trial, and two (with a third using one as an external cohort) used the LEAD trial; population-based cohorts included the Three Continent AMD Consortium and the Rotterdam Study (with external validation in ALIENOR), and several studies used clinic-based longitudinal OCT datasets, most notably from Moorfields Eye Hospital. Sample sizes ranged from 71 participants in the smallest imaging study to 10,106 at-risk individuals in pooled population-based analyses, with image-level sample sizes up to 67,401 fundus photographs or more than 5,500 OCT scans. Participants were predominantly older adults with mean ages typically in the late 60s to late 70s; race and ethnicity reporting was inconsistent, and when reported most cohorts were predominantly White, particularly AREDS- and HARBOR-based analyses, with only one large OCT cohort reporting substantial non-White representation [1]. The review’s central structural claim follows: this research area “appears broader in publication volume than in underlying population diversity” [1].
7.3 Outcomes and prediction horizons
Most models (16 of 20) predicted progression to late or advanced AMD, typically defined as incident neovascular AMD, geographic atrophy, or both. Four studies focused specifically on short-term exudative or neovascular conversion, occasionally using initiation of anti-VEGF therapy as a proxy outcome. Prediction horizons ranged from as short as three to 18 months for imminent exudative conversion in enriched fellow-eye or clinic cohorts, to as long as 15 years for incident advanced AMD in population-based cohorts. Because progression was defined differently and horizons, case mixes, and analytic units differed, reported discrimination values are not directly comparable [1].
7.4 Model types and predictors
Seven studies used traditional regression or survival models (including points-based risk scores and Cox models); two applied machine learning to structured, non-imaging data; and 11 used machine-learning or deep-learning approaches on retinal imaging (fundus photography, OCT volumes, or derived biomarkers). Two studies explicitly incorporated longitudinal modelling. Predictor domains were dominated by ocular and imaging variables, used far more frequently than demographic, clinical, lifestyle, or genetic predictors. Imaging-based models often used the image itself as the primary input, whereas cohort-based models more commonly used age, baseline AMD severity, smoking status, or genetic risk scores. Baseline ocular/imaging features and age were the most frequently retained predictors, but their independent prognostic contribution was not consistently confirmed in external data, so the review describes them as recurrent rather than confirmed [1].
7.5 Performance: discrimination without calibration
Discrimination (AUC or c-statistic) was the most frequently reported metric, ranging from approximately 0.65 to 0.97 across all eligible reported models, with many estimates between 0.80 and 0.92; when restricted to each study’s declared primary model, discrimination clustered around 0.80–0.92. The highest discrimination estimates (c-statistic above 0.92) occurred in selected clinical or imaging cohorts rather than population-based cohorts, consistent with narrower case mix. Calibration reporting varied: six studies mentioned calibration (30%), but reporting was heterogeneous and often did not meet the review’s criteria for quantitative calibration (calibration plots, calibration slope, calibration-in-the-large, Brier score, or equivalent). Where specified, evaluation relied on goodness-of-fit testing (for example, Hosmer–Lemeshow p=0.797 in one model) or study-specific approaches; most studies reported no quantitative calibration metric. Two studies reported an overall performance measure (a Brier score of 0.09 and an R² for explained variation) [1]. The review’s summary is precise: the literature has been “more consistent at demonstrating signal detection than at establishing reliable absolute risk estimation” [1].
7.6 Validation: development has outpaced confirmation
All included studies reported some form of internal evaluation, but reporting was often incomplete: internal validation used bootstrapping in five studies, cross-validation in six, and random split or holdout in two, while in seven studies the internal validation strategy was not described in sufficient detail to classify. External validation was reported in 7 of 20 studies (35%), but its rigor varied; several studies used related or similarly selected cohorts rather than clearly distinct health systems, devices, or demographic settings; calibration was seldom reported in the external sample; and no study reported recalibration or model updating. The review therefore cautions that “the presence of external validation … should not be equated with confirmed transportability” [1]. Where fully independent external data are unavailable, temporal validation, multisite collaboration, and federated analysis are feasible alternatives, and the reasons for not performing external validation should be reported transparently [1].
7.7 Risk of bias
Using PROBAST, ten of 20 studies were judged at low overall risk of bias, six at moderate risk, three at high risk, and one at unclear risk. Concerns were concentrated in the analysis domain and most frequently reflected limited events relative to the number of candidate predictors or model parameters, incomplete reporting of model development procedures, insufficient description of internal validation, unreported or unjustified handling of missing data, and an emphasis on discrimination at the expense of calibration or other measures of overall performance. Risk of bias in the participants and outcome domains was generally low, with moderate concern arising from incompletely described cohort selection or from outcome ascertainment by image grading without explicit masking or standardised procedures [1].
7.8 The review’s diagnosis and prescription
The anchor review concludes that the field shows “clear technical development but limited methodological maturity,” that conventional models remain competitive with machine learning, and that advances in model development have outpaced progress in model confirmation. Its prescription is a phased, resource-conscious pathway: specify target population, outcome, horizon, and decision context; develop in a representative or linked multisite cohort; prespecify predictors and size the study by expected events and model complexity; handle missing data, clustering, and overfitting with robust internal validation; report discrimination, calibration, uncertainty, and clinical utility; undertake temporal or geographic validation when fully independent data are unavailable; recalibrate or update before wider use; and evaluate transportability across sites, devices, and demographic groups before implementation. Outcome definitions should be standardised, with separate reporting for neovascular AMD, geographic atrophy, and composite late AMD, and fixed-horizon versus time-to-event tasks reported distinctly. The review states plainly that until such staged validation occurs, existing models should be regarded as exploratory rather than clinically actionable [1].
| Anchor-review dimension | Finding (20 studies) |
|---|---|
| Included studies | 20 (19 initial + 1 updated search) |
| Data sources | 8 AREDS; 3 HARBOR; 2–3 LEAD; 3CC; Rotterdam/ALIENOR; Moorfields OCT cohorts |
| Sample size range | 71 participants to 10,106 at risk; up to 67,401 images |
| Outcome definitions | Composite late AMD; neovascular/exudative conversion; GA/atrophy; anti-VEGF initiation proxy |
| Horizons | ~3 months to 15 years |
| Modelling approach | 7 regression/survival; 2 ML on structured data; 11 ML/DL on imaging |
| Discrimination range | ~0.65–0.97 (primary models ~0.80–0.92) |
| Calibration reported | 6/20 (30%); few quantitative |
| Internal validation | Bootstrap 5; cross-validation 6; split 2; unclear 7 |
| External validation | 7/20 (35%); seldom calibrated; no recalibration |
| Risk of bias (PROBAST) | 10 low; 6 moderate; 3 high; 1 unclear |
8. Adjacent Prognostic Tasks
8.1 Predicting anti-VEGF response
A distinct but related prognostic problem is predicting response to anti-VEGF therapy in neovascular AMD, where response is heterogeneous and treatment burden is high. Reviews have summarised the predictors of anti-VEGF treatment response and the morphological predictors used to modify treatment protocols [131,132,133]. Predictive biomarkers of response have been catalogued [134], and OCT features such as hyperreflective foci have been associated with treatment outcomes [135]. Baseline predictors for subretinal fibrosis have been identified [136]. Machine-learning and deep-learning models have been developed to predict treatment requirements, future visual change, and treatment burden, and to rank predictors of response [137,138,139,140]; a big-data analysis used an ML approach to predict anti-VEGF requirements in neovascular AMD [137]. Optical coherence tomography angiography has been used to build predictive models for treatment decisions [141], and a polygenic risk score has been associated with one-year outcomes after as-needed aflibercept [54]. Generative adversarial networks have been proposed to predict treatment response, with commentators questioning their clinical utility [142] (commentary). The KongMing transformer model represents the current high-water mark for anti-VEGF prognosis, validated prospectively across 18 hospitals [125]. Systematic and methodological reviews emphasise the need for careful validation and for distinguishing prediction of anatomy from prediction of function [143].
8.2 Home monitoring and early detection as a prognostic bridge
Prognostic models aim to identify who needs closer surveillance; home-monitoring technologies operationalise that surveillance. The HOme Monitoring of the Eye (HOME) study randomised participants at high risk of choroidal neovascularisation to the ForeseeHome preferential-hyperacuity-perimetry device plus standard care versus standard care alone, with the goal of detecting conversion earlier than symptom-driven monitoring [144]. Subsequent analyses examined the utility of ForeseeHome for detecting neovascularisation and the broader role of home monitoring in AMD management [145,146,147]. A 2024 study evaluated home-monitoring vision tests for detecting active neovascular AMD [148]. The clinical logic connecting prognosis to monitoring is that a calibrated individual risk estimate could triage which patients warrant home monitoring or intensified clinic follow-up (inference).
8.3 Risk stratification for clinical trials and prognosis-guided enrichment
Prognostic models have an explicit trial-design use case. In the AREDS-based CNV/GA prediction work, Bayesian trial simulations were used to compare the usefulness of predictive models for screening patients into prevention trials, and the incremental benefit of genetic factors depended on trial economics [90]. The PINNACLE programme, a multicentre UK/Austria/Switzerland effort pooling more than 400,000 OCT images with prospective recruitment of intermediate AMD patients, aims to identify and validate imaging biomarkers of progression specifically to enable shorter, more efficient trials [149]. Its systematic review (Pinnacle Report 2) searched 3,229 articles, found only six meeting eligibility, and concluded that study heterogeneity precluded meta-analysis and that poor methodological quality warranted cautious interpretation — the most commonly reported prognostic factors being baseline drusen volume/size (associated with neovascular progression) and outer-retinal thinning (associated with geographic atrophy) [18]. A separate overview of 17 systematic reviews and meta-analyses identified 218 prognostic factors, grouped into 20 distinct types across eight categories, with the modifiable factors having most evidence being antioxidant/multivitamin supplementation and smoking reduction; notably, no review studied social determinants of health as prognostic factors, a critical evidence gap [31].
8.4 Genetic risk scores and their clinical translation
Direct-to-clinic genetic risk testing for AMD has been discussed for over a decade, and genetic risk scores can now be computed from genome-wide data. Reviews have summarised the application and utility of GRS in AMD [39], and studies have demonstrated that GRS adds value over clinical grading stage in predicting progression [150], that GRS can support a joint view of AMD classification systems [49], and that updated PRS improve predictive performance [42]. The GRADE study is a prospective protocol to assess whether polygenic risk scores can guide prevention and monitoring [151]. However, the incremental value over phenotype is modest for many decisions [90], predictive genetics has been criticised as over-hyped [51], and ancestry dependence of effect sizes is substantial [40] — meaning a GRS developed in European-ancestry cohorts cannot be assumed to transport to other populations without ancestry-specific recalibration [1].
9. Data, Cohorts, Equity, and Regulatory Constraints
9.1 The recurring cohorts and their limitations
The dominance of AREDS and HARBOR is both a strength and a constraint. It is a strength because these cohorts are large, longitudinally followed, richly phenotyped, and have standardised grading — properties that make them ideal for model development and internal validation. It is a constraint because models developed in a few datasets with shared ancestry, imaging protocols, and eligibility criteria cannot be assumed to generalise. The anchor review emphasises that enriched trial or specialty-cohort settings offer standardised imaging, tight eligibility, and structured follow-up that can inflate apparent performance and reduce representativeness of routine care [1]. LEAD, the Three Continent Consortium, Rotterdam/ALIENOR, and Moorfields OCT cohorts broaden the evidence somewhat, but the overall picture remains narrow [1].
9.2 Ancestry, representation, and transportability
The ancestry composition of AMD modelling cohorts is a first-order equity and validity issue. Most AREDS- and HARBOR-based analyses are predominantly White [1], while the genetics of AMD differ markedly across ancestries: the 2024 multi-ancestry GWAS found reduced risk at CFH and ARMS2/HTRA1 in African-ancestry populations and a distinct MHC class II signal [40]. Because predictor distributions and baseline risks shift across ancestral groups, models and genetic scores developed in one population require ancestry-specific calibration before use in another [1]. A scoping review of transportability methods provides the methodological vocabulary for this problem [80]. The anchor review’s observation that only one large OCT cohort reported substantial non-White representation underscores how much work remains [1].
9.3 Imaging standards, datasets, and interoperability
Transportability is as much an engineering problem as a statistical one. Different fundus cameras and OCT devices produce images with different characteristics, and models trained on one vendor’s outputs may fail on another. The Joint American Academy of Ophthalmology and Association for Research in Vision and Ophthalmology policy statement argues that the time for digital imaging standards implementation is now [152], and a 2025 review documents the march toward harmonised imaging standards for retinal imaging [153]. The Collaborative Community on Ophthalmic Imaging (CCOI) roadmap for AI in AMD articulates data and deployment requirements [154]. A global review of publicly available ophthalmological imaging datasets documented barriers to access, usability, and generalisability that constrain reproducible model development and external validation [155]. These infrastructure gaps are a concrete reason why external validation is uncommon in AMD prognosis (inference).
9.4 Regulatory and deployment context
Retinal AI has already crossed regulatory thresholds in adjacent tasks, most visibly in diabetic-retinopathy screening, where AI systems have been deployed in real-world multicentre and primary-care settings, with economic evaluations and policy implications [156,157,158]. Reviews of AI in AMD describe the transition from diagnostic accuracy toward clinical application, prognosis, and treatment [159,160,161]. Complementary reporting and appraisal standards (CONSORT-AI, SPIRIT-AI, DECIDE-AI, TRIPOD-AI, PROBAST-AI) give regulators and reviewers the tools to judge whether an AMD prognostic model is ready for clinical evaluation [74,75,76,77,73]. The gap for AMD specifically is that, unlike diabetic-retinopathy detection, no AMD prognostic model has yet completed the prospective, calibrated, externally validated evaluation that deployment would require [1].
10. Synthesis: What the Evidence Supports
10.1 Progression is prognostically tractable
The most robust positive finding of this literature is that progression from non-late to late AMD can be predicted with useful discrimination. Across regression, machine-learning, and deep-learning models, AUC/c-statistics commonly fall in the 0.80–0.92 range, and the anchor review’s full range of approximately 0.65–0.97 indicates that some prediction tasks are easier than others but that the problem is far from intractable [1]. Population-based models such as the Three Continent Consortium (AUC 0.88/0.85) and the Rotterdam/ALIENOR machine-learning model (AUC 0.92) demonstrate that this is achievable in representative populations, not just in enriched cohorts [86,87].
10.2 Conventional models remain competitive
A counter-narrative in the modern literature is that deep learning has displaced traditional prognostic modelling. The anchor review finds the opposite: Cox and logistic models based on clinical phenotype, smoking, demographics, and genetic risk frequently achieved discrimination comparable to more complex approaches, particularly for medium- and long-term prediction, and conventional models retain advantages in transparency, parsimony, and ease of recalibration [1]. For clinical implementation, a slightly less discriminating but well-calibrated and interpretable model may be more useful than a complex system that performs well only under development conditions [1].
10.3 The recurrent predictors form a common biological core
Despite technical heterogeneity, the retained predictors are strikingly similar across model families: age, drusen burden and morphology (including maximum drusen size and drusen volume), pigmentary abnormalities, reticular pseudodrusen, outer-retinal thinning, hyperreflective foci, and selected genetic variants (CFH, ARMS2/HTRA1, C3, C2/CFB), with smoking and diet quality as modifiable contributors [1,20,26,113,118,162,34]. This convergence suggests that models are drawing on a common biological core of AMD progression [1].
10.4 But convergent predictors raise a validity question
The same convergence carries a warning. Because the dominant retained predictors overlap heavily with established markers of current AMD severity, it is unclear in some imaging studies whether the model is learning novel prognostic information or simply automating baseline severity assessment [1]. Diagnostic accuracy and prognostic utility are not the same, and models whose performance is driven by current-stage features may add less than their AUCs suggest (inference).
10.5 Performance figures are not interchangeable
Reported AUCs must be interpreted with the outcome definition, horizon, case mix, and analytic unit in mind. A six-month exudative-conversion model in a fellow-eye cohort, a five-year late-AMD model in a population cohort, and a fifteen-year incident-advanced-AMD model in a genetic cohort answer three different clinical questions [1]. Treating their AUCs as a single leaderboard of “best AMD model” is a category error (inference).
10.6 Calibration and validation, not novelty, are the bottlenecks
The decisive constraints on clinical usefulness are calibration and external validation, not algorithmic novelty. Calibration was addressed in only 6 of 20 studies, external validation in 7 of 20, and recalibration/updating in none [1]. Because decisions depend on calibrated absolute risk and because calibration degrades when baseline risk, imaging practice, referral thresholds, or patient mix change, this is the field’s central deficit [66,69].
10.7 Data and equity are limiting conditions
Finally, the evidence base’s narrow ancestry and cohort composition limits transportability [1,40], imaging-device heterogeneity and dataset-access barriers impede external validation [155,153], and social determinants of health are a documented gap [31]. These are not peripheral; they bound what any current AMD prognostic model can claim.
| Predictor domain | Representative factors | Consistency across models | Note |
|---|---|---|---|
| Demographic / lifestyle | Age, smoking, BMI, diet quality, education | High (age universal; lifestyle in structured models) | Modifiable: smoking, diet |
| Ocular phenotype (CFP) | Drusen size/area/volume, pigmentary abnormalities, RPD | Very high | Overlaps current-severity markers |
| OCT imaging | Outer-retinal thinning, hyperreflective foci, SDD, drusen volume, EZ/RPE loss | High in imaging models | Rim/junctional zone informative for GA |
| Genetic | CFH (Y402H), ARMS2/HTRA1, C3, C2/CFB; GRS/PRS | Moderate–high; increments modest | Ancestry-dependent |
| Functional | Rod-mediated dark adaptation, microperimetry, low-luminance | Lower / emerging | Potential incremental value |
| Metabolic / systemic | MVX, MetaboHealth, lipids, complement measures | Emerging | Interaction with PRS reported |
11. LIMITATIONS
This review has important limitations that should temper its conclusions.
Review-level limitations. This is a structured narrative synthesis, not a preregistered systematic review or meta-analysis. It was conducted headless, without duplicate independent screening or extraction, and its search, while broad (20 documented E-utilities queries plus snowballing of 16 landmark papers and ClinicalTrials.gov sweeps), was not exhaustive across Embase, Web of Science, or grey literature to the standard of the anchor review. Statistical synthesis was not attempted because the anchor evidence base is heterogeneous by construction [1]. Selection of cited sources, while driven by search and relevance scoring, involved judgment and is therefore not reproducible in the sense a systematic review would require. No structured statistics were extracted from figures or full texts beyond abstracts and the anchor review’s text, so some finer-grained performance or calibration numbers reported only in full-text tables were not captured.
Limitations of the underlying evidence. The evidence base itself carries limitations that no review can overcome. Studies are drawn from few, recurring, predominantly European-ancestry cohorts (AREDS, HARBOR, LEAD, Three Continent, Rotterdam/ALIENOR, Moorfields), so population diversity is limited and transportability is uncertain [1]. Outcome definitions vary (composite late AMD, neovascular conversion, geographic atrophy, anti-VEGF initiation as a proxy), prediction horizons range from months to more than a decade, and analytic units vary (eyes, patients, scans), so reported discrimination values are not directly comparable [1]. Calibration is reported in a minority of studies and rarely quantitatively; external validation is uncommon and, when present, often uses related rather than genuinely independent cohorts; and no included study reported recalibration or model updating [1]. Overfitting cannot be excluded in several studies because event counts, candidate-predictor numbers, shrinkage, and internal-validation procedures were incompletely reported [1]. Risk of bias is concentrated in the analysis domain [1]. For anti-VEGF and home-monitoring evidence, much of the relevant work concerns treatment response or detection rather than progression from non-late to late AMD, so it is adjacent rather than directly within scope.
Verification and currency limitations. Every PMID in the reference list was mechanically verified against NCBI E-utilities esummary, but verification confirms that an indexed record exists and that its title matches; it does not independently re-audit the underlying study’s statistical methods or its reported numbers. Several sources are registry entries, protocols, or non-indexed reports and are cited without PMIDs and labelled accordingly. The literature is moving quickly — notably in generative and self-supervised imaging models and in complement-inhibitor therapy — and a review prepared on the stated date may be overtaken within months. Where a widely repeated statistic could not be traced to a verifiable primary source, it has been flagged or omitted rather than repeated; readers should treat any clinical extrapolation from model performance as an inference, not an established fact.
12. Future Directions
Six priorities follow from the evidence and the anchor review’s prescription.
1. Standardise outcomes and horizons. Progression should be prespecified and reported separately for neovascular AMD, geographic atrophy, and composite late AMD, with fixed-horizon and time-to-event tasks reported as distinct, and anti-VEGF initiation proxies avoided or clearly labelled [1].
2. Make calibration routine. At minimum, studies should report calibration-in-the-large and calibration slope alongside discrimination, and should present decision-curve or other utility-based evaluation when models are positioned for clinical use [66,78,1].
3. Treat external validation and recalibration as the norm, not the exception. Where fully independent data are unavailable, temporal validation, multisite collaboration, and federated analysis are feasible substitutes, and reasons for their absence should be reported transparently; recalibration and updating should be planned rather than deferred indefinitely [69,1].
4. Broaden and diversify the data. Development and validation in diverse, multi-ancestry, routine-care populations is now a validity requirement rather than a nicety, given ancestry-dependent genetic effects and narrow cohort composition [40,1]. Improving dataset access and imaging standardisation is a prerequisite [155,153,152].
5. Clarify incremental value. Future studies should demonstrate that added complexity or additional data types (imaging, genomics, metabolomics) produce reproducible improvement over strong baseline predictors such as drusen characteristics, age, and smoking, rather than merely achieving high AUCs in enriched cohorts [1,90].
6. Move toward decision-relevant, prospectively evaluated tools. The endpoint of the field should be prospectively validated, calibrated, transportable tools embedded in care pathways — potentially linked to treatable targets such as complement inhibition for geographic atrophy and to home-monitoring triage — rather than isolated retrospective AUCs [16,17,144,1]. The PINNACLE and I-SCREEN programmes, and prospective anti-VEGF prognosis models such as KongMing, are early examples of this shift [149,124,125].
13. Conclusions
Multivariable prognostic models for progression of age-related macular degeneration have proliferated, driven by machine learning applied to retinal imaging, but the evidence base’s methodological maturity lags its technical output. Progression is prognostically tractable — discrimination commonly 0.80–0.92, ranging 0.65–0.97 across heterogeneous tasks — and conventional regression and risk-score models remain competitive with deep learning for medium- and long-term prediction [1]. Yet the evidence is drawn from a few recurring, predominantly European-ancestry cohorts; calibration is reported in a minority of studies; external validation is uncommon and often non-independent; recalibration is absent; and risk of bias clusters in the analysis domain [1]. The recurrent predictors — age, drusen burden and morphology, pigmentary abnormalities, reticular pseudodrusen, outer-retinal thinning, hyperreflective foci, and selected genetic variants — form a plausible common biological core, but their overlap with current-severity markers leaves open whether some imaging models add genuine prognostic information beyond severity staging (inference). Genetic risk scores improve stratification but add modest incremental value and are ancestry-dependent. The path to clinical usefulness is not a better algorithm but staged, feasible validation: standardised outcomes and horizons, robust internal validation, routine quantitative calibration, temporal or geographic external validation, and recalibration before use. Until then, current AMD prognostic models should be treated as exploratory research tools rather than clinically actionable instruments, and any decision made on their basis should be made with explicit awareness of their calibration and transportability limits [1].
References (Vancouver — all PMIDs mechanically verified via NCBI E-utilities esummary, 2026-10-12)
- Aljied R, Saggu S, Chaudhary V, Pigeyre M, Griffith LE, Raina P. Multivariable Prognostic Models for Age-Related Macular Degeneration: A Systematic Review. Journal of ophthalmology. 2026;2026:9231744. PMID 42763542.
- Mitchell P, Liew G, Gopinath B, Wong TY. Age-related macular degeneration. Lancet (London, England). 2018;392(10153):1147-1159. PMID 30303083.
- Schultz NM, Bhardwaj S, Barclay C, Gaspar L, Schwartz J. Global Burden of Dry Age-Related Macular Degeneration: A Targeted Literature Review. Clinical therapeutics. 2021;43(10):1792-1818. PMID 34548176.
- Zhang S, Ren J, Chai R, Yuan S, Hao Y. Global burden of low vision and blindness due to age-related macular degeneration from 1990 to 2021 and projections for 2050. BMC public health. 2024;24(1):3510. PMID 39695517.
- Zhang Z, Zhang X, Zhang R, Gao Y. Trends and drivers of blindness and vision loss burden attributable to age-related macular degeneration from 1990 to 2021: findings from the Global Burden of Disease Study 2021. BMJ open. 2025;15(7):e094953. PMID 40701607.
- Peng Y, Keenan TD, Chen Q, Agrón E, Allot A, Wong WT, et al. Predicting risk of late age-related macular degeneration using deep learning. NPJ digital medicine. 2020;3:111. PMID 32904246.
- Ganjdanesh A, Zhang J, Chew EY, Ding Y, Huang H, Chen W. LONGL-Net: temporal correlation structure guided deep learning model to predict longitudinal age-related macular degeneration severity. PNAS nexus. 2022;1(1):pgab003. PMID 35360552.
- Li JQ, Welchowski T, Schmid M, Mauschitz MM, Holz FG, Finger RP. Prevalence and incidence of age-related macular degeneration in Europe: a systematic review and meta-analysis. The British journal of ophthalmology. 2020;104(8):1077-1084. PMID 31712255.
- Rudnicka AR, Kapetanakis VV, Jarrar Z, Wathern AK, Wormald R, Fletcher AE, et al. Incidence of Late-Stage Age-Related Macular Degeneration in American Whites: Systematic Review and Meta-analysis. American journal of ophthalmology. 2015;160(1):85-93.e3. PMID 25857680.
- Zhou M, Duan PC, Liang JH, Zhang XF, Pan CW. Geographic distributions of age-related macular degeneration incidence: a systematic review and meta-analysis. The British journal of ophthalmology. 2021;105(10):1427-1434. PMID 32907810.
- Hendriksen JM, Geersing GJ, Moons KG, de Groot JA. Diagnostic and prognostic prediction models. Journal of thrombosis and haemostasis : JTH. 2013;11 Suppl 1:129-41. PMID 23809117.
- Cheung R, Chun J, Sheidow T, Motolko M, Malvankar-Mehta MS. Diagnostic accuracy of current machine learning classifiers for age-related macular degeneration: a systematic review and meta-analysis. Eye (London, England). 2022;36(5):994-1004. PMID 33958739.
- Leng X, Shi R, Wu Y, Zhu S, Cai X, Lu X, et al. Deep learning for detection of age-related macular degeneration: A systematic review and meta-analysis of diagnostic test accuracy studies. PloS one. 2023;18(4):e0284060. PMID 37023082.
- Pead E, Megaw R, Cameron J, Fleming A, Dhillon B, Trucco E, et al. Automated detection of age-related macular degeneration in color fundus photography: a systematic review. Survey of ophthalmology. 2019;64(4):498-511. PMID 30772363.
- Kang C, Lo JE, Zhang H, Ng SM, Lin JC, Scott IU, et al. Artificial intelligence for diagnosing exudative age-related macular degeneration. The Cochrane database of systematic reviews. 2024;10(10):CD015522. PMID 39417312.
- Heier JS, Lad EM, Holz FG, Rosenfeld PJ, Guymer RH, Boyer D, et al. Pegcetacoplan for the treatment of geographic atrophy secondary to age-related macular degeneration (OAKS and DERBY): two multicentre, randomised, double-masked, sham-controlled, phase 3 trials. Lancet (London, England). 2023;402(10411):1434-1448. PMID 37865470.
- Khanani AM, Patel SS, Staurenghi G, Tadayoni R, Danzig CJ, Eichenbaum DA, et al. Efficacy and safety of avacincaptad pegol in patients with geographic atrophy (GATHER2): 12-month results from a randomised, double-masked, phase 3 trial. Lancet (London, England). 2023;402(10411):1449-1458. PMID 37696275.
- Hagag AM, Kaye R, Hoang V, Riedl S, Anders P, Stuart B, et al. Systematic review of prognostic factors associated with progression to late age-related macular degeneration: Pinnacle study report 2. Survey of ophthalmology. 2024;69(2):165-172. PMID 37890677.
- Davis MD, Gangnon RE, Lee LY, Hubbard LD, Klein BE, Klein R, et al. The Age-Related Eye Disease Study severity scale for age-related macular degeneration: AREDS Report No. 17. Archives of ophthalmology (Chicago, Ill. : 1960). 2005;123(11):1484-98. PMID 16286610.
- Clemons TE, Milton RC, Klein R, Seddon JM, Ferris FL. Risk factors for the incidence of Advanced Age-Related Macular Degeneration in the Age-Related Eye Disease Study (AREDS) AREDS report no. 19. Ophthalmology. 2005;112(4):533-9. PMID 15808240.
- . Lutein + zeaxanthin and omega-3 fatty acids for age-related macular degeneration: the Age-Related Eye Disease Study 2 (AREDS2) randomized clinical trial. JAMA. 2013;309(19):2005-15. PMID 23644932.
- Chew EY, Clemons T, SanGiovanni JP, Danis R, Domalpally A, McBee W, et al. The Age-Related Eye Disease Study 2 (AREDS2): study design and baseline characteristics (AREDS2 report number 1). Ophthalmology. 2012;119(11):2282-9. PMID 22840421.
- Chew EY, Clemons TE, Agrón E, Domalpally A, Keenan TDL, Vitale S, et al. Long-term Outcomes of Adding Lutein/Zeaxanthin and ω-3 Fatty Acids to the AREDS Supplements on Age-Related Macular Degeneration Progression: AREDS2 Report 28. JAMA ophthalmology. 2022;140(7):692-698. PMID 35653117.
- Keenan TDL, Agrón E, Keane PA, Domalpally A, Chew EY. Oral Antioxidant and Lutein/Zeaxanthin Supplements Slow Geographic Atrophy Progression to the Fovea in Age-Related Macular Degeneration. Ophthalmology. 2025;132(1):14-29. PMID 39025435.
- Evans JR, Lawrenson JG. Antioxidant vitamin and mineral supplements for slowing the progression of age-related macular degeneration. The Cochrane database of systematic reviews. 2023;9(9):CD000254. PMID 37702300.
- Schmidt-Erfurth U, Waldstein SM, Klimscha S, Sadeghipour A, Hu X, Gerendas BS, et al. Prediction of Individual Disease Conversion in Early AMD Using Artificial Intelligence. Investigative ophthalmology & visual science. 2018;59(8):3199-3208. PMID 29971444.
- Curcio CA, Owsley C. Rod-Mediated Dark Adaptation as a Suitable Outcome for Early and Intermediate Age-Related Macular Degeneration. Ophthalmology. 2019;126(6):866-867. PMID 31122362.
- Owsley C, McGwin G, Clark ME, Gao L, Gooden L, Thomas TN, et al. Delayed Rod-Mediated Dark Adaptation Is Associated with Incidence and Early Progression of Age-Related Macular Degeneration: Alabama Study on Early Age-Related Macular Degeneration 2. Ophthalmology. 2025;132(11):1273-1283. PMID 40532855.
- Nigalye AK, Hess K, Pundlik SJ, Jeffrey BG, Cukras CA, Husain D. Dark Adaptation and Its Role in Age-Related Macular Degeneration. Journal of clinical medicine. 2022;11(5). PMID 35268448.
- Wu Z, Luu CD, Hodgson LA, Caruso E, Chen FK, Chakravarthy U, et al. Examining the added value of microperimetry and low luminance deficit for predicting progression in age-related macular degeneration. The British journal of ophthalmology. 2021;105(5):711-715. PMID 32606079.
- Holt CA, Niedert C, Gedtal M, Virgili G, Hogg RE, Qureshi R. Prognostic factors for age-related macular degeneration progression: An overview of systematic reviews. Survey of ophthalmology. 2026;71(3):759-773. PMID 41412300.
- Marques-Couto P, Coelho-Costa I, Ferreira-da-Silva R, Andrade JP, Carneiro Â. Mediterranean Diet on Development and Progression of Age-Related Macular Degeneration: Systematic Review and Meta-Analysis of Observational Studies. Nutrients. 2025;17(6). PMID 40292456.
- Saigal K, Salama JE, Pardo AA, Lopez SE, Gregori NZ. Modifiable Lifestyle Risk Factors and Strategies for Slowing the Progression of Age-Related Macular Degeneration. Vision (Basel, Switzerland). 2025;9(1). PMID 40137928.
- Seddon JM, De D, Rosner B. Quantifying Effects of Diet and Lifestyle Changes on Progression to Advanced Age Related Macular Degeneration in High Genetic Risk Individuals. Ophthalmology. 2026;133(3):333-349. PMID 41076031.
- Yu J, Zhang Y, Gao YL, Xu JM, Chen H, Ho M, et al. Metabolic vulnerability index and MetaboHealth score as risk factors for age-related macular degeneration in a large-scale prospective cohort. Journal of translational medicine. 2026;24(1). PMID 41703599.
- Klein RJ, Zeiss C, Chew EY, Tsai JY, Sackler RS, Haynes C, et al. Complement factor H polymorphism in age-related macular degeneration. Science (New York, N.Y.). 2005;308(5720):385-9. PMID 15761122.
- Seddon JM, Reynolds R, Maller J, Fagerness JA, Daly MJ, Rosner B. Prediction model for prevalence and incidence of advanced age-related macular degeneration based on genetic, demographic, and environmental variables. Investigative ophthalmology & visual science. 2009;50(5):2044-53. PMID 19117936.
- Grassmann F, Heid IM, Weber BH. Genetic risk models in age-related macular degeneration. Advances in experimental medicine and biology. 2014;801:291-300. PMID 24664710.
- Cooke Bailey JN, Hoffman JD, Sardell RJ, Scott WK, Pericak-Vance MA, Haines JL. The Application of Genetic Risk Scores in Age-Related Macular Degeneration: A Review. Journal of clinical medicine. 2016;5(3). PMID 26959068.
- Gorman BR, Voloudakis G, Igo RP, Kinzy T, Halladay CW, Bigdeli TB, et al. Genome-wide association analyses identify distinct genetic architectures for age-related macular degeneration across ancestries. Nature genetics. 2024;56(12):2659-2671. PMID 39623103.
- He W, Han X, Ong JS, Wu Y, Hewitt AW, Mackey DA, et al. Genome-Wide Meta-analysis Identifies Risk Loci and Improves Disease Prediction of Age-Related Macular Degeneration. Ophthalmology. 2024;131(1):16-29. PMID 37634759.
- Yu C, Robman L, He W, Woods RL, Phuong Thao LT, Wolfe R, et al. Predictive Performance of an Updated Polygenic Risk Score for Age-Related Macular Degeneration. Ophthalmology. 2024;131(8):880-891. PMID 38307466.
- Huang YC, Liao WL, Lin HJ, Huang YT, Chang YW, Liu TY, et al. Identification of Genetic Variants for Risk Prediction and Early Diagnosis of Age-Related Macular Degeneration in the Taiwanese Population. International journal of molecular sciences. 2024;25(6). PMID 38542204.
- Serra R, Rallo V, Pinna A, Steri M, Piras MG, Marongiu M, et al. Polygenic risk score and biochemical/environmental variables predict a low-risk profile of age-related macular degeneration in Sardinia. Graefe’s archive for clinical and experimental ophthalmology = Albrecht von Graefes Archiv fur klinische und experimentelle Ophthalmologie. 2023;261(3):691-698. PMID 36264335.
- Wąsowska A, Teper S, Matczyńska E, Łyszkiewicz P, Sendecki A, Machalińska A, et al. Polygenic Risk Score Impact on Susceptibility to Age-Related Macular Degeneration in Polish Patients. Journal of clinical medicine. 2022;12(1). PMID 36615095.
- Connolly E, Rhatigan M, O’Halloran AM, Muldrew KA, Chakravarthy U, Cahill M, et al. Prevalence of age-related macular degeneration associated genetic risk factors and 4-year progression data in the Irish population. The British journal of ophthalmology. 2018;102(12):1691-1695. PMID 29453225.
- Farinha C, Barreto P, Coimbra R, Cachulo ML, Melo JB, Cunha-Vaz J, et al. Common and rare genetic risk variants in age-related macular degeneration and genetic risk score in the Coimbra eye study. Acta ophthalmologica. 2023;101(2):185-199. PMID 36036675.
- Grunin M, Triffon D, Beykin G, Rahmani E, Schweiger R, Tiosano L, et al. Genome wide association study and genomic risk prediction of age related macular degeneration in Israel. Scientific reports. 2024;14(1):13034. PMID 38844476.
- Herold JM, Zimmermann ME, Gorski M, Günther F, Weber BHF, Helbig H, et al. Genetic Risk Score Analysis Supports a Joint View of Two Classification Systems for Age-Related Macular Degeneration. Investigative ophthalmology & visual science. 2023;64(12):31. PMID 37721739.
- Wąsowska A, Sendecki A, Boguszewska-Chachulska A, Teper S. Polygenic Risk Score and Rare Variant Burden Identified by Targeted Sequencing in a Group of Patients with Pigment Epithelial Detachment in Age-Related Macular Degeneration. Genes. 2023;14(9). PMID 37761846.
- Gorin MB, daSilva MJ. Predictive genetics for AMD: Hype and hopes for genetics-based strategies for treatment and prevention. Experimental eye research. 2020;191:107894. PMID 31862397.
- Perlee LT, Bansal AT, Gehrs K, Heier JS, Csaky K, Allikmets R, et al. Inclusion of genotype with fundus phenotype improves accuracy of predicting choroidal neovascularization and geographic atrophy. Ophthalmology. 2013;120(9):1880-92. PMID 23523162.
- Awh CC, Lane AM, Hawken S, Zanke B, Kim IK. CFH and ARMS2 genetic polymorphisms predict response to antioxidants and zinc in patients with age-related macular degeneration. Ophthalmology. 2013;120(11):2317-23. PMID 23972322.
- Shijo T, Sakurada Y, Yoneyama S, Kikushima W, Sugiyama A, Matsubara M, et al. Association between Polygenic Risk Score and One-Year Outcomes Following As-Needed Aflibercept Therapy for Exudative Age-Related Macular Degeneration. Pharmaceuticals (Basel, Switzerland). 2020;13(9). PMID 32962278.
- Liao DS, Grossi FV, El Mehdi D, Gerber MR, Brown DM, Heier JS, et al. Complement C3 Inhibitor Pegcetacoplan for Geographic Atrophy Secondary to Age-Related Macular Degeneration: A Randomized Phase 2 Trial. Ophthalmology. 2020;127(2):186-195. PMID 31474439.
- Jaffe GJ, Westby K, Csaky KG, Monés J, Pearlman JA, Patel SS, et al. C5 Inhibitor Avacincaptad Pegol for Geographic Atrophy Due to Age-Related Macular Degeneration: A Randomized Pivotal Phase 2/3 Trial. Ophthalmology. 2021;128(4):576-586. PMID 32882310.
- Holz FG, Sadda SR, Busbee B, Chew EY, Mitchell P, Tufail A, et al. Efficacy and Safety of Lampalizumab for Geographic Atrophy Due to Age-Related Macular Degeneration: Chroma and Spectri Phase 3 Randomized Clinical Trials. JAMA ophthalmology. 2018;136(6):666-677. PMID 29801123.
- Dascalu AM, Grigorescu CC, Serban D, Tudor C, Alexandrescu C, Stana D, et al. Complement Inhibitors for Geographic Atrophy in Age-Related Macular Degeneration-A Systematic Review. Journal of personalized medicine. 2024;14(9). PMID 39338244.
- Garg A, Nanji K, Tai F, Phillips M, Zeraatkar D, Garg SJ, et al. The effect of complement C3 or C5 inhibition on geographic atrophy secondary to age-related macular degeneration: A living systematic review and meta-analysis. Survey of ophthalmology. 2024;69(3):349-361. PMID 38008405.
- Tzoumas N, Riding G, Williams MA, Steel DH. Complement inhibitors for age-related macular degeneration. The Cochrane database of systematic reviews. 2023;6(6):CD009300. PMID 37314061.
- Schmidt-Erfurth U, Mai J, Reiter GS, Riedl S, Vogl WD, Sadeghipour A, et al. Disease Activity and Therapeutic Response to Pegcetacoplan for Geographic Atrophy Identified by Deep Learning-Based Analysis of OCT. Ophthalmology. 2025;132(2):181-193. PMID 39151755.
- Cao JA, Zhou AW, Teagle GM, Baumann LM, Sahraravand RA, Wong CW, et al. Geographic Atrophy Progression in Clinical Practice Before and After Pegcetacoplan Treatment. Vision (Basel, Switzerland). 2025;9(4). PMID 41283579.
- Seddon JM, Silver RE, Kwong M, Rosner B. Risk Prediction for Progression of Macular Degeneration: 10 Common and Rare Genetic Variants, Demographic, Environmental, and Macular Covariates. Investigative ophthalmology & visual science. 2015;56(4):2192-202. PMID 25655794.
- Ying GS, Maguire MG, Glynn RJ, Rosner B. Tutorial on Biostatistics: Longitudinal Analysis of Correlated Continuous Eye Data. Ophthalmic epidemiology. 2021;28(1):3-20. PMID 32744149.
- Alba AC, Agoritsas T, Walsh M, Hanna S, Iorio A, Devereaux PJ, et al. Discrimination and Calibration of Clinical Prediction Models: Users’ Guides to the Medical Literature. JAMA. 2017;318(14):1377-1384. PMID 29049590.
- Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW. Calibration: the Achilles heel of predictive analytics. BMC medicine. 2019;17(1):230. PMID 31842878.
- Jacobs JP, Shahian DM, O’Brien SM, Badhwar V. Assessing the Performance of Risk Models With Discrimination and Calibration. The Annals of thoracic surgery. 2023;115(1):282-286. PMID 35688206.
- Siontis GC, Tzoulaki I, Castaldi PJ, Ioannidis JP. External validation of new risk prediction models is infrequent and reveals worse prognostic discrimination. Journal of clinical epidemiology. 2015;68(1):25-34. PMID 25441703.
- Van Calster B, Steyerberg EW, Wynants L, van Smeden M. There is no such thing as a validated prediction model. BMC medicine. 2023;21(1):70. PMID 36829188.
- Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD): the TRIPOD Statement. The British journal of surgery. 2015;102(3):148-58. PMID 25627261.
- Wolff RF, Moons KGM, Riley RD, Whiting PF, Westwood M, Collins GS, et al. PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Annals of internal medicine. 2019;170(1):51-58. PMID 30596875.
- Moons KGM, Wolff RF, Riley RD, Whiting PF, Westwood M, Collins GS, et al. PROBAST: A Tool to Assess Risk of Bias and Applicability of Prediction Model Studies: Explanation and Elaboration. Annals of internal medicine. 2019;170(1):W1-W33. PMID 30596876.
- Collins GS, Dhiman P, Andaur Navarro CL, Ma J, Hooft L, Reitsma JB, et al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ open. 2021;11(7):e048008. PMID 34244270.
- Liu X, Cruz Rivera S, Moher D, Calvert MJ, Denniston AK. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. The Lancet. Digital health. 2020;2(10):e537-e548. PMID 33328048.
- Liu X, Cruz Rivera S, Moher D, Calvert MJ, Denniston AK. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nature medicine. 2020;26(9):1364-1374. PMID 32908283.
- Rivera SC, Liu X, Chan AW, Denniston AK, Calvert MJ. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ (Clinical research ed.). 2020;370:m3210. PMID 32907797.
- Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nature medicine. 2022;28(5):924-933. PMID 35585198.
- Piovani D, Sokou R, Tsantes AG, Vitello AS, Bonovas S. Optimizing Clinical Decision Making with Decision Curve Analysis: Insights for Clinical Investigators. Healthcare (Basel, Switzerland). 2023;11(16). PMID 37628442.
- Eddy DM, Hollingworth W, Caro JJ, Tsevat J, McDonald KM, Wong JB. Model transparency and validation: a report of the ISPOR-SMDM Modeling Good Research Practices Task Force-7. Medical decision making : an international journal of the Society for Medical Decision Making. 2012;32(5):733-43. PMID 22990088.
- Manke-Reimers F, Brugger V, Bärnighausen T, Kohler S. When, why and how are estimated effects transported between populations? A scoping review of studies applying transportability methods. European journal of epidemiology. 2025;40(3):255-273. PMID 40249515.
- Seddon JM, Reynolds R, Yu Y, Daly MJ, Rosner B. Risk models for progression to advanced age-related macular degeneration using demographic, environmental, genetic, and ocular factors. Ophthalmology. 2011;118(11):2203-11. PMID 21959373.
- Seddon JM, Reynolds R, Yu Y, Rosner B. Validation of a prediction algorithm for progression to advanced macular degeneration subtypes. JAMA ophthalmology. 2013;131(4):448-55. PMID 23411794.
- Seddon JM, Rosner B. Validated Prediction Models for Macular Degeneration Progression and Predictors of Visual Acuity Loss Identify High-Risk Individuals. American journal of ophthalmology. 2019;198:223-261. PMID 30389371.
- Chiu CJ, Mitchell P, Klein R, Klein BE, Chang ML, Gensler G, et al. A risk score for the prediction of advanced age-related macular degeneration: development and validation in 2 prospective cohorts. Ophthalmology. 2014;121(7):1421-7. PMID 24650555.
- Ying GS, Maguire MG. Development of a risk score for geographic atrophy in complications of the age-related macular degeneration prevention trial. Ophthalmology. 2011;118(2):332-8. PMID 20801521.
- Buitendijk GHS, Rochtchina E, Myers C, van Duijn CM, Lee KE, Klein BEK, et al. Prediction of age-related macular degeneration in the general population: the Three Continent AMD Consortium. Ophthalmology. 2013;120(12):2644-2655. PMID 24120328.
- Ajana S, Cougnard-Grégoire A, Colijn JM, Merle BMJ, Verzijden T, de Jong PTVM, et al. Predicting Progression to Advanced Age-Related Macular Degeneration from Clinical, Genetic, and Lifestyle Factors Using Machine Learning. Ophthalmology. 2021;128(4):587-597. PMID 32890546.
- Shin KU, Song SJ, Bae JH, Lee MY. Risk Prediction Model for Progression of Age-Related Macular Degeneration. Ophthalmic research. 2017;57(1):32-36. PMID 27732974.
- Delcourt C, Souied E, Sanchez A, Bandello F. Development and Validation of a Risk Score for Age-Related Macular Degeneration: The STARS Questionnaire. Investigative ophthalmology & visual science. 2017;58(14):6399-6407. PMID 29260196.
- McCarthy LC, Newcombe PJ, Whittaker JC, Wurzelmann JI, Fries MA, Burnham NR, et al. Predictive models of choroidal neovascularization and geographic atrophy incidence applied to clinical trial design. American journal of ophthalmology. 2012;154(3):568-578.e12. PMID 22704140.
- Trinh M, Duong A, Cheung R, Chen S, Ng D, Friedrich J, et al. Proposal of a Simpler Eye-Level Risk Model Incorporating Reticular Pseudodrusen for the Clinical Prediction of Late Age-Related Macular Degeneration. Clinical & experimental ophthalmology. 2025;53(8):936-945. PMID 40610024.
- Shen LL, Liu F, Nardini HG, Del Priore LV. RECLASSIFICATION OF FUNDUS AUTOFLUORESCENCE PATTERNS SURROUNDING GEOGRAPHIC ATROPHY BASED ON PROGRESSION RATE: A Systematic Review and Meta-Analysis. Retina (Philadelphia, Pa.). 2019;39(10):1829-1839. PMID 30829988.
- Shen LL, Liu F, Grossetta Nardini HK, Del Priore LV. Fellow Eye Status Is a Biomarker for the Progression Rate of Geographic Atrophy: A Systematic Review and Meta-analysis. Ophthalmology. Retina. 2019;3(4):305-315. PMID 31014681.
- Niu S, de Sisternes L, Chen Q, Rubin DL, Leng T. Fully Automated Prediction of Geographic Atrophy Growth Using Quantitative Spectral-Domain Optical Coherence Tomography Biomarkers. Ophthalmology. 2016;123(8):1737-1750. PMID 27262765.
- Cluceru J, Anegondi N, Gao SS, Lee AY, Lad EM, Chakravarthy U, et al. Topographic Clinical Insights From Deep Learning-Based Geographic Atrophy Progression Prediction. Translational vision science & technology. 2024;13(8):6. PMID 39102242.
- Arslan J, Benke KK. Progression of Geographic Atrophy: Epistemic Uncertainties Affecting Mathematical Models and Machine Learning. Translational vision science & technology. 2021;10(13):3. PMID 34727162.
- Sitnilska V, Kersten E, Altay L, Schick T, Enders P, de Jong EK, et al. Major Predictive Factors for Progression of Early to Late Age-Related Macular Degeneration. Ophthalmologica. Journal international d’ophtalmologie. International journal of ophthalmology. Zeitschrift fur Augenheilkunde. 2020;243(6):444-452. PMID 32172233.
- Burlina PM, Joshi N, Pacheco KD, Freund DE, Kong J, Bressler NM. Use of Deep Learning for Detailed Severity Characterization and Estimation of 5-Year Risk Among Patients With Age-Related Macular Degeneration. JAMA ophthalmology. 2018;136(12):1359-1366. PMID 30242349.
- Grassmann F, Mengelkamp J, Brandl C, Harsch S, Zimmermann ME, Linkohr B, et al. A Deep Learning Algorithm for Prediction of Age-Related Eye Disease Study Severity Scale for Age-Related Macular Degeneration from Color Fundus Photography. Ophthalmology. 2018;125(9):1410-1420. PMID 29653860.
- Peng Y, Dharssi S, Chen Q, Keenan TD, Agrón E, Wong WT, et al. DeepSeeNet: A Deep Learning Model for Automated Classification of Patient-based Age-related Macular Degeneration Severity from Color Fundus Photographs. Ophthalmology. 2019;126(4):565-575. PMID 30471319.
- Russakoff DB, Lamin A, Oakley JD, Dubis AM, Sivaprasad S. Deep Learning for Prediction of AMD Progression: A Pilot Study. Investigative ophthalmology & visual science. 2019;60(2):712-722. PMID 30786275.
- de Sisternes L, Simon N, Tibshirani R, Leng T, Rubin DL. Quantitative SD-OCT imaging biomarkers as indicators of age-related macular degeneration progression. Investigative ophthalmology & visual science. 2014;55(11):7093-103. PMID 25301882.
- Yim J, Chopra R, Spitz T, Winkens J, Obika A, Kelly C, et al. Predicting conversion to wet age-related macular degeneration using deep learning. Nature medicine. 2020;26(6):892-899. PMID 32424211.
- Yan Q, Weeks DE, Xin H, Swaroop A, Chew EY, Huang H, et al. Deep-learning-based Prediction of Late Age-Related Macular Degeneration Progression. Nature machine intelligence. 2020;2(2):141-150. PMID 32285025.
- Hallak JA, de Sisternes L, Osborne A, Yaspan B, Rubin DL, Leng T. Imaging, Genetic, and Demographic Factors Associated With Conversion to Neovascular Age-Related Macular Degeneration: Secondary Analysis of a Randomized Clinical Trial. JAMA ophthalmology. 2019;137(7):738-744. PMID 31021381.
- Nassisi M, Lei J, Abdelfattah NS, Karamat A, Balasubramanian S, Fan W, et al. OCT Risk Factors for Development of Late Age-Related Macular Degeneration in the Fellow Eyes of Patients Enrolled in the HARBOR Study. Ophthalmology. 2019;126(12):1667-1674. PMID 31281056.
- Banerjee I, de Sisternes L, Hallak JA, Leng T, Osborne A, Rosenfeld PJ, et al. Prediction of age-related macular degeneration disease using a sequential deep learning approach on longitudinal SD-OCT imaging biomarkers. Scientific reports. 2020;10(1):15434. PMID 32963300.
- Lee J, Wanyan T, Chen Q, Keenan TDL, Glicksberg BS, Chew EY, et al. Predicting Age-related Macular Degeneration Progression with Longitudinal Fundus Images Using Deep Learning. Machine learning in medical imaging. MLMI (Workshop). 2022;13583:11-20. PMID 36656604.
- Yin C, Moroi SE, Zhang P. Predicting Age-Related Macular Degeneration Progression with Contrastive Attention and Time-Aware LSTM. KDD : proceedings. International Conference on Knowledge Discovery & Data Mining. 2022;2022:4402-4412. PMID 36158613.
- Chakravarty A, Emre T, Leingang O, Riedl S, Mai J, Scholl HPN, et al. Morph-SSL: Self-Supervision With Longitudinal Morphing for Forecasting AMD Progression From OCT Volumes. IEEE transactions on medical imaging. 2024;43(9):3224-3239. PMID 38635383.
- Bridge J, Harding S, Zheng Y. Development and validation of a novel prognostic model for predicting AMD progression using longitudinal fundus images. BMJ open ophthalmology. 2020;5(1):e000569. PMID 33083553.
- Dow ER, Jeong HK, Katz EA, Toth CA, Wang D, Lee T, et al. A Deep-Learning Algorithm to Predict Short-Term Progression to Geographic Atrophy on Spectral-Domain Optical Coherence Tomography. JAMA ophthalmology. 2023;141(11):1052-1061. PMID 37856139.
- Schmidt-Erfurth U, Bogunovic H, Grechenig C, Bui P, Fabianska M, Waldstein S, et al. Role of Deep Learning-Quantified Hyperreflective Foci for the Prediction of Geographic Atrophy Progression. American journal of ophthalmology. 2020;216:257-270. PMID 32277942.
- Liefers B, Colijn JM, González-Gonzalo C, Verzijden T, Wang JJ, Joachim N, et al. A Deep Learning Model for Segmentation of Geographic Atrophy to Study Its Long-Term Natural History. Ophthalmology. 2020;127(8):1086-1096. PMID 32197912.
- Rivail A, Vogl WD, Riedl S, Grechenig C, Coulibaly LM, Reiter GS, et al. Deep survival modeling of longitudinal retinal OCT volumes for predicting the onset of atrophy in patients with intermediate AMD. Biomedical optics express. 2023;14(6):2449-2464. PMID 37342683.
- Pfau M, von der Emde L, de Sisternes L, Hallak JA, Leng T, Schmitz-Valckenberg S, et al. Progression of Photoreceptor Degeneration in Geographic Atrophy Secondary to Age-related Macular Degeneration. JAMA ophthalmology. 2020;138(10):1026-1034. PMID 32789526.
- Liu TYA, Liu Y, Gastonguay MS, Midgett D, Kuo N, Zhao Y, et al. Predicting Imminent Conversion to Exudative Age-Related Macular Degeneration Using Multimodal Data and Ensemble Machine Learning. Ophthalmology science. 2025;5(5):100785. PMID 40502295.
- Wu AL, Guo Y, Hormel TT, Flaxel CJ, Hwang TS, Huang D, et al. Drusen volume and reticular pseudodrusen volume from optical coherence tomography with deep learning as risk factors for progression to late age-related macular degeneration in eyes with reticular pseudodrusen and contralateral macular neovascularisation. The British journal of ophthalmology. 2026. PMID 42336620.
- Halev A, Huang D, Rezaei S, Banks S, McPherson JD, Shankar SP, et al. Genotype Prediction from Retinal Fundus Images Using Deep Learning in Eyes with Age-Related Macular Degeneration. Ophthalmology science. 2025;5(6):100836. PMID 40661176.
- Sendecki A, Ledwoń D, Tuszy A, Nycz J, Wąsowska A, Boguszewska-Chachulska A, et al. Fundus Image Deep Learning Study to Explore the Association of Retinal Morphology with Age-Related Macular Degeneration Polygenic Risk Score. Biomedicines. 2024;12(9). PMID 39335605.
- Govindaiah A, Baten A, Smith RT, Balasubramanian S, Bhuiyan A. Optimized Prediction Models from Fundus Imaging and Genetics for Late Age-Related Macular Degeneration. Journal of personalized medicine. 2021;11(11). PMID 34834479.
- Yan Q, Jiang Y, Huang H, Swaroop A, Chew EY, Weeks DE, et al. Genome-Wide Association Studies-Based Machine Learning for Prediction of Age-Related Macular Degeneration Risk. Translational vision science & technology. 2021;10(2):29. PMID 34003914.
- Sun T, Wei Y, Chen W, Ding Y. Genome-wide association study-based deep learning for survival prediction. Statistics in medicine. 2020;39(30):4605-4620. PMID 32974946.
- Enzendorfer ML, Reiter GS, Bogunović H, Riedl S, Mai J, Schrittwieser J, et al. I-SCREEN: Development of an AI-based infrastructure for community-wide screening and prediction of progression in age-related macular degeneration providing accessible shared care. Eye (London, England). 2026;40(12):1806-1814. PMID 42129346.
- Zhao X, Wen Y, Yang Z, Gu X, Zeng Y, Lai A, et al. Development and validation of a deep learning model to predict visual and anatomical prognosis of anti-VEGF therapy for neovascular age-related macular degeneration (KongMing Study): a prospective, nationwide, multicentre study. The Lancet. Digital health. 2026;8(4):100971. PMID 41896127.
- Sumer F, Toren M, Asan B, Solak M, Coskuner N, Ozkan B, et al. Comparative Evaluation of Deep Generative Models for Predicting 12-Month Neovascular AMD Progression Using OCT and Fundus Photography. Journal of imaging informatics in medicine. 2026. PMID 42552269.
- Muntean GA, Marginean A, Groza A, Damian I, Roman SA, Hapca MC, et al. The Predictive Capabilities of Artificial Intelligence-Based OCT Analysis for Age-Related Macular Degeneration Progression-A Systematic Review. Diagnostics (Basel, Switzerland). 2023;13(14). PMID 37510207.
- Damian I, Nicoară SD. SD-OCT Biomarkers and the Current Status of Artificial Intelligence in Predicting Progression from Intermediate to Advanced AMD. Life (Basel, Switzerland). 2022;12(3). PMID 35330205.
- Bhuiyan A, Wong TY, Ting DSW, Govindaiah A, Souied EH, Smith RT. Artificial Intelligence to Stratify Severity of Age-Related Macular Degeneration (AMD) and Predict Risk of Progression to Late AMD. Translational vision science & technology. 2020;9(2):25. PMID 32818086.
- Wu Z, Bogunović H, Asgari R, Schmidt-Erfurth U, Guymer RH. Predicting Progression of Age-Related Macular Degeneration Using OCT and Fundus Photography. Ophthalmology. Retina. 2021;5(2):118-125. PMID 32599175.
- Finger RP, Wickremasinghe SS, Baird PN, Guymer RH. Predictors of anti-VEGF treatment response in neovascular age-related macular degeneration. Survey of ophthalmology. 2014;59(1):1-18. PMID 24332379.
- Shah AR, Williams S, Baumal CR, Rosner B, Duker JS, Seddon JM. Predictors of Response to Intravitreal Anti-Vascular Endothelial Growth Factor Treatment of Age-Related Macular Degeneration. American journal of ophthalmology. 2016;163:154-166.e8. PMID 26705092.
- Ashraf M, Souka A, Adelman RA. Age-related macular degeneration: using morphological predictors to modify current treatment protocols. Acta ophthalmologica. 2018;96(2):120-133. PMID 29130626.
- Oca AI, Pérez-Sala Á, Pariente A, Ochoa R, Velilla S, Peláez R, et al. Predictive Biomarkers of Age-Related Macular Degeneration Response to Anti-VEGF Treatment. Journal of personalized medicine. 2021;11(12). PMID 34945801.
- Hsia Y, Yang CH, Hsieh YT, Yang CM, Ho TC, Lai TT. Hyperreflective foci in predicting the treatment outcome of antivascular endothelial growth factor in neovascular age-related macular degeneration. Graefe’s archive for clinical and experimental ophthalmology = Albrecht von Graefes Archiv fur klinische und experimentelle Ophthalmologie. 2020;258(2):273-280. PMID 31828425.
- Roberts PK, Schranz M, Motschi A, Desissaire S, Hacker V, Pircher M, et al. Baseline predictors for subretinal fibrosis in neovascular age-related macular degeneration. Scientific reports. 2022;12(1):88. PMID 34996934.
- Bogunovic H, Waldstein SM, Schlegl T, Langs G, Sadeghipour A, Liu X, et al. Prediction of Anti-VEGF Treatment Requirements in Neovascular AMD Using a Machine Learning Approach. Investigative ophthalmology & visual science. 2017;58(7):3240-3248. PMID 28660277.
- Fu DJ, Faes L, Wagner SK, Moraes G, Chopra R, Patel PJ, et al. Predicting Incremental and Future Visual Change in Neovascular Age-Related Macular Degeneration Using Deep Learning. Ophthalmology. Retina. 2021;5(11):1074-1084. PMID 33516917.
- Tan TE, Wong TY, Ting DSW. Artificial Intelligence for Prediction of Anti-VEGF Treatment Burden in Retinal Diseases: Towards Precision Medicine. Ophthalmology. Retina. 2021;5(7):601-603. PMID 34243967.
- Arslan J, Benke KK. Application of Machine Learning to Ranking Predictors of Anti-VEGF Response. Life (Basel, Switzerland). 2022;12(11). PMID 36431061.
- Coscas F, Lupidi M, Boulet JF, Sellam A, Cabral D, Serra R, et al. Optical coherence tomography angiography in exudative age-related macular degeneration: a predictive model for treatment decisions. The British journal of ophthalmology. 2019;103(9):1342-1346. PMID 30467129.
- Liu TYA, Farsiu S, Ting DS. Generative adversarial networks to predict treatment response for neovascular age-related macular degeneration: interesting, but is it useful?. The British journal of ophthalmology. 2020;104(12):1629-1630. PMID 32862129.
- Ferrara D, Newton EM, Lee AY. Artificial intelligence-based predictions in neovascular age-related macular degeneration. Current opinion in ophthalmology. 2021;32(5):389-396. PMID 34265783.
- Chew EY, Clemons TE, Bressler SB, Elman MJ, Danis RP, Domalpally A, et al. Randomized trial of the ForeseeHome monitoring device for early detection of neovascular age-related macular degeneration. The HOme Monitoring of the Eye (HOME) study design – HOME Study report number 1. Contemporary clinical trials. 2014;37(2):294-300. PMID 24530651.
- Yu HJ, Kiernan DF, Eichenbaum D, Sheth VS, Wykoff CC. Home Monitoring of Age-Related Macular Degeneration: Utility of the ForeseeHome Device for Detection of Neovascularization. Ophthalmology. Retina. 2021;5(4):348-356. PMID 32810682.
- Busquets MA, Sabbagh O. Current status of home monitoring technology for age-related macular degeneration. Current opinion in ophthalmology. 2021;32(3):240-246. PMID 33710010.
- Han DP. The ForeSeeHome device and the HOME study: a milestone in the self-detection of neovascular age-related macular degeneration. JAMA ophthalmology. 2014;132(10):1167-8. PMID 25058876.
- Hogg RE, Sivaprasad S, Wickens R, O’Connor S, Gidman E, Ward E, et al. Home-Monitoring Vision Tests to Detect Active Neovascular Age-Related Macular Degeneration. JAMA ophthalmology. 2024;142(6):512-520. PMID 38662399.
- Sutton J, Menten MJ, Riedl S, Bogunović H, Leingang O, Anders P, et al. Developing and validating a multivariable prediction model which predicts progression of intermediate to late age-related macular degeneration-the PINNACLE trial protocol. Eye (London, England). 2023;37(6):1275-1283. PMID 35614343.
- Heesterbeek TJ, de Jong EK, Acar IE, Groenewoud JMM, Liefers B, Sánchez CI, et al. Genetic risk score has added value over initial clinical grading stage in predicting disease progression in age-related macular degeneration. Scientific reports. 2019;9(1):6611. PMID 31036867.
- Hollitt GL, Qassim A, Thomson D, Schmidt JM, Nguyen TT, Landers J, et al. Genetic Risk Assessment of Degenerative Eye Disease (GRADE): study protocol of a prospective assessment of polygenic risk scores to predict diagnosis of glaucoma and age-related macular degeneration. BMC ophthalmology. 2023;23(1):431. PMID 37875865.
- Lee AY, Sadda S, Lum F. Joint American Academy of Ophthalmology and Association for Research in Vision and Ophthalmology Policy Statement: The Time for Digital Imaging Standards Implementation Is Now. Ophthalmology. 2022;129(11):1229-1231. PMID 36096893.
- Gim N, Ferguson AN, Blazes M, Lee CS, Lee AY. The march to harmonized imaging standards for retinal imaging. Progress in retinal and eye research. 2025;107:101363. PMID 40360070.
- Dow ER, Keenan TDL, Lad EM, Lee AY, Lee CS, Loewenstein A, et al. From Data to Deployment: The Collaborative Community on Ophthalmic Imaging Roadmap for Artificial Intelligence in Age-Related Macular Degeneration. Ophthalmology. 2022;129(5):e43-e59. PMID 35016892.
- Khan SM, Liu X, Nath S, Korot E, Faes L, Wagner SK, et al. A global review of publicly available datasets for ophthalmological imaging: barriers to access, usability, and generalisability. The Lancet. Digital health. 2021;3(1):e51-e66. PMID 33735069.
- Zhang Y, Shi J, Peng Y, Zhao Z, Zheng Q, Wang Z, et al. Artificial intelligence-enabled screening for diabetic retinopathy: a real-world, multicenter and prospective study. BMJ open diabetes research & care. 2020;8(1). PMID 33087340.
- Ruamviboonsuk P, Chantra S, Seresirikachorn K, Ruamviboonsuk V, Sangroongruangsri S. Economic Evaluations of Artificial Intelligence in Ophthalmology. Asia-Pacific journal of ophthalmology (Philadelphia, Pa.). 2021;10(3):307-316. PMID 34261102.
- Gunasekeran DV, Ting DSW, Tan GSW, Wong TY. Artificial intelligence for diabetic retinopathy screening, prediction and management. Current opinion in ophthalmology. 2020;31(5):357-365. PMID 32740069.
- Lee E, Hunt D, Cakir Y, Kuo D, Zhou Z, Pajic M, et al. Artificial intelligence in age-related macular degeneration: Advancing diagnosis, prognosis, and treatment. Survey of ophthalmology. 2026;71(2):560-571. PMID 40975441.
- Abd El-Khalek AA, Balaha HM, Sewelam A, Ghazal M, Khalil AT, Abo-Elsoud MEA, et al. A Comprehensive Review of AI Diagnosis Strategies for Age-Related Macular Degeneration (AMD). Bioengineering (Basel, Switzerland). 2024;11(7). PMID 39061793.
- Tan YY, Kang HG, Lee CJ, Kim SS, Park S, Thakur S, et al. Prognostic potentials of AI in ophthalmology: systemic disease forecasting via retinal imaging. Eye and vision (London, England). 2024;11(1):17. PMID 38711111.
- Agrón E, Domalpally A, Cukras CA, Clemons TE, Chen Q, Lu Z, et al. Reticular Pseudodrusen: The Third Macular Risk Feature for Progression to Late Age-Related Macular Degeneration: Age-Related Eye Disease Study 2 Report 30. Ophthalmology. 2022;129(10):1107-1119. PMID 35660417.
Review complete as of 2026-10-12. All cited PMIDs were mechanically verified via NCBI E-utilities esummary; the verification record and provenance are preserved alongside this file. Where sources could not corroborate a commonly repeated figure, the review states the verified figure or records the claim as unverified rather than repeating it.