Pith. sign in

REVIEW 4 major objections 4 minor 21 references

MOSCARD -- Causal Reasoning and De-confounding for Multimodal Opportunistic Screening of Cardiovascular Adverse Events

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fusing chest X-ray and ECG with causal de-confounding predicts heart-event risk better than single-modality baselines across internal, emergency, and ICU populations.

desk verdict Real candidate for peer review once temporal ordering is pinned down and causal claims are dialed back. read the letter →

arxiv 2506.19174 v1 pith:3RTCZYS5 submitted 2025-06-23 cs.CV

Jialu Pi , Juan Maria Farina , Rimita Lahiri , Jiwoong Jeong , Archana Gurudu , Hyung-Bok Park , Chieh-Ju Chao , Chadi Ayoub
show 2 more authors
Reza Arsanjani Imon Banerjee
This is my paper
classification cs.CV
keywords MajoradversecardiovasculareventsMultimodallearningChestX-rayElectrocardiogramCausalreasoningDe-confoundingOpportunisticscreeningDomainshift
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Major adverse cardiovascular events (MACE) — heart attacks, strokes, and related outcomes — are the leading cause of death worldwide. The paper proposes MOSCARD, a predictive framework that combines a chest X-ray (CXR) and a 12-lead electrocardiogram (ECG) from the same patient to estimate short-term (6-month) and long-term (5-year) MACE risk from routine check-up data. MOSCARD trains each modality with a confusion loss that suppresses age and gender confounders, then aligns the two modalities with a co-attention mechanism in which ECG guides attention over the X-ray, and adds causal-reasoning branches that model comorbidity pathways. On internal data, the combined-model area under the ROC curve reaches 0.75 for one-year MACE; on a shifted emergency-department population the CXR branch reaches 0.83, and on an external intensive-care cohort the ECG branch reaches 0.71, beating single-modality and standard multimodal baselines. The authors argue this makes cost-effective opportunistic screening possible, because the required images are already collected during routine visits.

What carries the argument

Three mechanisms carry the argument. First, de-confounding: each modality is encoded by a Vision Transformer whose early-layer features predict age and gender, and a confusion loss pushes those predictions toward a uniform distribution, removing the X←C→Y backdoor path. Second, co-attention alignment: ECG features serve as the query and CXR features as both key and value, so the fused representation selectively attends to X-ray regions most relevant given the electrical signal. Third, causal reasoning: a structural causal model with a directed acyclic graph connects image X, outcome Y, causal factors A (comorbidities) and confounder C, and auxiliary branches predict causal factor labels so the main branch learns the desired X→Y relationship rather than spurious correlations. The whole is trained in two steps, so if one modality is missing the de-confounder-only encoder can still be used.

What would settle it

Look up the acquisition timestamps of the CXR and ECG pairs in the internal cohort and compare them with the documented MACE event dates; if any pairs used for the 6-month prediction were acquired after the event, re-run the 6-month analysis restricted to pre-event images and check whether the AUC remains near 0.75 (internal) and 0.71 (external).

Watch

Extended reading notes

Core claim

The central claim is that a multimodal model explicitly designed to remove confounder shortcuts and to model causal structure can outperform both single-modality encoders and generic contrastive or knowledge-distillation fusion for MACE risk prediction, and that the gains are largest when the test population differs from the training population. In the paper's own comparison, the best internal result (AUC 0.750 at one year, combined CXR+ECG) comes from the full MOSCARD model with confounder and causal branches, while single-modality CXR reaches 0.739 at five years; on the emergency-department shift cohort the causal CXR branch reaches AUC 0.837 at one year, and on the external ICU cohort the ECG branch reaches AUC 0.71 at six months. The paper also reports that ECG-only predictions improve after alignment, that de-confounded encoders generalize better on external data, and that subgroup performance is balanced across demographics and comorbidities.

Load-bearing premise

The load-bearing premise is that every chest X-ray and ECG pair was acquired before the MACE outcome it is used to predict; the paper only specifies a ±6-month matching window, not the ordering, so for the 6-month endpoint a post-event image would leak the label and inflate the reported AUC.

Editorial extensions

If this is right

  • Routine chest X-rays and ECGs already captured during health check-ups could be repurposed to flag patients at elevated 6-month and 5-year MACE risk without additional testing.
  • De-confounding improves generalization to lower-risk populations, suggesting that standard single-modality models rely on shortcut features that do not transfer.
  • ECG-only predictions improve after co-attention alignment, so cross-modal guidance can transfer useful information even when one modality is weaker.
  • The two-step design means the de-confounded single-modality encoder remains usable when paired images are unavailable, preserving a fallback screening path.
  • If the reported AUCs hold, MOSCARD-style screening could identify high- and medium-risk patients earlier, enabling lifestyle and medical intervention before events occur.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct next test is to enforce strict temporal ordering: the paper matches CXR and ECG within a ±6-month window but does not state that imaging predates the 6-month outcome, so checking and re-running with pre-event images only would settle whether label leakage inflates the short-horizon AUC.
  • The causal chain the model encodes is a simplified linear disease pathway; testing alternative graphs (or learned causal structures) could reveal which comorbidity links actually drive the performance gain.
  • The same ECG-guided co-attention idea could transfer to other paired routine modalities, such as retinal photographs with blood pressure or ultrasound with lab values, whenever one time-series modality can guide spatial attention in another.
  • Calibration and decision-curve analysis would tell whether the AUC improvement translates into changed clinical actions, since opportunistic screening needs well-calibrated risk estimates, not just ranking accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes MOSCARD, a two-step multimodal framework that combines chest X-ray (CXR) and 12-lead ECG encoders with co-attention, a confusion-loss de-confounding branch, and a structural causal model for predicting Major Adverse Cardiovascular Events (MACE). The authors evaluate the model on a large multi-site internal cohort (12,612 patients), an emergency-department shift cohort, and a small external MIMIC-IV cohort, reporting AUCs of 0.75, 0.83, and 0.71 for internal, ED, and MIMIC settings, respectively, and claiming superiority over single-modality and multimodal baselines (MedCLIP and ALBEF). The paper also reports subgroup performance and saliency-map analyses.

Significance. If the central claims hold, MOSCARD would be a practically useful opportunistic screening tool: it uses routinely collected CXR and ECG, addresses confounding via a dedicated branch, and shows promising generalization to shifted external populations. The manuscript's strengths include the large multi-site internal cohort, the inclusion of two external test sets, the comparison against established multimodal baselines, and the subgroup and saliency analyses. However, three load-bearing issues currently prevent acceptance of the prognostic and causal claims: the temporal order of imaging and outcomes is not established, the causal factor label Y_ca is never defined, and the external headline numbers are selected as the best branch/variant per dataset. These issues are fixable with additional analyses and clarifications, but they are central rather than cosmetic.

major comments (4)
  1. [Section 3, Internal cohort] The cohort definition states that CXR and 12-lead ECG are matched within a +/-6 months time window, but it never states that both studies were acquired before the MACE outcome used as the label. For the 6-month endpoint, an ECG (or CXR) recorded after the event can contain acute or post-event findings, and the saliency analysis shows the model attends to such findings; this is label leakage rather than prognosis. The authors must define the index date, verify that every image pair precedes the outcome, and report a sensitivity analysis restricted to pairs acquired before the event; without this, the reported AUCs do not support the prognostic claim.
  2. [Section 2.2 and Eq. (L_total)] The causal factor label Y_ca is used throughout the loss L_total = L_causal_CXR + L_main_CXR + L_causal_ECG + L_main_ECG + L_main_concat, but it is never defined in the manuscript. Neither Section 2 nor Section 3 lists which comorbidities constitute A/Y_ca, how they were extracted, or how the DAG X<-A->Y was validated. Without this definition, the causal branch is unidentifiable and the causal-reasoning contribution cannot be evaluated; the authors should specify the causal factor set, its measurement, and preferably a causal-identification or placebo test.
  3. [Section 3, Table 2] The external performance numbers in the abstract and text are the best branch/model variant per dataset selected post hoc: ED uses the CXR branch without confounder (AUC 0.837) and MIMIC uses the ECG branch without confounder (AUC 0.710), while the internal best is the combined branch with confounder (AUC 0.750). The text's claim that 'MOSCARD with causality achieves the best performance on both internal and external datasets' is not supported by any pre-specified model-selection rule, and reporting the best variant per dataset inflates performance under multiple comparisons. The authors should report a fixed primary branch and variant, and treat de-confounding/no-de-confounding as an ablation rather than selecting per dataset.
  4. [Section 3, External (MIMIC)] The MIMIC external set contains only 175 patients with matched ECGs, and the 6-month MACE outcome is curated by parsing clinical notes and inpatient mortality. The outcome-curation procedure is described too briefly to exclude ascertainment bias, and with n=175 the comparison across ECG versus CXR branches is unstable despite the bootstrapped intervals. The authors should report the number of MACE events, describe the note-parsing procedure, and provide confidence intervals for the between-branch and between-model differences.
minor comments (4)
  1. [Section 2.2] The co-attention equations are inconsistent: the first formula uses Q_CXR and K_E (CXR query, ECG key), while the text describes ECG as query and CXR as key/value; the symbols W_E_q and W_CXR_k appear in a transposed order in the second expression. Please align the notation.
  2. [Section 3, Results] There is a typo in the phrase 'MODSCARD (ECG branch)' which should read 'MOSCARD (ECG branch)'; also, 'Mace' in the table headers should be capitalized as 'MACE'.
  3. [Section 1, Abstract] The abstract cites 'Global Disease Burden Study 2021' but reference [11] is the GBD 2017 report; please update the citation or the wording to match.
  4. [Section 3 and Fig. 2] The claim of optimal 'fair' or equalized performance across subgroups is supported only by a spoke plot with no numeric subgroup metrics or statistical tests; please add quantitative subgroup AUCs and confidence intervals.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline external AUCs are post-hoc selected maxima over three output branches, so the 'our model outperformed' claim is partly by construction.

  1. fitted input called prediction [Abstract and Section 3 (Results), Table 2; Discussion]
    "Evaluated on internal, shift data from emergency department (ED) and external MIMIC datasets, our model outperformed single modality and state-of-the-art foundational models - AUC: 0.75, 0.83, 0.71 respectively. ... MOSCARD with causality (CXR branch) achieves the best performance on external ED dataset(AUC = 0.837), while MODSCARD (ECG branch) performs best on MIMIC (AUC = 0.71) for MACE prediction at 6 months."

    The three headline AUCs correspond to three different model variants in Table 2: internal combined+confounder at 1yr (0.750), ED CXR-only without confounder at 1yr (0.837), and MIMIC ECG-only without confounder at 6M (0.710). Each external number is the maximum of the combined/CXR/ECG branches evaluated on that dataset; the variant label is selected after seeing the test-set AUCs. The reported external performance is therefore, by construction, the maximum over that selection, not the out-of-sample performance of a single fixed model. The conclusion that the causal/de-confounding variants generalize is partially forced by choosing whichever branch wins on each dataset, and the abstract mixes different follow-up horizons as if they were one model's results.

full rationale

The paper is an empirical ML evaluation rather than a first-principles derivation: the encoders are trained and evaluated on held-out and external data, so the core predictive results are not circular. The self-citations (refs 2, 8, 9, all including Banerjee) are background for shortcut learning and adversarial debiasing and are not load-bearing for the central claim. The main circularity-adjacent issue is the reporting of post-hoc selected best branches as if they were a single model's external AUCs; this is an evaluation-selection flaw that makes the headline superiority partly by construction. The undefined Yca ('causal factor' label) and the postulated DAG mean the causal claim is assumed rather than tested, but that is incompleteness, not circularity. The temporal-leakage concern (CXR-ECG matched within +/-6 months possibly after the 6-month MACE endpoint) is a serious validity threat to the AUCs, but it would inflate scores empirically rather than make the derivation equivalent to its inputs, so it does not by itself raise the circularity score.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claims rest on the assumed causal graph, the undefined causal label Yca, and the choice of age/sex as confounders. These are not derived or validated; they are modeling assumptions introduced in Section 2. The free parameters are design choices that affect the reported AUCs.

free parameters (3)
  • Confusion loss weight alpha
    Coefficient for the confusion loss term in LossConf (Section 2.1); value not reported, yet it controls the strength of debiasing.
  • Imaging-outcome matching window = +/-6 months
    Threshold chosen for pairing CXR and ECG (Section 3); if imaging occurred after the event it can leak the outcome.
  • Transformer layer selection for confounder branch = early layers
    Choice of which ViT layers feed the confounder predictor (Section 2.1); not justified or ablated.
assumptions (5)
  • domain assumption Causal graph X <- A -> Y and X <- C -> Y correctly represents the data generating process
    Assumed in Section 2 and Fig. 1; no test of faithfulness or identifiability.
  • domain assumption Age and gender are the salient confounders C, and removing them improves generalizability
    Used to define LossConf in Section 2.1; no sensitivity analysis or support from the data.
  • ad hoc to paper The causal factor label set Yca is well-defined and measurable
    Yca appears in the five-branch loss (Section 2.2) but is never defined in the paper.
  • domain assumption Co-attention mechanism from pathology/genomics transfers to CXR and ECG alignment
    Adapted from [7] in Section 2.2 without validation on this modality pair.
  • standard math Cross-entropy loss and softmax are standard
    Standard machine learning components used throughout.
invented entities (1)
  • Causal factor A / causal label Yca
    purpose: Model the causal chain from comorbidities to MACE and guide feature learning
    No definition, no measurement, and no ablation. It is a placeholder for a causal mechanism that is never operationalized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MOSCARD -- Causal Reasoning and De-confounding for Multimodal Opportunistic Screening of Cardiovascular Adverse Events." pith.science (2026). https://pith.science/paper/3RTCZYS5

@misc{pith2026250619174,
  author       = {Pith},
  title        = {Pith review of: MOSCARD -- Causal Reasoning and De-confounding for Multimodal Opportunistic Screening of Cardiovascular Adverse Events},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3RTCZYS5}},
  note         = {Machine review of arXiv:2506.19174}
}
read the original abstract

Major Adverse Cardiovascular Events (MACE) remain the leading cause of mortality globally, as reported in the Global Disease Burden Study 2021. Opportunistic screening leverages data collected from routine health check-ups and multimodal data can play a key role to identify at-risk individuals. Chest X-rays (CXR) provide insights into chronic conditions contributing to major adverse cardiovascular events (MACE), while 12-lead electrocardiogram (ECG) directly assesses cardiac electrical activity and structural abnormalities. Integrating CXR and ECG could offer a more comprehensive risk assessment than conventional models, which rely on clinical scores, computed tomography (CT) measurements, or biomarkers, which may be limited by sampling bias and single modality constraints. We propose a novel predictive modeling framework - MOSCARD, multimodal causal reasoning with co-attention to align two distinct modalities and simultaneously mitigate bias and confounders in opportunistic risk estimation. Primary technical contributions are - (i) multimodal alignment of CXR with ECG guidance; (ii) integration of causal reasoning; (iii) dual back-propagation graph for de-confounding. Evaluated on internal, shift data from emergency department (ED) and external MIMIC datasets, our model outperformed single modality and state-of-the-art foundational models - AUC: 0.75, 0.83, 0.71 respectively. Proposed cost-effective opportunistic screening enables early intervention, improving patient outcomes and reducing disparities.

Figures

Figures reproduced from arXiv: 2506.19174 by the authors.

Figure 1
Figure 1. Proposed MOSCARD architecture and de-confounding causal reasoning graph, input X, task label Y , causal factor A, confounder C, directed edges for causal con￾founder relations: (a) Step 1 - single modality encoder training with confusion loss; (b) Step 2, multimodal learning with co-attention and SCM; (c) Step 1 training – Single modality; (d) Step 2 Multimodal training with co-attention and causal intervention. MOS… view at source ↗
Figure 2
Figure 2. Spoke plot shows the comparative assessment of MOSCARD (green) against the multimodal co-alignment (blue) and with encoder after de-confounding (orange). Qualitative analysis with saliency maps - (a) Older male patient, true MACE event but single modality failed to predict (prob 0.38), MOSCARD correctly predicted (prob 0.78); (b) Young male patient, no MACE event, single and confounder model predicted false positive… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 11 canonical work pages

  1. [1]

    Nature Medicine28(9), 1773–1784 (2022)

    Acosta, J.N., Falcone, G.J., Rajpurkar, P., Topol, E.J.: Multimodal biomedical ai. Nature Medicine28(9), 1773–1784 (2022)

    work page 2022

  2. [2]

    short- cuts

    Banerjee, I., Bhattacharjee, K., Burns, J.L., Trivedi, H., Purkayastha, S., Seyyed-Kalantari, L., Patel, B.N., Shiradkar, R., Gichoya, J.: “short- cuts” causing bias in radiology artificial intelligence: Causes, evaluation, and mitigation. Journal of the American College of Radiology20(9), 842–851 (2023). https://doi.org/https://doi.org/10.1016/j.jacr.202...

    doi:10.1016/j.jacr.2023.06.025 2023

  3. [3]

    Behar, J.A., Levy, J., Celi, L.A.: Generalization in medical ai: a perspective on developing scalable models (2023), https://arxiv.org/abs/2311.05418

    arXiv 2023

  4. [4]

    British Journal of General Practice55(520), 838–845 (2005), https://bjgp.org/content/55/520/838

    Brindle, P.M., McConnachie, A., Upton, M.N., Hart, C.L., Smith, G.D., Watt, G.C.: The accuracy of the framingham risk-score in different socioeconomic groups: a prospective study. British Journal of General Practice55(520), 838–845 (2005), https://bjgp.org/content/55/520/838

    work page 2005

  5. [5]

    Nature communications14(1), 4314 (2023)

    Brown, A., Tomasev, N., Freyberg, J., Liu, Y., Karthikesalingam, A., Schrouff, J.: Detecting shortcut learning for fair medical ai using shortcut testing. Nature communications14(1), 4314 (2023)

    2023

  6. [6]

    Pi et al

    Budoff, M.J., Young, R., Burke, G., Carr, J., Detrano, R.C., Folsom, A.R., Kro- nmal, R., Lima, J.A.C., Liu, K.J., McClelland, R.L., Michos, E.D., Post, W.S., 10 J. Pi et al. Shea, S., Watson, K.E., Wong, N.D.: Ten-year association of coronary artery calcium with atherosclerotic cardiovascular disease (ascvd) events: the multi- ethnic study of atheroscler...

    work page doi:10.1093/eurheartj/ehy217 2018

  7. [7]

    In: 2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV)

    Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for sur- vival prediction in gigapixel whole slide images. In: 2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV). pp. 3995–4005 (2021). https://doi.org/10.1109/ICCV48922.2021.00398

    arXiv 2021

  8. [8]

    Journal of Biomedical Informatics149, 104548 (2024)

    Correa, R., Pahwa, K., Patel, B., Vachon, C.M., Gichoya, J.W., Baner- jee, I.: Efficient adversarial debiasing with concept activation vector — medical image case-studies. Journal of Biomedical Informatics149, 104548 (2024). https://doi.org/https://doi.org/10.1016/j.jbi.2023.104548, https://www.sciencedirect.com/science/article/pii/S1532046423002691

    work page arXiv 2024

Show all 21 references
  1. [9]

    Journal of Medical and Biological Engineering42(6), 816–827 (2022)

    Correa,R.,Shaan,M.,Trivedi,H.,Patel,B.,Celi,L.A.G.,Gichoya,J.W.,Banerjee, I.: A systematic review of ‘fair’ai model development for image classification and prediction. Journal of Medical and Biological Engineering42(6), 816–827 (2022)

    2022

  2. [10]

    BMC medicine17, 1–16 (2019)

    Damen, J.A., Pajouheshnia, R., Heus, P., Moons, K.G., Reitsma, J.B., Scholten, R.J., Hooft, L., Debray, T.P.: Performance of the framingham risk models and pooled cohort equations for predicting 10-year risk of cardiovascular disease: a systematic review and meta-analysis. BMC...

    2019

  3. [11]

    Lancet (London, England) 392(10159), 1736–1788 (2018)

    of Death Collaborators, G..C.: Global, regional, and national age-sex-specific mor- tality for 282 causes of death in 195 countries and territories, 1980-2017: a system- atic analysis for the global burden of disease study 2017. Lancet (London, England) 392(10159), 1736–1788 (...

    2018 doi

  4. [12]

    https://doi.org/https://doi.org/10.1016/j.neuroimage.2020.117689, https://www.sciencedirect.com/science/article/pii/S1053811920311745

    Dinsdale, N.K., Jenkinson, M., Namburete, A.I.: Deep learning-based unlearning of datasetbiasformriharmonisationandconfoundremoval.NeuroImage228,117689 (2021). https://doi.org/https://doi.org/10.1016/j.neuroimage.2020.117689, https://www.sciencedirect.com/science/article/pii/S...

    2021

  5. [13]

    PhysioNet

    Johnson, A., Bulgarelli, L., Pollard, T., Horng, S., Celi, L.A., Mark, R.: Mimic-iv. PhysioNet. Available online at: https://physionet. org/content/mimiciv/1.0/(accessed August 23, 2021) pp. 49–55 (2020)

    2020

  6. [14]

    arXiv preprint arXiv:2309.14381 (2023)

    Lee, N., Bang, Y., Lovenia, H., Cahyawijaya, S., Dai, W., Fung, P.: Survey of social bias in vision-language models. arXiv preprint arXiv:2309.14381 (2023)

    2023 arXiv

  7. [15]

    Advances in neural information processing systems34, 9694–9705 (2021)

    Li,J.,Selvaraju,R.,Gotmare,A.,Joty,S.,Xiong,C.,Hoi,S.C.H.:Alignbeforefuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems34, 9694–9705 (2021)

    2021

  8. [16]

    The American Journal of Cardiology94(1), 20–24 (2004)

    Lloyd-Jones, D.M., Wilson, P.W., Larson, M.G., Beiser, A., Leip, E.P., D’Agostino, R.B., Levy, D.: Framingham risk score and prediction of life- time risk for coronary heart disease. The American Journal of Cardiology94(1), 20–24 (2004). https://doi.org/https://doi.org/10.1016...

    2004 doi

  9. [17]

    Cambridge university press (2009)

    Pearl, J.: Causality. Cambridge university press (2009)

    2009

  10. [18]

    Canadian Journal of Cardiology31(5), 613–619 (2015)

    Preiss, D., Kristensen, S.L.: The new pooled cohort equations risk calculator. Canadian Journal of Cardiology31(5), 613–619 (2015). https://doi.org/https://doi.org/10.1016/j.cjca.2015.02.001, https://www.sciencedirect.com/science/article/pii/S0828282X15000860

    2015 doi

  11. [19]

    Int J Cardiol 199, 460–466 (2015)

    Suh, Y.J., Hong, Y.J., Lee, H.J., Hur, J., Kim, Y.J., Lee, H.S., Hong, S.R., Im, D.J., Kim, Y.J., Park, C.H., Kim, T.H., Yoo, K.J., Choi, B.W.: Prognostic value of syntax score based on coronary computed tomography angiography. Int J Cardiol 199, 460–466 (2015). https://doi.or...

    2015 doi

  12. [20]

    arXiv preprint arXiv:2210.10163 (2022)

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. arXiv preprint arXiv:2210.10163 (2022)

    2022 arXiv

  13. [21]

    J Am Coll Cardiol73(3), 291–301 (2019)

    Williams, M.C., Moss, A.J., Dweck, M., Adamson, P.D., Alam, S., Hunter, A., Shah, A.S.V., Pawade, T., Weir-McCall, J.R., Roditi, G., van Beek, E.J.R., Newby, D.E., Nicol, E.D.: Coronary artery plaque characteristics associated with adverse outcomes in the scot-heart study. J A...

    2019 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.