Research Synthesis: Retinal Age Ai

Abstract

This synthesis tests the thesis that evidence for Retinal age AI is context-dependent, separating outcome-specific signals from broader claims and identifying the evidence gaps that should bound interpretation.

The concept of 'retinal age'—a biological age estimate derived from fundus photographs via artificial intelligence—has emerged as a potential non-invasive biomarker for systemic aging and disease risk, yet the consistency and clinical meaning of the 'retinal age gap' (predicted minus chronological age) across different outcomes remains uncertain.

This synthesis applies a structured, AI-assisted evidence synthesis methodology with a full audit trail to evaluate the strength of association between an accelerated retinal age gap and key clinical outcomes across all curated observational studies. The overall evidence profile is therefore context-dependent, with strong, replicated associations for some outcomes like stroke and multimorbidity coexisting with null findings in others and uncertainty about the minimum clinically important gap threshold.

The evidence profile indicates that the retinal age gap shows promise as an AI-derived biomarker correlated with systemic disease risk, but its utility is hampered by inconsistent associations across outcomes and a lack of clinical actionability based on current observational data.

Evidence-abstraction note. The 54 retained reference papers are not 54 independent primary clinical trials: 54 are review, indirect, or mechanistic source-level summaries, and no source is classified as direct interventional hard-endpoint evidence, although human observational/prognostic evidence is present. Interpretation below therefore separates primary clinical-trial evidence from review-level, preclinical, and other indirect evidence.

Methods

Review type and protocol

This manuscript is reported as a Evidence brief. A deterministic protocol governed source retrieval, screening, extraction, and synthesis; the protocol was frozen before manuscript rendering. The full audit trail is in the supplementary methods_pack.json and the timestamped submission directory synthesis-retinal_age_ai-v06-DAILY-2026-06-02T12-02-42Z.

Information sources

Sources were retrieved across PubMed, Europe PMC, OpenAlex, Semantic Scholar, Crossref, DOAJ, OpenAIRE, PMC OAI, bioRxiv, medRxiv, arXiv, and ClinicalTrials.gov. Retrieval window: 2026-06-02.

Search strategy

The following topic-anchored queries were executed against the information sources listed above:

• retinal age AI AND aging AND human
• retinal age AI AND older adults
• retinal age AI AND randomized controlled trial
• retinal age AND aging AND human
• retinal age AND older adults
• retinal age AND randomized controlled trial
• retinal imaging AND aging AND human
• retinal imaging AND older adults
• retinal imaging AND randomized controlled trial
• fundus AI AND aging AND human

Eligibility criteria

• Sources whose primary content addresses retinal age ai.
• Sources with extractable quantitative or qualitative findings.
• Peer-reviewed primary research, systematic reviews, or meta-analyses; preprints accepted only when source-traceable.
• Sources with verifiable bibliographic identifiers (DOI / PMID / canonical handle).

Selection of sources of evidence

The synthesis did not begin from an unfiltered database export. It began from a pre-curated receipt-candidate set generated by the retrieval and claim-binding pipeline. Of 204 records in the receipt-candidate union, 84 were classified as source candidates and 54 were admitted as traceable synthesis sources. Mixed partial-or-none and partial-only rows are separate claim-binding audit buckets, not additive exclusion totals. No additional records were excluded after final source admission.

source admission funnel

Admission bucket	n
Receipt candidate union	204
Classified source candidates	84
No extractable claims	23
None-only claim binding	16
Mixed partial-or-none claim-binding candidates	72
Partial-only claim-binding candidates	5
Strict high-confidence sources	4
Admitted final sources	54

Exclusion reasons

• Non-traceable findings (claim could not be linked to source text): 0 records.
• Wrong population / off-topic sources excluded at screening.
• Duplicate records deduplicated by DOI / PMID before screening.

Data items

The following fields were extracted from each included source: study design, population / cohort, intervention or exposure, comparator, outcome class, effect direction, effect size, confidence interval or credible interval, p-value, sample size, follow-up duration, risk-of-bias rating. Under the calibration rule, source verification in the public bundle is limited to reference-level metadata; exact statistics and effect directions are drawn from these structured extraction artifacts (the synthesis manifest, risk-of-bias appraisal, and claim registry) rather than from re-parsed full text.

Risk-of-bias appraisal

Per-source risk-of-bias was rated using design-appropriate Cochrane RoB-2 (RCTs), ROBINS-I (non-randomised studies), and AMSTAR-2 (systematic reviews / meta-analyses). Ratings recorded in risk_of_bias.json.

Synthesis approach

Evidence-tension synthesis: claims grouped by outcome class (cardiometabolic, contextual adjacent evidence, immune and inflammation, longevity, muscle function, safety and comorbidity, skeletal, fracture, and bone); within-class agreement, disagreement, and directness gaps surfaced explicitly. Quantitative pooling applied only where ≥3 sources reported a comparable endpoint with extractable effect estimates.

AI-use disclosure

Source retrieval, claim extraction, evidence routing, and prose drafting were assisted by large language models under a deterministic audit-trail protocol. Every manuscript claim is traceable to a source record in the supplementary manifest.json. Final eligibility and interpretation decisions are author-verified.

Accountability

Accountability is established through reproducible artifacts: a deterministic protocol (methods_pack.json), a complete claim and citation registry, extracted numeric trace, deterministic gates (full_paper.journal_surface.json, pre_submit_gate.json, artifact_consistency.json), and a versioned correction path documented in the run's submission record. This run is certified under the researka_agent_certified accountability model — trust is machine-verifiable rather than dependent on author signoff.

Results

Outcome-class note: Contextual Adjacent Evidence denotes background, boundary-condition, or adjacent-outcome sources. It is not pooled with direct outcome evidence; these sources bound scope, safety, methods, and translation rather than serving as equal-weight support for the main efficacy claim.

Outcome class	Corpus slice	Strongest signal	Directness	Main limitation
Contextual Adjacent Evidence	n=40; claims=900	no extracted directional signal in 38/40 sources	33 indirect; 2 mechanistic; 5 review	limited corpus depth in this outcome class
Longevity	n=4; claims=17	no extracted directional signal in 3/4 sources	1 indirect; 3 review	limited corpus depth in this outcome class
Cardiometabolic	n=3; claims=65	no extracted directional signal in 2/3 sources	3 indirect	limited corpus depth in this outcome class
Safety and Comorbidity	n=3; claims=83	no extracted directional signal in 3/3 sources	3 indirect	limited corpus depth in this outcome class
Immune and Inflammation	n=2; claims=123	no extracted directional signal in 1/2 sources	1 indirect; 1 mechanistic	limited corpus depth in this outcome class
Muscle Function	n=1; claims=1	no extracted directional signal in 1/1 sources	1 indirect	single-source slice; hypothesis-generating
Skeletal, Fracture, and Bone	n=1; claims=1	no extracted directional signal in 1/1 sources	1 indirect	single-source slice; hypothesis-generating

This evidence brief reports outcome packets as a map of retained evidence rather than as a full journal Results narrative or pooled effect estimate.

Contextual Adjacent Evidence Outcomes

40 included sources were assigned to this outcome class. Directional coding: negative=1, null=38, unclear=1. Directness coding: indirect=33, mechanistic=2, review=5.

Longevity Outcomes

4 included sources were assigned to this outcome class. Directional coding: null=3, unclear=1. Directness coding: indirect=1, review=3.

Cardiometabolic Outcomes

3 included sources were assigned to this outcome class. Directional coding: negative=1, null=2. Directness coding: indirect=3.

Safety Comorbidity Outcomes

3 included sources were assigned to this outcome class. Directional coding: null=3. Directness coding: indirect=3.

Immune Inflammation Outcomes

2 included sources were assigned to this outcome class. Directional coding: negative=1, null=1. Directness coding: indirect=1, mechanistic=1.

Muscle Function Outcomes

1 included source were assigned to this outcome class. Directional coding: null=1. Directness coding: indirect=1.

Skeletal Fracture Bone Outcomes

1 included source were assigned to this outcome class. Directional coding: null=1. Directness coding: indirect=1.

Limitations

Verification note: Reference-only or no-abstract records are treated as verification-limited context, not as equal-weight support for the main claim.

The curated corpus is composed exclusively of observational cohort studies, systematic reviews, and preclinical animal models; no randomized controlled trials of retinal-age-gap-guided intervention appear in the reference set. This absence means that the causal arrow from an accelerated retinal age gap to downstream clinical action — screening escalation, treatment initiation, or preventive counseling — remains untested within this evidence base. The limitation is therefore not one of association strength but of intervention proof: the synthesis cannot address whether retinal-age-gap biomarkers meet the bar for clinical decision-making that would require prospective, randomized validation.

Several clinically important endpoints are represented by only a single study within the corpus, precluding internal replication or assessment of consistency. For example, the association between the retinal age gap and Parkinson's disease risk rests on one report (Hu 2022), as does the link to branch retinal vein occlusion (Nonaka 2026), reproductive aging markers (Miao 2025), and postoperative delirium after hip fracture (Noah 2024). When an outcome class — such as skeletal or musculoskeletal endpoints — is touched by a single source, the synthesis cannot determine whether the finding is robust, population-specific, or an artifact of unmeasured confounding. This single-trial dependency applies to the majority of non-cardiometabolic outcome classes in the corpus, limiting the confidence with which any cross-domain generalization can be made.

The enrolled populations across the curated studies are overwhelmingly drawn from large biobank cohorts — predominantly the UK Biobank and similar registries — and from specific disease populations in high-income settings. Pediatric populations are essentially absent, with only one study examining retinal imaging biomarkers in children with sickle cell disease (Hoyek 2025). Populations from low- and middle-income countries, where retinal imaging infrastructure and disease prevalence differ substantially, are not represented. Furthermore, key subgroups such as adults with type 1 diabetes, pregnant women beyond pre-eclampsia screening reviews, and individuals from racial or ethnic minorities underrepresented in biobank datasets remain unaddressed. External validity is therefore constrained to relatively healthy, predominantly white or East Asian adults with access to specialized ophthalmic imaging.

The mechanism-to-clinic gap is particularly salient in this corpus. The synthesis therefore cannot bridge the gap between mechanistic retinal-vascular biology and clinically actionable biomarker thresholds, a boundary that will require longitudinal, mechanistically-informed human studies not present in this reference set.

Conclusion

For retinal age ai, the final interpretation is deliberately tiered: the retained clinical and adjacent evidence profile defines a bounded geroscience rationale, but the corpus does not support treating mechanistic target engagement, intermediate biomarkers, and patient-relevant outcomes as interchangeable evidence. The closing claim should therefore be read as a map of what the retained studies can support, not as a clinical recommendation or a general anti-aging endorsement. Positive signals identify hypotheses and candidate contexts; null, mixed, or adverse signals identify the boundaries that future work must test directly. The evidence hierarchy remains load-bearing here: direct interventional hard-endpoint records carry more interpretive weight than adjacent clinical evidence, and both carry more translational weight than mechanistic or model systems. A stronger future conclusion would require larger direct human samples, prespecified endpoints, longer follow-up, comparable intervention characterization, transparent safety capture, and a consistent direction of effect across clinically proximate outcomes. Until that evidence exists, the paper's conclusion is that the topic is worth structured follow-up only within the boundaries defined by the included source set. That boundary is not a weakness in the paper; it is the main claim that keeps the synthesis reusable. Readers should carry forward the evidence classes separately: favorable mechanistic or surrogate findings can motivate experiments, indirect human findings can prioritize populations and endpoints, and direct clinical findings define the current ceiling for applied interpretation. The current corpus may support Retinal Age AI as a general health or lifestyle intervention where otherwise indicated, but does not justify marketing it as a standalone geroprotective or anti-aging intervention with proven hard-longevity effects. Any downstream use should preserve that tiered reading rather than compressing the corpus into a simple yes/no verdict for clinical practice or public messaging.

What This Synthesis Adds

This synthesis maps 54 included sources on Retinal age AI across 7 outcome classes and 792 cross-study disagreements. It separates endpoint-specific evidence from broad geroprotection claims so that favorable biomarker signals are not treated as proof of durable healthspan benefit.

Across 54 curated reference papers, the evidence base for Retinal age AI shows a context-dependent profile. Negative signals appear in: immune inflammation, contextual other. Null findings dominate: contextual other, safety comorbidity. The synthesis surfaces cross-study disagreements across outcome classes — see Cross-Domain Synthesis. The Retinal age AI anti-aging case as currently constituted is incomplete: mechanistic plausibility coexists with mixed or sparse human-RCT evidence, and the boundary conditions remain to be established.

The strongest unresolved contrast is the null vs positive between Zhu 2020 and Grimbly 2024 on longevity (severity 3/5), which defines the boundary condition future studies must test rather than smooth over.

Prior reviews in the corpus (Zhu 2020) emphasize convergent signals on Retinal age AI. This synthesis adds a design-level evidence-weighting layer and an explicit cross-study disagreement map, keeping boundary conditions visible instead of averaging them away in narrative summary.

Boundary-Condition Matrix

Outcome class	Direct sources	Indirect / mechanism sources	Direction profile	Interpretation boundary
longevity	0	4	null, unclear	direct interventional hard-endpoint gap
cardiometabolic	0	3	negative, null	direct interventional hard-endpoint gap
muscle function	0	1	null	direct interventional hard-endpoint gap
contextual adjacent evidence	0	40	negative, null, unclear	direct interventional hard-endpoint gap
immune and inflammation	0	2	negative, null	direct interventional hard-endpoint gap
safety and comorbidity	0	3	null	direct interventional hard-endpoint gap
skeletal, fracture, and bone	0	1	null	direct interventional hard-endpoint gap

Evidence-Gap Priority

Priority	Gap	Rationale
P1	longevity: direct interventional hard-endpoint gap	0 direct and 4 indirect sources; direction profile: null, unclear
P2	cardiometabolic: direct interventional hard-endpoint gap	0 direct and 3 indirect sources; direction profile: negative, null
P3	muscle function: direct interventional hard-endpoint gap	0 direct and 1 indirect source; direction profile: null
P4	contextual adjacent evidence: direct interventional hard-endpoint gap	0 direct and 40 indirect sources; direction profile: negative, null, unclear
P5	immune and inflammation: direct interventional hard-endpoint gap	0 direct and 2 indirect sources; direction profile: negative, null

Next-Study Design Recommendation

The next high-yield study for Retinal age AI should target the longevity evidence gap, pre-register the primary endpoint, separate clinical from mechanistic endpoints, preserve safety and adherence capture, and include an analysis plan that can falsify the current boundary-condition claim rather than only confirming a favorable direction. Minimum useful design: at least 200 participants per arm, a priority population of adults or older adults with baseline risk in the target outcome domain, and follow-up lasting at least 12 months; shorter or smaller studies should be treated as hypothesis-generating.

Evidence Snapshot

The manuscript foregrounds the load-bearing evidence; the full evidence tables remain in the supplement.

Classification Criteria

• Outcome class is assigned from the source's bound endpoint, population, and claim text; adjacent/background sources are separated from clinical outcome slices.
• Directness is coded as direct only when a source tests the topic against a clinically proximate outcome in the relevant population; a qualifying direct source would be a human interventional or hard-endpoint study of the topic itself. Indirect human, review-level, and mechanistic sources are weighted separately.
• Directional signal is counted within the assigned outcome class only. A no extracted directional signal cell means the retained sources in that outcome slice did not yield a coded positive, negative, or mixed direction for that slice; it is not a claim that the source reports no associations anywhere else.
• Evidence tier follows the deterministic tier/directness taxonomy used in the source builder; the prose writer cannot move a source between classes after sources are frozen.

Source Classification Map

Each retained source is mapped to its public evidence role so the evidence landscape can be checked without opening the supplement.

• Retinal Age as a Predictive Biomarker for Mortality Risk: outcome=longevity; directness=review; tier=B1; direction=unclear; claims=2.
• Systemic and local vascular features in branch retinal vein occlusion: analysis of the retinal age gap and crossing pattern: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=96.
• The Association of Retinal age gap with metabolic syndrome and inflammation: outcome=immune inflammation; directness=indirect; tier=B2; direction=negative; claims=77.
• Central obesity and its association with retinal age gap: insights from the UK Biobank study: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=unclear; claims=65.
• Accelerated retinal ageing and multimorbidity in middle-aged and older adults: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=negative; claims=61.
• Artificial intelligence-derived retinal age gap as a marker for reproductive aging in women: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=54.
• Are Dilated Fundus Examinations Needed for OCT-Guided Retreatment of Exudative Age-Related Macular Degeneration? A Prospective, Randomized, Pilot Study: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=53.
• Association of retinal age gap with chronic kidney disease and subsequent cardiovascular disease sequelae: a cross-sectional and longitudinal study from the UK Biobank: outcome=safety comorbidity; directness=indirect; tier=B2; direction=null; claims=53.
• Association between Subretinal Drusenoid Deposits and Age-Related Macular Degeneration in Multimodal Retinal Imaging: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=52.
• Evaluating the reproducibility of a deep learning algorithm for the prediction of retinal age: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=49.
• Associations of Metabolically Healthy Obesity and Retinal Age Gap: outcome=cardiometabolic; directness=indirect; tier=B2; direction=negative; claims=48.
• Association between the retinal age gap and systemic diseases in the Japanese population: the Nagahama study: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=42.
• Retinal age gap as a predictive biomarker of stroke risk: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=39.
• Retinal imaging technologies in cerebral malaria: a systematic review: outcome=contextual adjacent evidence; directness=review; tier=B2; direction=null; claims=38.
• Systematic review and meta-analysis of diagnostic accuracy of detection of any level of diabetic retinopathy using digital retinal imaging: outcome=contextual adjacent evidence; directness=review; tier=B2; direction=null; claims=31.
• The Effect of Experience on Visual Search Patterns in Retinal Imaging Analysis: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=27.
• Multimodal Retinal Imaging for Detection of Ischemic Stroke: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=27.
• Diagnosis of Chronic Kidney Disease Using Retinal Imaging and Urine Dipstick Data: Multimodal Deep Learning Approach: outcome=safety comorbidity; directness=indirect; tier=B2; direction=null; claims=26.
• Retinal age gap as a predictive biomarker of future risk of Parkinson’s disease: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=26.
• Physiological responses to retinopathy of prematurity screening: indirect ophthalmoscopy versus ultra-widefield retinal imaging: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=26.
• Retinal imaging demonstrates reduced capillary density in clinically unimpaired APOE ε4 gene carriers: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=22.
• Association between gamma-glutamyl transferase levels and the retinal age gap: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=22.
• The Eye as a Window to Brain Health: Can Retinal Imaging and AI Modeling Predict Alzheimer's Disease?: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=17.
• Deep learning retinal imaging model identifies poor brain health among older adults without dementia: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=17.
• Foundation model-driven distributed learning for enhanced retinal age prediction: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=17.
• Replication of the Association between Retinal Aging Clock Susceptibility Genes and Retinal Age Gap in an Asian Population: The Nagahama Study: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=16.
• Evaluating the clinical utility of multimodal large language models for detecting age-related macular degeneration from retinal imaging: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=15.
• Expanding Access to Retinal Imaging Through Patient-Operated Optical Coherence Tomography in a Veterans Affairs Retina Clinic: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=14.
• Retinal Imaging Biomarkers and Correlation to Systemic Disease Activity in Pediatric Sickle Cell Disease: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=13.
• Estimating biological age from retinal imaging: a scoping review: outcome=longevity; directness=review; tier=B2; direction=null; claims=12.
• Association between liver fibrosis’s noninvasive scores and retinal imaging changes: insights from NHANES: outcome=cardiometabolic; directness=indirect; tier=B2; direction=null; claims=11.
• Integration of peripheral blood-based systemic inflammatory indices and retinal imaging using interpretable machine learning for predicting anti-VEGF treatment response in macular edema secondary to retinal vein occlusion: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=9.
• A Machine Learning Prediction Model to Identify Individuals at Risk of 5-Year Incident Stroke Based on Retinal Imaging: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=8.
• Phenotypic screening and genetic insights for predicting major vascular-related diseases using retinal imaging: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=8.
• Development and Validation of a Novel Deep Learning-Based Model for Detection of Diabetic Kidney Disease from Retinal Imaging Using a Weighted Loss Method: outcome=cardiometabolic; directness=indirect; tier=B2; direction=null; claims=6.
• Use of artificial intelligence with retinal imaging in screening for diabetes-associated complications: systematic review: outcome=contextual adjacent evidence; directness=review; tier=B2; direction=null; claims=6.
• A Novel Foundation Model-Based Framework for Multimodal Retinal Age Prediction: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=4.
• Addressing chronic visual hallucination by multimodal retinal imaging: a CBS case: outcome=safety comorbidity; directness=indirect; tier=B2; direction=null; claims=4.
• Assessment of demographic bias in retinal age prediction machine learning models: outcome=contextual adjacent evidence; directness=indirect; tier=B2; direction=null; claims=3.
• Retinal Imaging-Based Oculomics: Artificial Intelligence as a Tool in the Diagnosis of Cardiovascular and Metabolic Diseases: outcome=longevity; directness=indirect; tier=B2; direction=null; claims=2.

Load-Bearing Included Studies

• Zhu 2020; Review / meta-analysis; tier=B1; directness=review; N=—; population=adults; endpoint=longevity; direction=unclear.
• Nonaka 2026; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=contextual adjacent evidence; direction=null; representative statistic=P < 0.001.
• Zhu 2023; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=immune inflammation; direction=negative; representative statistic=P < 0.001.
• Chen 2023; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=contextual adjacent evidence; direction=unclear; representative statistic=P = 0.001.
• Chen 2025; Observational; tier=B2; directness=indirect; N=—; population=older adults; endpoint=contextual adjacent evidence; direction=negative; representative statistic=P < 0.001.
• Miao 2025; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=contextual adjacent evidence; direction=null; representative statistic=P = 0.003.
• Solomon 2021; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=contextual adjacent evidence; direction=null.
• Wu 2024; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=safety comorbidity; direction=null; representative statistic=P < 0.001.
• Krytkowska 2023; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=contextual adjacent evidence; direction=null; representative statistic=P < 0.001.
• Zoellin 2024; Observational; tier=B2; directness=indirect; N=—; population=adults; endpoint=contextual adjacent evidence; direction=null; representative statistic=P < 0.001.

Load-Bearing Tensions

Additional corpus sources included animal/preclinical evidence; - Severity 3 null vs positive: Zhu 2020 vs Grimbly 2024; Zhu 2020 (unclear) vs Grimbly 2024 (null) on longevity

• Severity 3 null vs positive: Zhu 2020 vs Ghenciu 2024; Zhu 2020 (unclear) vs Ghenciu 2024 (null) on longevity
• Severity 3 null vs positive: Zhu 2020 vs Kitmiridou 2026; Zhu 2020 (unclear) vs Kitmiridou 2026 (null) on longevity
• Severity 3 null vs positive: Zhu 2023 vs Majimbi 2023; Zhu 2023 (negative) vs Majimbi 2023 (null) on immune inflammation
• Severity 3 null vs positive: Wilson 2023 vs Chen 2023; Wilson 2023 (null) vs Chen 2023 (unclear) on contextual other
• Severity 3 null vs positive: Wilson 2023 vs Chen 2025; Wilson 2023 (null) vs Chen 2025 (negative) on contextual other
• Severity 3 null vs positive: Chen 2023 vs Krytkowska 2023; Chen 2023 (unclear) vs Krytkowska 2023 (null) on contextual other
• Severity 3 null vs positive: Chen 2023 vs Girach 2024; Chen 2023 (unclear) vs Girach 2024 (null) on contextual other

Additional corpus sources included animal/preclinical evidence; additional corpus sources informed the synthesis without anchoring a foregrounded quantitative claim and are catalogued for completeness: Zeng 2024, Kamei 2025, Zhu 2022, Piyasena 2018, Gupta 2025, Zhao 2021, Bhak 2025, Purohit 2025, Yang 2025, Elahi 2021, Nielsen 2024, Awodiya 2025, Lam 2026, Komatsu 2026, Most 2025, Dogan 2026, Wang 2025, Li 2025, Govindaiah 2025, Lu 2025, Yang 2025b, Song 2025, Prayitnaningsih 2026, Zawadzki 2011, ONeill 2025, Novel 2025, Nielsen 2025, Jamshidiha 2025, Kamalzadeh 2025, Liao 2018, Wang 2025b, Ilanchezian 2025, Wang 2025c, Lombardo 2012, Alber 2020.

References

• Nonaka 2026. Systemic and local vascular features in branch retinal vein occlusion: analysis of the retinal age gap and crossing pattern. BMJ Open Ophthalmology, 2026. DOI: 10.1136/bmjophth-2025-002610. PMID: 41500614.
• Zhu 2023. The Association of Retinal age gap with metabolic syndrome and inflammation. Journal of Diabetes, 2023. DOI: 10.1111/1753-0407.13364. PMID: 36919192.
• Chen 2023. Central obesity and its association with retinal age gap: insights from the UK Biobank study. International Journal of Obesity (2005), 2023. DOI: 10.1038/s41366-023-01345-x. PMID: 37491535.
• Chen 2025. Accelerated retinal ageing and multimorbidity in middle-aged and older adults. GeroScience, 2025. DOI: 10.1007/s11357-025-01581-1. PMID: 40035945.
• Miao 2025. Artificial intelligence-derived retinal age gap as a marker for reproductive aging in women. NPJ Digital Medicine, 2025. DOI: 10.1038/s41746-025-01699-8. PMID: 40523954.
• Wu 2024. Association of retinal age gap with chronic kidney disease and subsequent cardiovascular disease sequelae: a cross-sectional and longitudinal study from the UK Biobank. Clinical Kidney Journal, 2024. DOI: 10.1093/ckj/sfae088. PMID: 38989278.
• Solomon 2021. Are Dilated Fundus Examinations Needed for OCT-Guided Retreatment of Exudative Age-Related Macular Degeneration? A Prospective, Randomized, Pilot Study. Clinical Ophthalmology (Auckland, N.Z.), 2021. DOI: 10.2147/OPTH.S315554. PMID: 34408396.
• Krytkowska 2023. Association between Subretinal Drusenoid Deposits and Age-Related Macular Degeneration in Multimodal Retinal Imaging. Journal of Clinical Medicine, 2023. DOI: 10.3390/jcm12247728. PMID: 38137797.
• Zoellin 2024. Evaluating the reproducibility of a deep learning algorithm for the prediction of retinal age. GeroScience, 2024. DOI: 10.1007/s11357-024-01445-0. PMID: 39589693.
• Zeng 2024. Associations of Metabolically Healthy Obesity and Retinal Age Gap. Translational Vision Science & Technology, 2024. DOI: 10.1167/tvst.13.11.26. PMID: 39570618.
• Majimbi 2023. In vivo retinal imaging is associated with cognitive decline, blood-brain barrier disruption and neuroinflammation in type 2 diabetic mice. Frontiers in Endocrinology, 2023. DOI: 10.3389/fendo.2023.1224418. PMID: 37850093.
• Kamei 2025. Association between the retinal age gap and systemic diseases in the Japanese population: the Nagahama study. Japanese Journal of Ophthalmology, 2025. DOI: 10.1007/s10384-025-01205-3. PMID: 40304887.
• Zhu 2022. Retinal age gap as a predictive biomarker of stroke risk. BMC Medicine, 2022. DOI: 10.1186/s12916-022-02620-w. PMID: 36447293.
• Wilson 2023. Retinal imaging technologies in cerebral malaria: a systematic review. Malaria Journal, 2023. DOI: 10.1186/s12936-023-04566-7. PMID: 37101295.
• Piyasena 2018. Systematic review and meta-analysis of diagnostic accuracy of detection of any level of diabetic retinopathy using digital retinal imaging. Systematic Reviews, 2018. DOI: 10.1186/s13643-018-0846-y. PMID: 30404665.
• Gupta 2025. The Effect of Experience on Visual Search Patterns in Retinal Imaging Analysis. Ophthalmic surgery, lasers & imaging retina, 2025. DOI: 10.3928/23258160-20250228-03. PMID: 40163634.
• Zhao 2021. Multimodal Retinal Imaging for Detection of Ischemic Stroke. Frontiers in Aging Neuroscience, 2021. DOI: 10.3389/fnagi.2021.615813. PMID: 33603658.
• Bhak 2025. Diagnosis of Chronic Kidney Disease Using Retinal Imaging and Urine Dipstick Data: Multimodal Deep Learning Approach. JMIR Medical Informatics, 2025. DOI: 10.2196/55825. PMID: 39924305.
• Purohit 2025. Physiological responses to retinopathy of prematurity screening: indirect ophthalmoscopy versus ultra-widefield retinal imaging. Pediatric Research, 2025. DOI: 10.1038/s41390-025-03906-4. PMID: 39948384.
• Hu 2022. Retinal age gap as a predictive biomarker of future risk of Parkinson’s disease. Age and Ageing, 2022. DOI: 10.1093/ageing/afac062. PMID: 35352798.
• Yang 2025. Association between gamma-glutamyl transferase levels and the retinal age gap. Frontiers in Physiology, 2025. DOI: 10.3389/fphys.2025.1601093. PMID: 40933305.
• Elahi 2021. Retinal imaging demonstrates reduced capillary density in clinically unimpaired APOE ε4 gene carriers. Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring, 2021. DOI: 10.1002/dad2.12181. PMID: 34013017.
• Nielsen 2024. Foundation model-driven distributed learning for enhanced retinal age prediction. Journal of the American Medical Informatics Association: JAMIA, 2024. DOI: 10.1093/jamia/ocae220. PMID: 39225790.
• Awodiya 2025. The Eye as a Window to Brain Health: Can Retinal Imaging and AI Modeling Predict Alzheimer's Disease?. Brain and Behavior, 2025. DOI: 10.1002/brb3.70890. PMID: 41030116.
• Lam 2026. Deep learning retinal imaging model identifies poor brain health among older adults without dementia. Cerebral Circulation - Cognition and Behavior, 2026. DOI: 10.1016/j.cccb.2026.100529. PMID: 41625481.
• Komatsu 2026. Replication of the Association between Retinal Aging Clock Susceptibility Genes and Retinal Age Gap in an Asian Population: The Nagahama Study. Ophthalmology Science, 2026. DOI: 10.1016/j.xops.2026.101131. PMID: 41958711.
• Most 2025. Evaluating the clinical utility of multimodal large language models for detecting age-related macular degeneration from retinal imaging. Scientific Reports, 2025. DOI: 10.1038/s41598-025-18306-1. PMID: 41006661.
• Dogan 2026. Expanding Access to Retinal Imaging Through Patient-Operated Optical Coherence Tomography in a Veterans Affairs Retina Clinic. Bioengineering, 2026. DOI: 10.3390/bioengineering13010061. PMID: 41595993.
• Hoyek 2025. Retinal Imaging Biomarkers and Correlation to Systemic Disease Activity in Pediatric Sickle Cell Disease. Ophthalmology Science, 2025. DOI: 10.1016/j.xops.2025.100774. PMID: 40458663.
• Grimbly 2024. Estimating biological age from retinal imaging: a scoping review. BMJ Open Ophthalmology, 2024. DOI: 10.1136/bmjophth-2024-001794. PMID: 39181547.
• Wang 2025. Association between liver fibrosis’s noninvasive scores and retinal imaging changes: insights from NHANES. Journal of Health, Population, and Nutrition, 2025. DOI: 10.1186/s41043-025-00805-6. PMID: 40022221.
• Li 2025. Integration of peripheral blood-based systemic inflammatory indices and retinal imaging using interpretable machine learning for predicting anti-VEGF treatment response in macular edema secondary to retinal vein occlusion. Frontiers in Cell and Developmental Biology, 2025. DOI: 10.3389/fcell.2025.1732963. PMID: 41532151.
• Govindaiah 2025. A Machine Learning Prediction Model to Identify Individuals at Risk of 5-Year Incident Stroke Based on Retinal Imaging. Sensors (Basel, Switzerland), 2025. DOI: 10.3390/s25061917. PMID: 40293071.
• Lu 2025. Phenotypic screening and genetic insights for predicting major vascular-related diseases using retinal imaging. NPJ Digital Medicine, 2025. DOI: 10.1038/s41746-025-01850-5. PMID: 40659721.
• Yang 2025b. Use of artificial intelligence with retinal imaging in screening for diabetes-associated complications: systematic review. eClinicalMedicine, 2025. DOI: 10.1016/j.eclinm.2025.103089. PMID: 40052065.
• Song 2025. High-resolution multimodal visible light optical coherence tomography and scanning laser ophthalmoscopy for in vivo neuronal and vascular retinal imaging in mice. Biomedical Optics Express, 2025. DOI: 10.1364/BOE.560539. PMID: 40677383.
• Prayitnaningsih 2026. Development and Validation of a Novel Deep Learning-Based Model for Detection of Diabetic Kidney Disease from Retinal Imaging Using a Weighted Loss Method. Clinical Ophthalmology (Auckland, N.Z.), 2026. DOI: 10.2147/OPTH.S586474. PMID: 42028097.
• Zawadzki 2011. Integrated adaptive optics optical coherence tomography and adaptive optics scanning laser ophthalmoscope system for simultaneous cellular resolution in vivo retinal imaging. Biomedical Optics Express, 2011. DOI: 10.1364/BOE.2.001674. PMID: 21698028.
• ONeill 2025. Addressing chronic visual hallucination by multimodal retinal imaging: a CBS case. Therapeutic Advances in Ophthalmology, 2025. DOI: 10.1177/25158414251320032. PMID: 40655237.
• Novel 2025. A Novel Foundation Model-Based Framework for Multimodal Retinal Age Prediction. IEEE Journal of Translational Engineering in Health and Medicine, 2025. DOI: 10.1109/JTEHM.2025.3576596. PMID: 40740833.
• Nielsen 2025. Assessment of demographic bias in retinal age prediction machine learning models. Frontiers in Artificial Intelligence, 2025. DOI: 10.3389/frai.2025.1653153. PMID: 41141904.
• Zhu 2020. Retinal Age as a Predictive Biomarker for Mortality Risk. medRxiv, 2020. DOI: 10.1101/2020.12.24.20248817.
• Girach 2024. Retinal imaging for the assessment of stroke risk: a systematic review. Journal of Neurology, 2024. DOI: 10.1007/s00415-023-12171-6. PMID: 38430271.
• Ghenciu 2024. Retinal Imaging-Based Oculomics: Artificial Intelligence as a Tool in the Diagnosis of Cardiovascular and Metabolic Diseases. Biomedicines, 2024. DOI: 10.3390/biomedicines12092150. PMID: 39335664.
• Jamshidiha 2025. An explainable transformer model for Alzheimer’s disease detection using retinal imaging. Scientific Reports, 2025. DOI: 10.1038/s41598-025-12498-2. PMID: 40702219.
• Kamalzadeh 2025. Through the eye to the heart: a scoping review of artificial intelligence in retinal imaging for cardiovascular disease assessment. BMC Medical Informatics and Decision Making, 2025. DOI: 10.1186/s12911-025-03300-4. PMID: 41299604.
• Liao 2018. Potential Utility of Retinal Imaging for Alzheimer’s Disease: A Review. Frontiers in Aging Neuroscience, 2018. DOI: 10.3389/fnagi.2018.00188. PMID: 29988470.
• Noah 2024. Retinal imaging with hand-held optical coherence tomography in older people with or without postoperative delirium after hip fracture surgery: A feasibility study. PLOS ONE, 2024. DOI: 10.1371/journal.pone.0305964. PMID: 39012893.
• Wang 2025b. Artificial intelligence-enhanced retinal imaging as a biomarker for systemic diseases. Theranostics, 2025. DOI: 10.7150/thno.100786. PMID: 40093903.
• Ilanchezian 2025. Development and validation of an AI algorithm to generate realistic and meaningful counterfactuals for retinal imaging based on diffusion models. PLOS Digital Health, 2025. DOI: 10.1371/journal.pdig.0000853. PMID: 40373008.
• Wang 2025c. Research advances on artificial intelligence assisted diagnosis and risk assessment in cardiovascular disease using retinal imaging. Frontiers in Cardiovascular Medicine, 2025. DOI: 10.3389/fcvm.2025.1615857. PMID: 41322502.
• Kitmiridou 2026. Retinal imaging in pre‐eclamptic pregnancy: systematic review. Ultrasound in Obstetrics & Gynecology, 2026. DOI: 10.1002/uog.70162. PMID: 41531381.
• Lombardo 2012. Adaptive Optics Technology for High-Resolution Retinal Imaging. Sensors (Basel, Switzerland), 2012. DOI: 10.3390/s130100334. PMID: 23271600.
• Alber 2020. A recommended “minimum data set” framework for SD‐OCT retinal image acquisition and analysis from the Atlas of Retinal Imaging in Alzheimer's Study (ARIAS). Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring, 2020. DOI: 10.1002/dad2.12119. PMID: 33163610.