{"id":"1015a7d7-a617-4a29-a36d-af57e7a48f7d","arxiv_id":"2505.20810","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"This is a narrative review of AI-enhanced retinal imaging for detecting systemic diseases; it presents no new data and is undermined by multiple citation errors.","lead":"This paper reviews how AI and retinal imaging are used to detect diseases like diabetes, Alzheimer's, and heart disease. It compiles many studies, but contains frequent citation errors, a nonsensical claim, and a self-promotional section on vessel segmentation.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central novelty claim is contradicted by its own cited references: reference [29] (Wagner et al. 2020) is a prior oculomics review spanning neurological and cardiovascular disease, undercutting the 'no prior review' assertion.","rationale":"The reader's REJECT verdict remains appropriate, but my stress-test identifies a sharper, more fundamental flaw than the specific citation-number mismatches highlighted in the reader's weakest_assumption. The central claim is a novelty assertion: that no prior review has unified AI-enhanced retinal oculomics across neurological and cardiovascular domains. That assertion is directly falsified by the paper's own reference list, most notably Wagner et al. (2020), which is an oculomics review spanning systemic diseases including both categories. This is not a matter of imprecise reporting of a performance figure; it is the collapse of the paper's stated reason for existing. Even if every individual study summary were accurate, a review that is not the first cross-domain synthesis and that offers no reproducible search methodology cannot support the advertised contribution. The reader's examples of mismatched citations (e.g., hemoglobin prediction attributed to a central serous chorioretinopathy paper) are symptoms of the same broader unreliability, but the novelty contradiction is the load-bearing failure. Therefore, I recommend keeping the reader's REJECT verdict unchanged, and I partially agree with the reader's identification of the weakest assumption: the general assumption that cited studies support the claims is violated, and one particularly consequential violation is the contradiction of the novelty claim by the paper's own references.","tokens_in":36319,"tokens_out":4387,"duration_ms":44108,"concrete_test":"Read reference [29] (Wagner et al., Transl Vis Sci Technol 2020) and reference [62] (Li et al., Eur Heart J Digit Health 2024). Determine whether either review covers both neurological and cardiovascular disease prediction from retinal imaging. If either does, the Section 1 claim that 'No prior review has systematically unified AI-enhanced retinal oculomics across both neurological and cardiovascular disease domains' is false. A simple verification is to search each paper for disease-domain coverage (e.g., Alzheimer's, stroke, coronary artery disease, hypertension) and note the presence of both categories in a single review.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated contribution is the Section 1 claim that 'No prior review has systematically unified AI-enhanced retinal oculomics across both neurological and cardiovascular disease domains to provide an integrated framework for clinical translation.' This is the central claim on which the review's novelty rests. It is directly contradicted by the paper's own bibliography. Reference [29] is S. K. Wagner et al., 'Insights into systemic disease through retinal imaging-based oculomics' (Translational Vision Science & Technology, 2020), a review that explicitly surveys retinal imaging-based oculomics across multiple systemic disease domains, including cardiovascular and neurological conditions. Reference [62] is Li et al., 'Prediction of cardiovascular markers and diseases using retinal fundus images and deep learning: A systematic scoping review' (European Heart Journal Digital Health, 2024), which systematically reviews deep-learning retinal imaging for cardiovascular markers/diseases, and reference [30] (Barriada & Masip, Diagnostics 2022) reviews deep-learning methods for cardiovascular risk assessment from retinal images. The 'no prior review' assertion is thus falsifiable from the paper's own citation list. Additionally, the paper does not provide a reproducible search method for its 'systematic' synthesis, so the novelty claim cannot be defended as a precise scope distinction. The load-bearing condition—that this is the first cross-domain synthesis—fails, and with it the paper's primary contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a narrative review claiming to be the first systematic unification of AI-enhanced retinal oculomics across neurological and cardiovascular disease domains, with additional coverage of metabolic, renal, hepatobiliary, and hematological applications. It surveys imaging modalities (fundus photography, OCT/OCTA, adaptive optics), AI/ML approaches, and clinical-translation challenges, and proposes a roadmap for standardization, external validation, and workflow integration. The paper includes several summary tables of reported performance figures and a cross-domain challenge table.","tokens_in":36566,"tokens_out":1857,"duration_ms":19292,"significance":"If the review's synthesis and performance figures were reliable, a cross-domain survey of retinal oculomics could be a useful resource for researchers and clinicians. The paper explicitly identifies important challenges such as external validation, dataset bias, and protocol standardization, and it correctly emphasizes the need for prospective multicenter studies. However, the evidentiary value of the review is heavily compromised by numerous mismatches between cited references and the claims they are supposed to support, and the stated novelty is contradicted by references within the paper's own bibliography. The tables therefore cannot be trusted as accurate summaries of the literature, and the central contribution is not established.","major_comments":[{"comment":"The novelty claim that 'No prior review has systematically unified AI-enhanced retinal oculomics across both neurological and cardiovascular disease domains' is contradicted by the paper's own references. Reference [29] is Wagner et al., 'Insights into systemic disease through retinal imaging-based oculomics' (Translational Vision Science & Technology, 2020), a review that explicitly spans cardiovascular and neurological conditions. Reference [30] (Barriada & Masip) reviews deep-learning methods for cardiovascular risk assessment from retinal images, and reference [62] is a systematic scoping review of retinal fundus imaging for cardiovascular markers. The authors must either remove the 'no prior review' claim or restrict the novelty to a precise, reproducible scope that remains after acknowledging these works.","section":"§1, p. 4 (end of Introduction)"},{"comment":"The claim that 'Zhang et al. [18]' reported fundus-image-based models with low R2 (0.06–0.57) for hemoglobin and RBC counts is unsupported. Reference [18] is X. Zhang, C. Z. F. Lim, J. Chhablani, Y. M. Wong, 'Central serous chorioretinopathy: Updates in the pathogenesis, diagnosis and therapeutic strategies' (Eye and Vision, 2023), which contains no hemoglobin prediction results. The R2 values appear to be misattributed, and the same reference is later used in §7.2 for CKD risk classification, where it is also inappropriate. Without a correct citation, the hematological parameter discussion and Table 4 lose their evidentiary basis.","section":"§6.1, Table 4 and surrounding text"},{"comment":"The statement that 'Lieb et al. [81]' analyzed retinal images from diabetic patients to classify atherosclerotic risk categories and predicted atherosclerosis up to five years in advance is contradicted by reference [81], which is W. Lieb and R. S. Vasan, 'Genetics of coronary artery disease' (Circulation, 2013). This is a genetics review, not a retinal imaging study. This misattribution undermines the credibility of the AI-for-cardiovascular-risk section and indicates that the reference list does not reliably correspond to the claims made.","section":"§5.3, paragraph on Lieb et al."},{"comment":"This subsection is largely a list of the authors' own segmentation networks (references [105]–[117]), with more than ten self-citations and no comparative assessment against other published methods or benchmarks. The text describes general benefits of vessel segmentation but does not critically evaluate these specific methods or justify their selection over alternative approaches. At minimum, the authors should replace this self-citation-heavy passage with a balanced discussion that includes non-self references and quantitative comparisons, or remove it from a review whose scope is systemic disease detection.","section":"§7.5, 'Importance of Vessels Segmentaion'"},{"comment":"The abstract and introduction describe the paper as a systematic synthesis, but no search strategy, inclusion criteria, or PRISMA-style methodology is provided. Given that the paper claims to 'systematically synthesize' a broad literature and quantify diagnostic performance, the absence of a reproducible search method makes the selection of references and the reported performance figures unverifiable. This is a load-bearing methodological gap for a review that asserts systematic coverage.","section":"Overall, 'systematic' synthesis and methods"}],"minor_comments":[{"comment":"The sentence 'A theoretical model proposed by Sabanayagam [35] suggests that nearly all individuals with PD exhibit telangiectasia in the mega-reflex' is unclear and appears to misattribute a retinal imaging finding; the cited reference [35] is a chronic kidney disease deep learning paper, not a Parkinson's disease study.","section":"§3.2.1"},{"comment":"'Retinal Omics' is used to mean multimodal imaging and AI analysis, which is a nonstandard use of the term 'omics'; this should be defined or replaced with clearer terminology.","section":"§3.3"},{"comment":"The text cites 'Wang et al. [29]' for a mechanistic connection between oxidative-stress-driven nephron dysfunction and CAD-related stenosis, but reference [29] is Wagner et al., an oculomics review; the citation does not match the stated claim.","section":"§5.2.1"},{"comment":"The text attributes both 'Almonte et al. [89]' observations to the same reference, but the second finding (microvascular density loss in generalized anxiety disorder) is not supported by the title/scope of the cited Almonte and Capellà review; please verify and correct the citation.","section":"§6.2"},{"comment":"The reported AUROC of 0.93 with sensitivity 93.2% and specificity 82.0% is given without confidence intervals or external-validation context; consider adding these details or tempering the summary statement.","section":"Table 2 (Cheung et al. row)"},{"comment":"Several reference numbers are used inconsistently: for example, [78] is cited for a meta-analysis by McGeechan et al., but the reference list at [78] is McGeechan et al., so numbering is correct, yet the same number is also used in §5.2.1 for 'Wang et al. (2018)'. Please recheck all citation numbering and ensure each claim is supported by the cited source.","section":"Throughout"},{"comment":"The concluding section introduces a sentence about pregnant women with pro-atherosclerotic changes that is not connected to any prior analysis in the review; either elaborate or remove this tangential statement.","section":"§9 Conclusions"}],"recommendation":"reject","confidential_remarks":"The paper is a review with a significant number of citation–claim mismatches that directly affect its central content (e.g., misattribution of hemoglobin R2 values to a central serous chorioretinopathy paper, and misattribution of a retinal-imaging atherosclerosis study to a genetics paper). These are not isolated typos: they affect the tables and performance summaries that form the paper's main contribution. In addition, the novelty claim is contradicted by the paper's own references. Given that the evidentiary basis is compromised and no systematic methodology is provided, rejection is appropriate. The authors could revise and resubmit a substantially different manuscript with corrected citations, a defensible novelty claim, and a clear search methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a review, and as a review it fails where it must succeed: accurate reporting of the literature. The central claim that no prior review has unified AI-enhanced retinal oculomics across neurological and cardiovascular domains is directly contradicted by the paper's own reference [29], Wagner et al.'s 'Insights into systemic disease through retinal imaging-based oculomics.' That alone sinks the novelty claim. The stress-test note is right, and I see the same problem on reading the text.\n\nCredit where it's due: the paper does cover an admirably broad range of systemic diseases, and the summary tables and figures (Table 2, Table 3, Figure 1) give a passable birds-eye map of the field. The cross-domain challenges table (Table 6) is a reasonable checklist of obstacles. If the references were accurate and the contribution reframed as a tutorial overview rather than a novel synthesis, there would be some value for a clinician wanting a quick orientation.\n\nThe soft spots are not minor. Section 6.1 cites [18], a central serous chorioretinopathy paper, for hemoglobin prediction R² values—a plain mis-citation. Section 5.3 cites [81], a coronary artery disease genetics paper, for an AI atherosclerosis prediction result. Section 3.2.1 attributes a 'mega-reflex' model to Sabanayagam [35], which is a chronic kidney disease paper and the sentence is garbled. Section 3.2.2 contains an empty citation. And Section 7.5 is a list of the authors' own segmentation networks presented as key enabling tools, with no comparative assessment whatsoever—it reads as self-promotion, not synthesis. These are not edge cases; they are the evidentiary backbone of the review.\n\nThere is no reproducible search method, so the 'systematic' label is not defensible. The organization is also messy, with Section 8 and 9 somewhat redundant and misnumbered subsections.\n\nMy recommendation: do not send this to peer review in its current form. The citation errors and false novelty claim would waste referee time. I would desk-reject with an invitation to resubmit after a thorough rewrite: each claim re-checked against its source, the contribution honestly scoped, and the self-citation section either removed or replaced with a comparative evaluation. If the authors fix those issues, this could become a serviceable, if not groundbreaking, review.","headline":"A broad but carelessly assembled review whose central novelty claim is contradicted by its own references and whose citation errors undermine its evidentiary value.","tokens_in":37116,"tokens_out":1800,"would_cite":false,"duration_ms":20156,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review argues that AI-enhanced retinal imaging can be developed into a unified, non-invasive screening platform for life-threatening systemic diseases, and positions itself as the first cross-domain synthesis of retinal oculomics…","keywords":["retinal imaging","oculomics","deep learning","cardiovascular risk prediction","neurodegenerative biomarkers","optical coherence tomography","systemic disease screening","clinical translation"],"falsifier":"Check whether reference [18], cited for hemoglobin and red-blood-cell prediction, is actually a study of central serous chorioretinopathy, and whether reference [81], cited for five-year atherosclerosis risk prediction, is a coronary artery disease genetics paper; if either mismatch is confirmed, the corresponding claims in Sections 6.1 and 5.3 lack the stated evidentiary support.","tokens_in":36100,"feed_emoji":"👁️","tokens_out":3795,"duration_ms":41082,"temperature":0.7,"pith_summary":"The paper tries to establish that the retina is a practical window into systemic health, and that AI applied to retinal images can detect early signs of cardiovascular, neurodegenerative, metabolic, and hematological diseases. It claims to be the first review to unify AI-enhanced retinal oculomics across both neurological and cardiovascular domains into a single framework for clinical translation. If correct, routine eye scans could eventually supplement blood tests, brain imaging, and cardiovascular risk scores, enabling earlier and cheaper screening. The review synthesizes performance figures from many studies, including deep-learning sensitivity above 90 percent for diabetic retinopathy and an AUC of 0.89 for cardiovascular risk prediction from fundus photos.","feed_headline":"AI eye scans could screen for heart disease and dementia early","feed_subtitle":"Retinal blood vessels and nerve layers mirror cardiovascular and neurodegenerative disease, the review finds.","key_machinery":"The central object is oculomics: the treatment of the retina as a surrogate tissue whose microvasculature and neural layers mirror systemic and cerebral pathology. The engine is AI/deep learning, which learns to map fundus photographs, OCT, and OCTA scans to disease labels and risk factors without relying only on predefined features. The imaging modalities—fundus photography, optical coherence tomography, OCT angiography, and adaptive optics—supply the high-resolution structural and vascular measurements that the AI models analyze.","core_discovery":"On its own terms, the paper's central discovery is that retinal biomarkers—arteriolar narrowing, venular dilation, retinal nerve fiber layer thinning, microvascular density loss, and even amyloid deposits—are consistently linked to systemic disease across multiple organ systems, and that modern imaging plus deep learning can extract these signals at scale. The authors assert that no prior review has systematically unified AI-enhanced retinal oculomics across neurological and cardiovascular disease domains, and they offer this review as the integrated framework that maps modality strengths, quantifies diagnostic performance, and outlines a roadmap for multicenter standardization and prospective validation. Their conclusion is that retinal imaging, powered by AI, is poised to become a cornerstone of precision medicine for early detection and risk stratification.","pith_inferences":["The claim that this is the first unified review is a literature-positioning statement; its truth depends on the search scope and inclusion criteria, which the paper does not independently verify against a systematic protocol.","Because the review aggregates performance numbers from heterogeneous, mostly retrospective studies without meta-analysis, the pooled figures should be treated as indicative ranges rather than settled effect sizes.","If retinal oculomics matures, the same imaging infrastructure could be reused across multiple specialties, potentially changing the cost structure of population screening for cardiovascular and neurodegenerative disease.","A testable extension would be a prospective study asking whether adding an AI retinal risk score to conventional risk models improves reclassification of patients for statin therapy or dementia workup."],"forward_implications":["Retinal fundus photography could be added to primary-care screening for cardiovascular risk, since deep learning models already predict risk factors like age, blood pressure, and major adverse cardiovascular events from such images.","OCT-based measures of retinal nerve fiber layer and macular thickness could become non-invasive biomarkers for early Alzheimer's disease and for tracking multiple sclerosis progression.","AI analysis of retinal images could support non-invasive screening for anemia, chronic kidney disease, and liver disease, though the review notes predictive accuracy for continuous lab values such as hemoglobin remains modest.","Standardized imaging protocols and external validation in diverse populations would be required before any of these retinal biomarkers enter routine clinical workflows.","A unified cross-domain framework could accelerate translation by aligning regulatory, interoperability, and data-sharing standards across cardiology, neurology, nephrology, and primary care."],"supporting_citations":[{"why":"Supplies the deep-learning evidence that cardiovascular risk factors and major adverse cardiovascular events can be predicted from retinal fundus photographs.","marker":"[15]"},{"why":"Provides the deep-learning algorithm for chronic kidney disease detection from retinal photographs in community populations, a key renal oculomics anchor.","marker":"[35]"},{"why":"Establishes the retinal amyloid plaque finding in Alzheimer's patients and in vivo imaging in a mouse model, underpinning the AD biomarker discussion.","marker":"[55]"},{"why":"Introduces the oculomics concept of deriving systemic disease insights from retinal imaging, framing the review's central perspective.","marker":"[29]"},{"why":"Supplies evidence linking retinal microvascular abnormalities to incident stroke, a central cardiovascular claim of the review.","marker":"[12]"},{"why":"Demonstrates deep-learning detection of anemia from fundus images, supporting the hematological prediction section.","marker":"[14]"},{"why":"Meta-analysis showing retinal nerve fiber layer and macular thinning in Alzheimer's disease and mild cognitive impairment, grounding the neurodegenerative biomarker claims.","marker":"[66]"},{"why":"Meta-analysis of retinal layer segmentation in multiple sclerosis, supporting OCT as a prognostic biomarker for disability worsening.","marker":"[71]"}],"fun_headline_variants":["AI eye scans catch heart disease and dementia early","Retinal AI predicts heart disease and dementia risk","Deep learning reads eye scans for early disease signs","AI on retinal images spots early heart and brain disease"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review's load-bearing premise is that each cited study actually supports the specific statement attached to it, so the performance figures, disease associations, and summary tables rest on accurate sourcing.","fun_headline_variants_meta":{"raw":{"variants":["AI eye scans catch heart disease and dementia early","Retinal AI predicts heart disease and dementia risk","Deep learning reads eye scans for early disease signs","AI on retinal images spots early heart and brain disease"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1756,"prompt_tokens":954,"completion_tokens":802,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":742}},"tokens_in":570,"tokens_out":802,"duration_ms":9272,"temperature":1.0,"reasoning_tokens":742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:45:52.760321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether reference [18], cited for hemoglobin and red-blood-cell prediction, is actually a study of central serous chorioretinopathy, and whether reference [81], cited for five-year atherosclerosis risk prediction, is a coronary artery disease genetics paper; if either mismatch is confirmed, the corresponding claims in Sections 6.1 and 5.3 lack the stated evidentiary support.","supporting_citations":[{"cited_title":"Sabanayagam, D","cited_arxiv_id":null,"evidence_quote":"Provides the deep-learning algorithm for chronic kidney disease detection from retinal photographs in community populations, a key renal oculomics anchor."},{"cited_title":"Mitani, A","cited_arxiv_id":null,"evidence_quote":"Demonstrates deep-learning detection of anemia from fundus images, supporting the hematological prediction section."},{"cited_title":"den Haan, F","cited_arxiv_id":null,"evidence_quote":"Meta-analysis showing retinal nerve fiber layer and macular thinning in Alzheimer's disease and mild cognitive impairment, grounding the neurodegenerative biomarker claims."},{"cited_title":"Petzold, L","cited_arxiv_id":null,"evidence_quote":"Meta-analysis of retinal layer segmentation in multiple sclerosis, supporting OCT as a prognostic biomarker for disability worsening."}],"review_version":1}