{"id":"c5114368-e9c3-4736-bddb-c4d7fa19bcae","arxiv_id":"2507.20017","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"VAMPIRE predicts CVD risk and four blood-related conditions from OCTA images using a vessel-following Mamba module and morphology text enhancement, outperforming existing backbones on a new OCTA-CVD dataset.","lead":"This paper introduces a new dataset of OCTA retinal images from 843 patients and a deep learning model that predicts cardiovascular disease risk and related blood conditions from these images. The model combines a Mamba architecture that follows vessel paths with text descriptions generated by GPT-4o to improve prediction accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4's ablation baseline exactly matches VisionFM in Table 2, but VAMPIRE is described as Mamba-based; the module ablation does not isolate the proposed components, so the central claim is not yet supported.","rationale":"The reader's weakest assumption concerned SAM-OCTA segmentation reliability. That is a plausible external dependency, but the more directly checkable and load-bearing issue is internal inconsistency in the ablation: the Table 4 baseline exactly reproduces VisionFM's Table 2 metrics while Section 2.1 says VAMPIRE is Vim-based. The module attribution in Section 3.3 is therefore uninterpretable until the baseline backbone is clarified and re-run. If the ablation baseline is different from the architecture used in the main comparison, the claimed gains from MBD and IEM do not support the central claim. A secondary concern is that demographic features (age and gender) are concatenated with image features without a demographic-only control, which could also confound the comparison, but the ablation inconsistency is the primary issue because it can be settled by code inspection and directly affects the paper's main evidence.","tokens_in":8785,"tokens_out":7031,"duration_ms":87984,"concrete_test":"Inspect the released code at https://github.com/xmed-lab/VAMPIRE to identify the exact backbone used for the Table 4 'Baseline' row and confirm it matches the backbone of VAMPIRE with the MBD and IEM modules removed. If the baseline is VisionFM instead of the Vim-based architecture described in Section 2.1, re-run the full ablation (Baseline, w/ MBD, w/ IEM, Both) using the actual Mamba backbone and report new F1/AUC/AUPR values. Also verify that the 'Both' row in Table 4 corresponds to the same architecture as the VAMPIRE row in Table 2; if the numbers differ due to a typo, correct the table and re-evaluate the comparison.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3.3's ablation is the key evidence that the MBD and IEM modules improve performance. In Table 4, the 'Baseline' row reports F1=0.5522, AUC=0.8061, AUPR=0.6009, which are exactly the values reported for VisionFM in Table 2. Yet Section 2.1 describes VAMPIRE as a Mamba-based architecture following Vim [31], whose Table 2 row has F1=0.4851, AUC=0.7971, AUPR=0.5712. If the Table 4 baseline is VisionFM rather than the Mamba backbone used in the full model, then adding MBD and IEM to that baseline does not measure the modules' contribution to VAMPIRE, and the architecture described in the paper is inconsistent with the evaluated system. If the Table 4 baseline label is simply a typo, the printed ablation numbers are not interpretable as printed. Either way, the central claim that both modules significantly enhance VAMPIRE's CVD prediction is not currently supported by the reported experiments.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VAMPIRE, a multi-task deep learning framework that jointly predicts 10-year cardiovascular disease (CVD) risk and four CVD-related conditions (high blood glucose, cholesterol, triglycerides, and hypertension) from OCTA enface images. The framework combines a Mamba-Based Directional (MBD) module, which uses vessel segmentation maps and depth-first search to define a vessel-following scanning order for the state-space model, and an Information-Enhanced Morphological (IEM) module, which integrates textual vessel-shape descriptions generated by prompting GPT-4o with the outputs of a disease classifier. The authors introduce a new in-house dataset, OCTA-CVD, of 1,659 images from 843 patients, and report experiments with five-fold cross-validation and a second-hospital extension. The main claimed results are improvements over CNN backbones, OCTA-specific detection methods, and fine-tuned ophthalmologic foundation models on F1 score, AUC, and AUPR.","tokens_in":8970,"tokens_out":4984,"duration_ms":53182,"significance":"If the central claims were supported, the paper would make a useful contribution in at least three respects: it releases a first-of-its-kind OCTA dataset for CVD risk assessment; it introduces a multi-task paradigm that goes beyond binary risk classification and aligns with clinical blood-panel factors; and it proposes a vessel-aware scanning mechanism for Mamba that is conceptually well-motivated for vascular trajectory modeling. The external validation on a second hospital is a positive feature, as is the reported patient-level cross-validation split. However, the current evidence is weakened by a fundamental inconsistency in the ablation baseline (Table 4 exactly reproduces the VisionFM row of Table 2), which prevents the module contributions from being attributed to the described VAMPIRE architecture, and by the absence of any statistical significance testing for the claimed improvements. These issues are load-bearing for the paper's central claim that the MBD and IEM modules enhance VAMPIRE's CVD prediction.","major_comments":[{"comment":"The 'Baseline' row in Table 4 reports F1=0.5522, AUC=0.8061, AUPR=0.6009, which are exactly the values reported for VisionFM in Table 2. Section 2.1 describes VAMPIRE as a Mamba-based architecture following Vim [31], and Vim's Table 2 row is F1=0.4851, AUC=0.7971, AUPR=0.5712. Therefore the ablation baseline is not the backbone used in the described VAMPIRE model; it is a different model (VisionFM). Consequently, the reported improvements from adding MBD and IEM do not measure the modules' contribution to VAMPIRE, and the claim that both modules significantly enhance VAMPIRE's performance is not supported by the printed experiments. The authors must clarify the baseline architecture and re-run the ablation on the actual VAMPIRE backbone (or else explicitly describe and motivate a different composition).","section":"§3.3, Table 4 vs. Table 2"},{"comment":"The textual descriptions integrated in the IEM module are generated from a classification model trained on OCTA-500 that reads the same input OCTA image; the diagnostic results are then used to prompt GPT-4o for vessel morphology descriptions. This makes the 'morphological knowledge' a function of the input itself, rather than independent prior knowledge, which confounds the interpretation of the IEM module's contribution. The improvements attributed to IEM could stem from the auxiliary classifier's features or from label leakage, not from genuine vessel-shape knowledge. Please provide an analysis that separates these factors—for example, by using a fixed set of text prompts independent of the input, or by directly comparing against feeding the classifier's prediction into the model—to demonstrate that the IEM mechanism itself is responsible for the gains.","section":"§2.3 (IEM Module)"},{"comment":"The text states that applying MBD and IEM 'significantly enhance' performance, but no statistical significance tests are reported. For instance, the IEM F1 improvement over baseline in Table 4 is 0.0579, while the standard deviations are 0.0319 and 0.0255, making the effect within one standard deviation. Overlapping standard errors and the absence of paired tests (e.g., across folds, or bootstrap confidence intervals) mean the current claims of significance are not supported. Please add appropriate statistical tests or confidence intervals for the key comparisons in Tables 2 and 4, and temper the language accordingly.","section":"§3.3 (significance claims)"}],"minor_comments":[{"comment":"The sentence 'applying MBD and IEM independently can still improve the overall result by 6.4%, 2.92% in F1 score and 2.92%, 3.89% in AUPR' is numerically inconsistent with Table 4: the F1 improvements are 0.0640 (MBD) and 0.0579 (IEM), while the 2.92% and 3.89% are the AUPR improvements for MBD and IEM, respectively. Please correct the text to match the table.","section":"§3.3"},{"comment":"The column header 'NTG' appears to be a typo for 'HTG' (high triglycerides).","section":"Table 3"},{"comment":"The supplementary hospital dataset is described only as '765 images from 379 patients'; no demographic breakdown, acquisition protocol, or image layer details are provided, which limits the interpretability of the external validation in Table 3.","section":"§3.2 (Extended Experiment)"},{"comment":"The phrase 'As proved by [28]' is too strong for a single epidemiological association study; consider replacing with 'as shown by' or 'as suggested by'.","section":"Introduction"},{"comment":"There is a typographical double period in 'mean age of 45±12 years..'.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The ablation baseline inconsistency in Table 4 is the most serious issue; it is likely a fixable error (e.g., relabeling or rerunning on the correct backbone), but as printed it makes the central module-effectiveness claim unsupported. The IEM circularity concern is also substantive and would benefit from a controlled experiment. The dataset release and multi-task formulation are potentially valuable, and the external validation is a strength. If the authors can clarify the architecture and provide significance testing, the paper could become acceptable. No concerns about author conduct are implied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the OCTA-CVD dataset, and that is worth something. A 1,659-image, 843-patient OCTA collection with 10-year CVD risk and four blood-factor labels, split at the patient level and backed by an external 765-image validation set, is a real resource. The multi-task framing on OCTA is also a reasonable step beyond binary risk classification. The reported trends are consistent: VAMPIRE beats the listed CNN, OCTA-specific, and foundation-model baselines on most metrics, and the direction of the gains is plausible.\n\nBut the stress-test note is right, and it lands hard. Table 4, the ablation that supposedly isolates the MBD and IEM modules, has a baseline row with F1=0.5522, AUC=0.8061, AUPR=0.6009. Those are the exact numbers for VisionFM in Table 2. The paper describes VAMPIRE as built on Vim, whose Table 2 row is F1=0.4851. So the ablation compares the modules against VisionFM, not against the Mamba backbone actually used in the full model. That means the reported module contributions are not attributable to the architecture in the paper. It could be a typo in the table label, but as printed, the experiment does not support the claim that MBD and IEM improve VAMPIRE. This is the load-bearing evidence, and it is broken.\n\nOther soft spots are less severe but still real. There are no significance tests; some gains fall within overlapping standard errors, though the F1 and AUPR differences from VisionFM in Table 2 look larger than the error bars. The IEM module generates text from a classifier that reads the same input image, so the 'morphological knowledge' is not external information, and the ablation on IEM alone could simply reflect extra capacity. Table 2 also shows the model incorporates age and gender as inputs, but there is no demographic-only baseline, so how much of the predictive signal comes from the image is unclear. The dataset is called 'available' but hospital data are likely access-restricted; the authors should say exactly what is released.\n\nWho is this for? Researchers working on OCTA-based screening or retinal biomarkers for CVD. They will find the dataset and the multi-task formulation useful even if the model contributions remain muddy. The paper deserves a serious referee, but it needs a corrected ablation tied to the real backbone, significance testing, a demographic-only baseline, and a clearer data-access statement before the central claim can be trusted.","headline":"Useful new OCTA dataset and a sensible multi-task framing, but the module ablation is broken: the Table 4 baseline is VisionFM, not the Mamba backbone the model is built on, so the central claim is currently unsupported.","tokens_in":9507,"tokens_out":2665,"would_cite":false,"duration_ms":34877,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A vessel-following OCTA model jointly predicts 10-year CVD risk and four blood-factor conditions, beating general and ophthalmologic baselines.","keywords":["OCTA","cardiovascular disease risk prediction","Mamba","vessel trajectory","vessel morphology","multimodal language model","retinal imaging","OCTA-CVD dataset"],"falsifier":"Run VAMPIRE with the vessel-following DFS scan replaced by a fixed raster or diagonal scan, keeping the IEM module and all training settings identical: the paper's mechanism predicts a clear drop in F1 and AUPR, so an experiment showing no such drop would falsify the claim that vascular trajectory order is what drives the gain.","tokens_in":8597,"feed_emoji":"🫀","tokens_out":8411,"duration_ms":89606,"temperature":0.7,"pith_summary":"Retinal vessels mirror coronary circulation, and OCT angiography shows them in far more detail than fundus photos or OCT. The paper's claim is that a model can use that detail to do more than binary high/low CVD risk: VAMPIRE jointly predicts 10-year CVD risk and four CVD-related conditions (high blood glucose, high blood cholesterol, high blood triglycerides, and high blood pressure). To make this work, the authors introduce the OCTA-CVD dataset of 1,659 en face projection images from 843 patients and a two-part architecture that follows vascular trajectories and injects vessel-morphology descriptions. The paper reports that VAMPIRE surpasses standard CNN and ViT backbones, OCTA-based detection models, and fine-tuned ophthalmologic foundation models, with the advantage also appearing on a second hospital's 765 images. If the results hold, non-invasive OCTA screening can give patients actionable blood-factor warnings rather than a single risk label.","feed_headline":"OCTA vessel-tracking model beats foundation models for CVD risk","feed_subtitle":"Scanning OCTA along real vessel paths plus vessel-shape text lifts F1 and AUC on five CVD targets.","key_machinery":"The load-bearing object is the vessel-following scanning order inside a Mamba (selective state-space) backbone. Instead of raster or diagonal patch scans, the MBD module takes a SAM-OCTA vessel segmentation map, filters noise, runs depth-first search over the vessel tree, and concatenates patch blocks along each branch, with background patches inserted between branches, yielding $S = \\operatorname{concat}_{i,j \\in T}[V_i; B_{i,j}; V_j]$. This gives the state-space model a sequence whose local order is vascular continuity. The second component, the IEM module, computes a disease-diagnosis prompt from an OCTA-500-trained classifier, asks a multimodal language model for vessel-morphology descriptions, encodes them with a frozen text encoder, and fuses them into image features with cross-attention. The two modules together are what the ablation study credits for the performance gain.","core_discovery":"The central discovery is that vascular direction and vessel morphology, extracted explicitly from OCTA en face images, are stronger predictors of CVD risk factors than generic image features learned by general-purpose backbones. VAMPIRE's Mamba-Based Directional module uses a pre-trained vessel segmentation to build a scanning sequence $S = \\operatorname{concat}_{i,j \\in T}[V_i; B_{i,j}; V_j]$ that follows continuous vessel branches, so each patch is contextualized along a real vascular trajectory; its Information-Enhanced Morphological module adds text descriptions of vessel shape, generated from a disease classifier and a multimodal language model, through cross-attention. The paper argues these two signals are complementary, and its experiments on the new OCTA-CVD dataset report F1 0.627, AUPR 0.647, and AUC 0.824, above fine-tuned ophthalmologic foundation models, with consistent AUC gains across all five risk categories on an external hospital set.","pith_inferences":["The same vessel-following Mamba scan could transfer to other tubular-structure medical images, such as coronary angiography or OCTA-based diabetic retinopathy, where continuity along a branch is the informative axis.","Ablating the MLLM-generated text with fixed, hand-written vessel-shape templates would isolate whether the gain comes from semantic morphology knowledge or from the extra text-conditioning signal itself.","Because OCTA-CVD skews female (74.3%) and young (mean age 45), a demographic subgroup analysis on older or male-heavy cohorts would test whether the reported margins persist in the populations most at risk.","The joint-task framing suggests any vascular biomarker label, such as vessel caliber, could serve as an auxiliary task; testing additional auxiliary outputs would reveal whether the benefit is specific to blood-factor labels or a general multi-task effect."],"forward_implications":["Jointly predicting four blood-factor conditions alongside 10-year CVD risk improves the F1 score by more than 4% over binary-only risk classification, so clinical screening can report which factor is abnormal, not just high/low risk.","On OCTA-CVD, VAMPIRE reaches F1 0.627, AUPR 0.647, and AUC 0.824, beating CNN backbones, OCTA-specific detectors, and fine-tuned ophthalmologic foundation models.","The ablation shows each module alone lifts F1 and AUPR, and both together give the best result, implying trajectory and morphology are complementary signals.","The improvement transfers to a second hospital's 765 images, where VAMPIRE achieves the highest AUC on all five categories against ViT, RETFound, and VisionFM.","Adding the IEM module to other backbones improves their performance too, so vessel-shape descriptions act as portable prior knowledge rather than being tied to VAMPIRE's architecture."],"supporting_citations":[{"why":"Supplies the SAM-OCTA pre-trained segmentation network whose vessel maps define the depth-first search scanning order.","marker":"[25]"},{"why":"Supplies the bidirectional Mamba (Vim) backbone and the patch-sequence formulation that the MBD scanning modifies.","marker":"[31]"},{"why":"Supplies the multimodal language model used to generate vessel-morphology text descriptions for the IEM module.","marker":"[20]"},{"why":"Supplies OCTA-500, the dataset used to train the retinal-disease classifier that conditions the text descriptions.","marker":"[12]"},{"why":"Provides RETFound, a fine-tuned ophthalmologic foundation-model baseline that VAMPIRE must surpass.","marker":"[30]"},{"why":"Provides VisionFM, another ophthalmologic foundation-model baseline and the strongest competitor in the comparison tables.","marker":"[22]"},{"why":"Provides Eye-AD, an OCTA-based detection method used as a baseline.","marker":"[8]"},{"why":"Provides PolarNet, an OCTA-based detection method used as a baseline.","marker":"[17]"},{"why":"Supplies the selective state-space Mamba sequence model whose long-range modeling motivates the trajectory-following design.","marker":"[7]"}],"fun_headline_variants":["Vessel path and shape text lift CVD risk prediction from OCTA","VAMPIRE tracks vessel trajectories to predict CVD risk factors","New OCTA dataset enables vessel-aware CVD risk prediction","Vessel morphology text boosts OCTA-based CVD risk models","VAMPIRE outperforms foundation models on OCTA CVD risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing premise is that the SAM-OCTA segmentation maps are reliable enough that depth-first traversal yields clean, continuous vessel trajectories; if the maps are fragmented or noisy, the vessel-following scan order will not match real vascular anatomy and the MBD module's contribution will shrink.","fun_headline_variants_meta":{"raw":{"variants":["Vessel path and shape text lift CVD risk prediction from OCTA","VAMPIRE tracks vessel trajectories to predict CVD risk factors","New OCTA dataset enables vessel-aware CVD risk prediction","Vessel morphology text boosts OCTA-based CVD risk models","VAMPIRE outperforms foundation models on OCTA CVD risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001064,"raw_usage":{"total_tokens":4503,"prompt_tokens":1033,"completion_tokens":3470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":3385}},"tokens_in":649,"tokens_out":3470,"duration_ms":24500,"temperature":1.0,"reasoning_tokens":3385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:49:54.564760+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VAMPIRE with the vessel-following DFS scan replaced by a fixed raster or diagonal scan, keeping the IEM module and all training settings identical: the paper's mechanism predicts a clear drop in F1 and AUPR, so an experiment showing no such drop would falsify the claim that vascular trajectory order is what drives the gain.","supporting_citations":[{"cited_title":"In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)","cited_arxiv_id":null,"evidence_quote":"Supplies the SAM-OCTA pre-trained segmentation network whose vessel maps define the depth-first search scanning order."},{"cited_title":"In: Forty-first International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies the bidirectional Mamba (Vim) backbone and the patch-sequence formulation that the MBD scanning modifies."},{"cited_title":"Medical image analysis93, 103092 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies OCTA-500, the dataset used to train the retinal-disease classifier that conditions the text descriptions."},{"cited_title":"New England Journal of Medicine – Artificial Intelligence1(12), AIoa2300221 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides VisionFM, another ophthalmologic foundation-model baseline and the strongest competitor in the comparison tables."},{"cited_title":"npj Digital Medicine7(1), 294 (2024) 10 L","cited_arxiv_id":null,"evidence_quote":"Provides Eye-AD, an OCTA-based detection method used as a baseline."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Provides PolarNet, an OCTA-based detection method used as a baseline."}],"review_version":1}