Pith. sign in

REVIEW 4 major objections 5 minor 9 references

This paper claims that cardiovascular risk can be predicted from a single low-dose chest CT by explicitly reasoning from pulmonary findings to cardiac mechanisms, with AUCs of 0.919 for screening and 0.838 for mortality.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 23:13 UTC pith:4MJLQOH6

load-bearing objection Promising integration with a real leakage risk: the cross-disease reasoning claim needs a lung-specific perception check before the AUCs mean what they seem. the 4 major comments →

arxiv 2511.06625 v5 pith:4MJLQOH6 submitted 2025-11-10 cs.CV cs.AIcs.LG

Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from Low-Dose Computed Tomography

Yifei Zhang , Jiashuo Zhang , Mojtaba Safari , Xiaofeng Yang , Liang Zhao This is my paper
classification cs.CV cs.AIcs.LG
keywords low-dose CTcardiovascular disease predictionpulmonary-to-cardiac reasoningexplainable AIopportunistic screeningmultimodal fusionagentic reasoningmedical image analysis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper is trying to establish that a single low-dose chest CT, already used for lung cancer screening, can also predict cardiovascular risk—and that the best way to do so is to explicitly model the physiological chain from lung abnormalities to cardiac stress, rather than treating heart and lungs as separate tasks. The proposed framework first summarizes lung findings, then uses a knowledge-guided agentic reasoning step to turn those findings into a natural-language cardiovascular judgment with rationale, and finally fuses that text with a heart-centered 3D image embedding and a frozen lung-risk score. On a large multi-site lung-screening cohort, the paper reports AUCs of 0.919 for cardiovascular event screening and 0.838 for cardiovascular mortality, above cardiac-specific, single-disease, and foundation-model baselines. The authors interpret the results as evidence that the reasoning stage adds signal that raw image features and raw lung findings do not, and that the resulting model is more auditable because it states why a lung finding would matter for the heart. A sympathetic reader would care because, if true, this makes dual-purpose screening available from an existing scan and provides a template for interpretable cross-disease imaging.

Core claim

The central claim is that cardiovascular risk can be read from a lung-cancer screening CT by forcing the model through a constrained clinical-information pathway: perceive pulmonary abnormalities, reason through established lung–heart mechanisms, then fuse that reasoning with localized cardiac morphology and a malignancy risk prior. The authors report that the reasoning stage is not a decorative wrapper: adding raw lung findings without reasoning yields only marginal gains over the cardiac branch, whereas adding the reasoning stage raises screening AUC from 0.904 to 0.919 and mortality AUC from 0.806 to 0.824. The output is a binary CVD risk score accompanied by a generated rationale that tr

What carries the argument

The load-bearing component is the agentic pulmonary-to-cardiac reasoner: a three-step chain in which a pulmonary perception agent converts the CT volume into structured lung findings, a knowledge reasoning agent recalls physiological mechanisms linking lung to heart, and a diagnostic reasoning agent synthesizes those mechanisms into a CVD judgment with a natural-language rationale. This textual reasoning is encoded and concatenated with a heart-centered 3D CNN embedding and a frozen lung-cancer risk score; the fusion head is trained end-to-end. The comparison between variants that include versus omit this reasoning text is what isolates the paper's claimed cross-disease contribution.

Load-bearing premise

The pulmonary perception module must summarize only lung abnormalities; if it secretly uses cardiac structures visible in the same CT images, the measured benefit of the reasoning stage could come from cardiac information, not cross-disease reasoning.

What would settle it

Mask the cardiac silhouette in the CT before the pulmonary perception step and rerun the Reasoning-Only variant. If its AUC stays near 0.832, the pulmonary signal is independent; if it falls toward the cardiac-only AUC, the reasoning pathway was reading the heart. A second probe: replace the knowledge-generated rationale with a fixed-length template and measure the AUC change; no drop would indicate the text content itself is not carrying the signal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Lung cancer screening CTs could be reused to identify high cardiovascular risk, adding a second clinically actionable output from a scan people already receive.
  • The generated rationales give clinicians a concrete text to check, so a risk score is no longer a black-box number; it comes with a claim about which lung findings drive the risk.
  • The ablation pattern—raw findings add little, reasoned findings add more—implies that how pulmonary information is represented matters as much as whether it is used.
  • Mortality prediction up to an AUC of 0.838 suggests the framework captures prognostic information, not just association with prevalent disease.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A decisive outside test would mask the cardiac region before generating pulmonary findings; if the reasoning branch still holds, the cross-disease pathway is real, and if it drops, the vision-language perception step may be leaking cardiac information.
  • The same perception–knowledge–reasoning template could be ported to other opportunistic cross-organ predictions (e.g., liver fat to cardiac risk), a generalization the paper does not claim.
  • The explicit rationales could be used to define testable physiological hypotheses—for example, that patients with fibrosis-plus-effusion show higher right-ventricular strain on dedicated imaging—connecting the model to independent measurements.
  • A compression probe replacing the generated rationale with a fixed-length template of the same text length would show whether the semantic content of the rationale is what drives the gain or merely the presence of an additional text feature.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes an Explainable Cross-Disease Reasoning Framework for cardiovascular risk prediction from low-dose chest CT (LDCT). The framework combines a frozen lung-cancer risk prior (Sybil), a CT-CLIP-based pulmonary perception module, an agentic LLM-based pulmonary-to-cardiac reasoning module, and a localized cardiac 3D CNN encoder, fused by a trainable head. On the NLST cohort, the authors report AUCs of 0.919 for CVD screening and 0.838 for CVD mortality, claiming state-of-the-art performance. Ablations across four MedicalNet backbones show monotonic gains as lung risk and reasoning text are added to the cardiac branch, and visualizations are presented as evidence of physiologically grounded reasoning.

Significance. If the reported results are reproducible and the claimed mechanisms are validated, the framework would be a meaningful advance in opportunistic cardiovascular risk assessment from LDCT, with an interpretable reasoning pathway that goes beyond purely image-based discriminative models. The paper has notable strengths: a large NLST cohort (32,136 scans), a subject-disjoint split, evaluation across multiple 3D backbones, and an ablation design that isolates the contribution of each module. The proposed architecture is clearly motivated by the clinical literature linking pulmonary findings to cardiovascular outcomes. However, the central claim of state-of-the-art performance and the interpretation of the reasoning module as 'cross-disease' rely on several unverified assumptions, including comparability of baselines taken from prior publications, the pulmonary specificity of the CT-CLIP perception step, and the existence of the targeted controls promised in the abstract. These issues need to be resolved before the findings can be accepted at face value.

major comments (4)
  1. [Table 1, §4.2.1] The caption states that AE+SVM, DeepCAC, and KAMP-Net baseline numbers are 'taken from published results.' These AUCs were not computed on the same subject-disjoint test split used for the proposed method, so the comparison is not controlled. This directly affects the claim that the framework 'outperforms' established baselines. For CVD mortality, the reported 95% CI for the best model (ResNet-200, 0.838 [0.815, 0.859]) overlaps with the M3FM CI (0.816 [0.759, 0.874]), so the mortality gain is not statistically evident. The authors should re-run all baselines on their exact test split (or justify why published numbers are compatible) and report paired significance tests (e.g., DeLong).
  2. [§3.2, §4.1.4, Eq. (3)] The Pulmonary Perception Agent is implemented with CT-CLIP operating on the full LDCT volume. CT-CLIP is a generalist 3D CT vision-language model, not a lung-specific analyzer, and chest CT reports commonly include cardiac findings (cardiomegaly, aortic calcification, pericardial effusion). Therefore the structured findings R in Eq. (3) may contain cardiac information, and the 'Reasoning Only' branch (Table 3, AUC 0.832/0.776) may be driven by direct cardiac signal rather than by cross-disease pulmonary-to-cardiac reasoning. The novelty claim in the title and abstract depends on this distinction. The authors should either restrict the perception step to a lung mask/ROI, demonstrate empirically that CT-CLIP findings are exclusively or predominantly pulmonary, or compare against a lung-segmented variant to show that cardiac leakage does not explain the reasoning gains.
  3. [Abstract and §4.2] The abstract states: 'Targeted controls indicate that the gains are not explained by additional thoracic visual features alone, fixed rule propagation, or a single reasoning backend.' No such controls appear in the full text. Section 4.2.5 compares 'Lung-Findings' (concatenation) against 'Reasoning,' but it does not implement a fixed-rule propagation baseline or a single-reasoning-backend control. This is an unsupported empirical claim in the abstract and must either be backed by the corresponding experiments or removed/qualified.
  4. [§4.1.3, Tables 2–4] Statistical rigor is insufficient for the central comparison and for the ablation increments. Only bootstrapped CIs are reported; no DeLong or other paired tests are provided. For the mortality endpoint, the AUC differences across models are often within the reported CIs. Additionally, the dataset contains 32,136 scans from 10,395 subjects, and the positive/negative counts (6,339/25,797) sum to the total scan count, indicating scan-level labels with multiple scans per subject. The bootstrap should account for subject-level clustering to avoid overconfident CIs. Reporting clustered paired tests would materially strengthen the conclusions.
minor comments (5)
  1. [Eq. (5)] The notation '(h, w, d)(H, W, D)' is missing a relation symbol; it should be '(h, w, d) ≪ (H, W, D)' to indicate subvolume dimensions.
  2. [References] The ClinicalBERT citation in §4.1.4 is listed as Wang et al. (2023), but the cited reference is 'Optimized glycemic control of type 2 diabetes with reinforcement learning,' which is not the ClinicalBERT model description. The appropriate citation should be provided.
  3. [§4.1.4] The exact prompts, model versions (GPT-DeepResearch, GPT-5-nano), and sampling parameters are not provided. This limits reproducibility of the reasoning branch. Please include prompts and version identifiers in an appendix or supplement.
  4. [§3.3, Eq. (4)] Notation is inconsistent: the text uses z_lung for the lung-risk embedding, while Eq. (4) defines \hat{y}_lung; clarify how the temporal trajectory \hat{y}_lung is aggregated when used in the fusion vector of Eq. (1).
  5. [Figure 5] The textual attribution heatmap is described as coming from 'token-level attribution,' but no method is specified. Please describe how the token attributions are computed (e.g., gradients, attention, or another method).

Circularity Check

0 steps flagged

No significant circularity: prediction is a trained fusion over external pretrained features; missing targeted controls and CT-CLIP anatomy specificity are correctness concerns, not circularity.

full rationale

The derivation chain is not circular. Equation (1) defines the CVD prediction as a trained sigmoid over a concatenation of three embeddings: cardiac (MedicalNet on the localized subvolume), reasoning (ClinicalBERT encoding of the LLM-generated text), and lung-risk (frozen Sybil). Equation (2) is standard BCE supervised training; no equation sets y_hat equal to any input by construction. The reasoning text E is produced zero-shot by GPT-DeepResearch/GPT-5-nano from CT-CLIP pulmonary findings, so it is an external feature, not a fitted parameter or a renamed training label. The 'Reasoning Only' ablation is a supervised classifier on that text; even though E contains a CVD judgment, the judgment is not the NLST label and the LLM is not fine-tuned on those labels, so this is feature-leakage/construct-validity risk, not derivation-level circularity. Baselines are external published methods (DeepCAC, KAMP-Net, Tri2D-Net, Sybil, M3FM); no load-bearing self-citation or imported uniqueness theorem appears. Two legitimate concerns do not change the circularity score: the abstract promises 'targeted controls' (additional thoracic visual features, fixed rule propagation, single reasoning backend) that are absent from the full text, and the CT-CLIP-based 'pulmonary' perception may encode cardiac structures visible in the same LDCT. Both are omitted-support or measurement-validity issues; neither involves an equation reducing the prediction to its input or a self-citation chain forcing the result.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework introduces no new physical or biological entities; all components are borrowed from prior models. The main epistemic commitments are domain assumptions about the specificity of the perception module, the validity of the causal pathways, and the comparability of external baselines. The free parameters are mostly hand-chosen thresholds and unspecified architecture details that affect reproducibility.

free parameters (4)
  • Finding confidence threshold = 0.5
    Findings with s_i >= 0.5 are retained to form the pulmonary report (§3.2), chosen by hand without sensitivity analysis.
  • Cardiac ROI size = 128^3 voxels
    The cardiac subvolume is cropped to 128^3 voxels (§4.1.4); no ablation on ROI size is shown.
  • Encoder and fusion dimensions
    ClinicalBERT output dimension, cardiac feature dimension, and MLP head dimensions in Eq. (1) are unspecified, which are needed for reproduction.
  • LLM prompt templates and sampling parameters
    Prompts for GPT-DeepResearch and GPT-5-nano and their decoding settings are not described, making the reasoning outputs irreproducible.
axioms (4)
  • domain assumption The causal pathways used by the knowledge reasoning agent (emphysema→hypoxemia→pulmonary hypertension; fibrosis→inflammation→endothelial injury) are correct and relevant for this cohort.
    The knowledge reasoning agent propagates these mechanisms (§3.2) without clinical validation in the NLST population; if the physiology is wrong, the reasoning text encodes misleading cues.
  • domain assumption CT-CLIP's structured findings from the full LDCT volume are pulmonary-specific and exclude cardiac features.
    The Pulmonary Perception Agent runs on the whole scan (§4.1.4), so cardiac structures are visible; the paper assumes findings isolate lung abnormalities, which is load-bearing for the reasoning interpretation.
  • domain assumption Per-scan labels taken from subject-level outcomes are valid despite multiple scans per subject.
    The cohort has 32,136 scans from 10,395 subjects (§4.1.1); within-subject correlation is not modeled in AUC/CI estimation, which can inflate apparent discrimination.
  • ad hoc to paper Published baseline AUCs (AE+SVM, DeepCAC, KAMP-Net) are comparable to the authors' test split.
    Table 1 takes baseline numbers from published results (§4.2.1), which may use different splits, inclusion criteria, or follow-up definitions, so the comparison is not strictly head-to-head.

pith-pipeline@v1.3.0-alltime-deepseek · 14279 in / 13329 out tokens · 133764 ms · 2026-08-03T23:13:06.986649+00:00 · methodology

0 comments
read the original abstract

Low-dose chest computed tomography (LDCT) captures pulmonary and cardiac structures in a single scan, enabling joint assessment of lung and cardiovascular health. Existing approaches typically model these domains independently and do not explicitly represent their physiological interactions. We propose an Explainable Cross-Disease Reasoning Framework for cardiovascular risk assessment from LDCT. The framework follows a constrained clinical-information pathway: it extracts pulmonary findings, grounds cross-organ mechanisms in medical knowledge, and produces a cardiovascular prediction with a natural-language rationale. It combines four components: a frozen lung-risk prior, a pulmonary perception module, an agentic reasoning module, and a cardiac subvolume feature extractor. Their outputs are fused to integrate localized cardiac evidence with mechanism-level pulmonary context. On the National Lung Screening Trial cohort, the framework achieves an AUC of 0.919 for CVD screening and up to 0.838 for CVD mortality prediction, outperforming cardiac-specific, single-disease, and foundation-model baselines. Targeted controls indicate that the gains are not explained by additional thoracic visual features alone, fixed rule propagation, or a single reasoning backend. The proposed framework thus provides an auditable approach to cross-disease cardiovascular risk assessment from LDCT.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 5 linked inside Pith

  1. [3]

    Developing gener- alist foundation models from a multimodal dataset for 3d computed tomography.arXiv preprint arXiv:2403.17834,

    Ibrahim Ethem Hamamci, Sezgin Er, Chenyu Wang, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Doga, Omer Faruk Durugol, Weicheng Dai, Murong Xu, et al. Developing gener- alist foundation models from a multimodal dataset for 3d computed tomography.arXiv preprint arXiv:2403.17834,

    arXiv

  2. [5]

    Cpathagent: An agent-based foundation model for interpretable high-resolution pathology image analysis mimicking pathologists’ diagnostic logic.arXiv preprint arXiv:2505.20510,

    Yuxuan Sun, Yixuan Si, Chenglu Zhu, Kai Zhang, Zhongyi Shui, Bowen Ding, Tao Lin, and Lin Yang. Cpathagent: An agent-based foundation model for interpretable high-resolution pathology image analysis mimicking pathologists’ diagnostic logic.arXiv preprint arXiv:2505.20510,

    arXiv

  3. [8]

    Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs.arXiv preprint arXiv:2303.00915,

    Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs.arXiv preprint arXiv:2303.00915,

    Pith/arXiv arXiv

  4. [9]

    How wellcanmodernllmsactasagentcoresinradiologyenvironments?arXivpreprintarXiv:2412.09529,

    Qiaoyu Zheng, Chaoyi Wu, Pengcheng Qiu, Lisong Dai, Ya Zhang, Yanfeng Wang, and Weidi Xie. How wellcanmodernllmsactasagentcoresinradiologyenvironments?arXivpreprintarXiv:2412.09529,

    Pith/arXiv arXiv

  5. [2017]

    Agent hospital: A simulacrum of hospital with evolvable medical agents

    Junkai Li, Yunghwei Lai, Weitao Li, Jingyi Ren, Meng Zhang, Xinhui Kang, Siyu Wang, Peng Li, Ya-Qin Zhang, Weizhi Ma, et al. Agent hospital: A simulacrum of hospital with evolvable medical agents. arXiv preprint arXiv:2405.02957,

    Pith/arXiv arXiv

  6. [2019]

    Chexagent: Towards a foundation model for chest x-ray interpretation.arXiv preprint arXiv:2401.12208,

    Zhihong Chen, Maya Varma, Jean-Benoit Delbrouck, Magdalini Paschali, Louis Blankemeier, Dave Van Veen, Jeya Maria Jose Valanarasu, Alaa Youssef, Joseph Paul Cohen, Eduardo Pontes Reis, 15 Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from LDCT et al. Chexagent: Towards a foundation model for chest x-ray interpretation.arXiv prepri...

    Pith/arXiv arXiv

  7. [2021]

    Med3d: Transfer learning for 3d medical image analysis

    Sihong Chen, Kai Ma, and Yefeng Zheng. Med3d: Transfer learning for 3d medical image analysis. arXiv preprint arXiv:1904.00625,

    Pith/arXiv arXiv 1904

  8. [2022]

    Cardiovasculardiseases(cvds),July2025

    WorldHealthOrganization. Cardiovasculardiseases(cvds),July2025. URL https://www.who.int/ news-room/fact-sheets/detail/cardiovascular-diseases-(cvds). Accessed: 2025- 09-09. Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Hui Hui, Yanfeng Wang, and Weidi Xie. Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data.Nature Communica-...

    2025

  9. [2023]

    Medclip: Contrastive learning from unpaired medical images and text

    Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. InProceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, volume 2022, page 3876,

    2022