{"id":"6e3ce065-b564-43a5-a9e1-97f522d85c24","arxiv_id":"2501.04504","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A biventricular statistical shape model from 271 healthy CT scans shows heart size and elongation are the main shape variations and correlate with ECG intervals and amplitudes, and provides 100 synthetic heart meshes for electrophysiology simulations.","lead":"This paper builds a statistical shape model of the two heart ventricles from 271 healthy CT scans, using a new coordinate-based method to align shapes without expensive image registration. It then links the main shape differences, mainly heart size and elongation, to ECG and demographic measurements, and releases 100 synthetic heart meshes for simulation studies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"UVC-based correspondence is the load-bearing assumption: it is only indirectly validated, and the largest reconstruction errors sit at the RV apex and base where UVC alignment is least reliable, so PC scores and their ECG correlations may encode artifacts.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: UVC-based correspondence is the unvalidated foundation of the SSM, and all downstream PC-score correlations inherit its validity. I agree with the reader's conditional verdict rather than escalating to rejection because there is substantial supporting evidence: the published SSM and 100 synthetic volumetric meshes are concrete reproducible artifacts, the LOOCV error is competitive with prior work, the main PC interpretations qualitatively match established SSMs, and the anatomy-ECG correlations are explicitly framed as exploratory with Holm-Bonferroni correction. However, the absence of a registration-based validation is not merely a missing nice-to-have. The correspondence method is the central methodological novelty, and the paper's own error maps concentrate error at the RV apex and the ventricular base, precisely where UVC coordinate definitions are most sensitive to apex placement, valve-ring geometry, and the manual/semi-automatic generation pipeline. Because the reconstruction error is measured on the UVC-sampled point clouds themselves, it cannot detect systematic correspondence bias: a perfectly self-consistent but anatomically wrong correspondence would yield low errors. The proposed reference-SSM comparison directly tests whether the PC space and the Table 3 correlations are stable under an independent correspondence assumption. If they are, the central claim stands and the conditional acceptance is appropriate; if they are not, the paper's anatomical and clinical conclusions would need to be re-derived. This is a concrete, feasible check on the same data and does not require new imaging data.","tokens_in":26220,"tokens_out":4220,"duration_ms":47659,"concrete_test":"Build a reference SSM on the same 271 geometries using a standard registration-based correspondence method (e.g., nonrigid B-spline or symmetric diffeomorphic registration of the 0.9 mm meshes, or the pipeline of Bai et al. 2015). Compare the first three PC modes (cosine similarity after sign alignment) and the per-subject PC scores (Pearson correlation) between the UVC-SSM and the registration-SSM. If the score correlations for PC1-PC3 are below roughly 0.9, or if the significant Table 3 correlations do not reproduce in the registration-based SSM, then UVC correspondence is introducing artifacts that compromise the anatomy-ECG claim. A complementary check is to recompute LOOCV reconstruction error as point-to-surface distance to the original meshes and report errors separately for the RV apex and basal regions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that fixed UVC tuples (z_i, rho, phi_j, nu) correspond to the same anatomical location across subjects, so PCA modes and PC scores reflect true shape differences rather than coordinate-construction artifacts. Section 2.2.1 builds correspondence by extracting points at fixed UVC values and triangulating them directly, with no direct comparison to a registration-based SSM. The evidence offered is indirect (qualitative agreement with prior SSMs, Section 4.2), and the LOOCV reconstruction error of 0.61 mm is computed between corresponding UVC-sampled point clouds, not as point-to-surface distance to the original 0.9 mm anatomical meshes. The pointwise error map (Figure 8) shows the largest errors at the RV apex (8.83 mm for a random subject) and at the ventricular base/LV-RV junction, exactly the regions where UVC alignment is hardest: the apex has singular coordinate geometry and the base depends on valve-ring placement. Additionally, Laplacian smoothing is applied only in selected LV basal regions, based on visual inspection, which can itself shift corresponding points. If UVC correspondence is biased in these regions, PC2 ('elongation') and PC3 ('relative LV/RV position') contain non-anatomical variance, and the reported ECG/demographic correlations (Table 3) become correlations with a mixture of anatomy and UVC artifacts. This is not an accusation that the method is wrong; it is an unvalidated assumption that needs a direct check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a biventricular statistical shape model (SSM) built from 271 healthy cardiac CT scans, using universal ventricular coordinates (UVCs) to sample corresponding point clouds, followed by generalized Procrustes alignment and PCA. The authors report that the first three PCs encode ventricular size, elongation, and relative LV/RV positioning, jointly explaining 61.12% of the variance; a leave-one-out cross-validation reconstruction error of 0.61 mm; a synthetic cohort of 100 volumetric meshes with fiber orientations and anatomical labels; and statistically significant correlations between PC scores and ECG-derived and demographic features, notably PC1 with RR interval (R = -0.50) and height (R = -0.62). The paper also studies dataset-size effects on reconstruction error and sex-related representativeness of the SSM, and it releases the SSM, code, and synthetic meshes.","tokens_in":26485,"tokens_out":4550,"duration_ms":46838,"significance":"If the UVC-based correspondence is valid, this is a valuable contribution: the dataset is larger and at higher spatial resolution than most prior cardiac SSM studies, the anatomical twinning pipeline is highly automated, the LOOCV evaluation is carefully executed, multiple-testing correction is applied, and the public release of ready-to-simulate meshes with fibers and labels directly supports downstream electrophysiological simulation studies. The reported correlations between anatomy and clinical ECG features are also of practical interest for anchoring digital-twin validation. However, the central methodological premise—that fixed UVC tuples correspond to the same anatomical location across subjects—is only indirectly validated, and the correlation analysis is performed in-sample without adjustment for body size or sex. These issues affect the strength of the paper's main claims and require further analysis before the anatomy-to-ECG link can be considered established.","major_comments":[{"comment":"The reconstruction error reported as 0.61 mm is defined as the mean distance between corresponding points in the original and reconstructed UVC-resampled point clouds, not as a point-to-surface distance to the original 0.9 mm anatomical meshes. Since both point clouds are generated by the same UVC sampling and triangulation, this metric measures how well the SSM reproduces UVC-sampled coordinates, not how faithfully it reconstructs the underlying anatomy. The validation in Appendix A against the original meshes concerns sampling error for choosing the mesh size, not the SSM reconstruction error. The pointwise error map in Figure 8 shows the largest errors at the RV apex and the ventricular base/LV-RV junction, which are precisely the regions where UVC alignment is hardest (singular coordinate geometry at the apex, valve-ring dependence at the base). Because PC2 and PC3 encode elongation and relative LV/RV position, these modes—and hence the correlations in Table 3—may contain variance that is an artifact of the correspondence construction rather than true anatomical variation. I ask the authors to validate the correspondence more directly, for example by computing point-to-surface reconstruction errors of the reconstructed meshes against the original anatomical surfaces, or by comparing PC modes with a registration-based SSM on a subset of the data.","section":"Sections 2.2.1, 2.2.3, and 3.1 (reconstruction error and UVC correspondence)"},{"comment":"The correlation analysis uses PC scores estimated from the same 271 subjects that define the PCA, and the reported p-values do not account for the uncertainty in the estimated PC loadings. More importantly, no adjustment is made for sex, height, or body surface area. Since PC1 is described as overall ventricular size, the strong correlations of PC1 with height (R = -0.62) and BSA (R = -0.55) are partly expected from the construction of the size mode, and the PC1-ECG correlations such as RRint (R = -0.50) could be driven by body size and sex differences rather than by ventricular shape per se. To support the claimed anatomy-ECG link in Section 4.7, the authors should report partial correlations controlling for sex, height, and BSA, or validate the correlations in an independent cohort. Without this, the directional claims such as 'larger ventricles are associated with longer RR and QT intervals' remain vulnerable to confounding.","section":"Section 2.4.2 and Table 3 (in-sample correlation analysis)"},{"comment":"The manuscript states in Section 2.1.3 that RV myocardial thickness variations are not captured but fixed, and Section 4.3 acknowledges this limitation. This means the SSM is not a complete biventricular shape model, and the synthetic cohort inherits the same restriction; correlations with ECG features cannot be interpreted as reflecting full biventricular anatomical variability. In addition, Section 2.2.1 applies Laplacian smoothing to LV basal regions selected by visual inspection. Because the same smoothing is applied to all meshes to preserve correspondence, it can systematically shift corresponding points in exactly the regions where reconstruction errors are largest, potentially affecting PC loadings and downstream correlations. I ask the authors to quantify the sensitivity of the PC scores and of the main correlations to the smoothing procedure, or to justify the smoothing region choice with a quantitative criterion rather than visual inspection alone.","section":"Sections 2.1.3, 2.2.1, and 4.3 (fixed RV wall thickness and Laplacian smoothing)"}],"minor_comments":[{"comment":"The electrical axis is listed as an ECG-derived feature in Section 2.4.2 but is not included in Table 2; please add it to the table or clarify its source and measurement lead configuration.","section":"Table 2 and Section 2.4.1"},{"comment":"The p-values in Table 3 are reported in units of 10^-3, but the caption does not state this, and the text refers to p = 0.00; please clarify the scaling and avoid presenting rounded values as exactly zero.","section":"Table 3"},{"comment":"Synthetic geometries are sampled from 94 PCs under the assumption that PC scores are normally distributed, but the normality verification in Appendix E is reported only for the first 16 PCs; please either verify the remaining PCs or explicitly state the assumption for PCs 17-94.","section":"Section 2.3.1 and Appendix E"},{"comment":"The left panel is described as a 'random subject'; please clarify whether the same subject is used for the pointwise error map and for the average error map, or specify how the representative subject was chosen.","section":"Figure 8"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the released resources are potentially useful to the cardiac modeling community. The main risk is that the UVC correspondence is not validated against an independent anatomical standard, and the correlation analysis is in-sample and unadjusted for strong confounders. These are fixable within the manuscript's scope by adding point-to-surface reconstruction errors, a registration-based comparison or sensitivity analysis, and partial-correlation analyses. I would not reject at this stage, but the revisions are substantive rather than purely editorial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a useful paper: a 271-subject healthy biventricular SSM from high-resolution CT, a new lightweight correspondence method based on universal ventricular coordinates (UVCs), and a public synthetic cohort with fibers and anatomical labels. Second, the load-bearing assumption—that fixed UVC tuples correspond to the same anatomical location across subjects—is not directly tested against a registration-based SSM, and the largest reconstruction errors sit exactly where UVC alignment is hardest (RV apex, base). That is a real soft spot, but not a fatal one, and the authors are honest about many of their limitations.\n\nWhat it does well. The dataset is large and well-characterized, the pipeline is fully automated, and the evaluation is more thorough than most SSM papers: LOOCV across demographic subgroups, balanced datasets, sex-biased versus unbiased SSMs, and a size-independent analysis. The 0.61 mm LOOCV error is competitive, and the first three PCs (size, elongation, relative LV/RV position) match prior work, which is indirect evidence that the correspondence is not wildly off. They release the model, 100 volumetric meshes with fiber architecture, and code—that is a real, citable resource. The anatomy-ECG correlations survive Holm-Bonferroni correction, are framed as exploratory, and align with simulation-based findings (e.g., heart size vs. R-wave amplitude and QT interval). Good citation practice throughout.\n\nSoft spots, in order of severity. (1) The UVC correspondence validation is indirect. The authors point to qualitative agreement with earlier SSMs, but they never compare their PC modes or reconstruction errors against a registration-based SSM on the same data. The pointwise error map (Figure 8) shows 8.8 mm at the RV apex and elevated errors at the base. A direct comparison—even on a subset—would settle whether PC2 and PC3 carry true shape variance or coordinate artifacts. (2) The female and male cohorts were recruited under different criteria: males from Master@Heart with strict exclusions, females from clinical care with looser exclusions. The paper acknowledges this in the limitations, but it weakens the sex and size comparisons. (3) The correlation analysis is in-sample and does not adjust for sex, height, or BSA. PC1 correlates with height at R = -0.62; the ECG correlations might partly reflect body-size effects. The authors call it exploratory, so this is a minor issue. (4) Minor: normality was verified for only 16 of the 94 PCs used to sample the synthetic cohort, and only visually.\n\nWho it is for: anyone building cardiac digital twins or population-level anatomical models, and simulation groups wanting ready-to-use healthy ventricular meshes. It deserves a serious referee. The UVC validation gap should be addressed (direct comparison to registration, or at least a sensitivity analysis), and the cohort-recruitment mismatch should be discussed more prominently. Neither is a deal-breaker.","headline":"Solid biventricular SSM with a genuinely new UVC-based correspondence method and useful released resources; the main soft spot is that the correspondence is only indirectly validated, but the paper is worth refereeing.","tokens_in":27087,"tokens_out":2440,"would_cite":true,"duration_ms":24063,"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 biventricular statistical shape model of 271 healthy hearts, built with universal-coordinate correspondence, identifies ventricular size, elongation, and chamber position as the leading anatomical modes and links them to real ECG and…","keywords":["statistical shape model","universal ventricular coordinates","biventricular anatomy","cardiac digital twin","electrocardiogram","principal component analysis","synthetic cohort","healthy population"],"falsifier":"Rebuild the SSM from the same 271 CT scans using a conventional nonlinear registration method for correspondence instead of UVC sampling, then compare the top three PCs and the PC1-RR-interval correlation: if the leading modes or the sign and significance of the reported correlations shift materially, the UVC correspondence assumption is the cause.","tokens_in":25982,"feed_emoji":"🫀","tokens_out":5638,"duration_ms":56041,"temperature":0.7,"pith_summary":"This paper sets out to show that a statistical shape model of the healthy biventricular heart can be built cheaply and reliably by using universal ventricular coordinates to place corresponding points on every heart, bypassing the costly nonlinear registration step that earlier shape models rely on. Trained on 271 high-resolution CT scans of healthy individuals, the model finds that ventricular size (40.34% of variance), elongation (11.19%), and the relative apical-basal position of the left and right ventricles (9.59%) are the three leading modes, jointly explaining 61.12% of shape variability. The authors then connect these modes to clinical measurements: after multiple-testing correction, larger hearts correlate with shorter RR intervals, larger R-wave amplitudes, taller height, and larger body surface area, among other significant associations. If these correlations are correct, they give digital-twin builders a real-world, population-level link between anatomy and electrocardiography that can be used to calibrate and validate electrophysiological simulations, and they make a synthetic cohort of 100 ready-to-simulate volumetric meshes with fiber orientations available.","feed_headline":"Healthy heart shape predicts ECG features in 271 CT scans","feed_subtitle":"Ventricular size, elongation and chamber position drive anatomy, tying it to RR interval, height and R-wave amplitude.","key_machinery":"The central object is the universal ventricular coordinate (UVC) system, which assigns every ventricular location four coordinates: an apico-basal coordinate $z \\in [0,1]$, a transmural coordinate $\\rho \\in [0,1]$ from endocardium to epicardium, a rotational coordinate $\\phi$, and a chamber label $\\nu \\in \\{-1,1\\}$ for left or right ventricle. The paper samples a regular grid of $(z, \\rho, \\phi, \\nu)$ tuples, extracts the corresponding point from each subject's volumetric mesh (by interpolation when needed), and triangulates the resulting point clouds directly, producing 31,724 corresponding points per heart. This turns shape correspondence into a simple coordinate lookup, so ordinary PCA on translation- and rotation-aligned point clouds yields modes that describe biological shape differences rather than artifacts of a registration procedure.","core_discovery":"The central claim is that biventricular anatomy in a healthy population varies primarily along three interpretable axes: overall ventricular size, ventricular elongation (most visible in the right ventricle), and the relative apical-basal positioning of the left and right ventricles. These axes are the first three principal components of a statistical shape model trained on 271 healthy CT scans, built by sampling a fixed grid of universal ventricular coordinates on every geometry and applying PCA to the resulting point clouds. The paper further claims that these anatomical modes carry real electrophysiological signal: after Holm-Bonferroni correction, the size mode correlates significantly with ECG-derived features such as the RR interval ($R = -0.50$), QT interval, T-wave duration, and R-wave amplitudes, and with demographic variables including height ($R = -0.62$), weight, body surface area, BMI, and age. A leave-one-out reconstruction error of 0.61 mm indicates that the model represents unseen healthy geometries about as well as the CT resolution allows.","pith_inferences":["If the same UVC-grid approach were extended to the atria, as the authors suggest as future work, the analogous shape-to-P-wave correlations could be tested directly with the released workflow.","Because the size mode is tightly correlated with height and body surface area, a follow-up analysis that regresses out body-size variables before correlating PC1 with ECG features would separate anatomical heart size from systemic body scaling.","One could generate synthetic meshes at extreme PC1 values and run reaction-eikonal simulations to check whether the predicted R-wave and RR-interval changes reproduce the observed correlations in magnitude, not just in sign."],"forward_implications":["A digital-twin pipeline can obtain patient-specific biventricular anatomy from CT without a nonlinear registration step, because UVC grids give point correspondence directly.","The significant anatomy-to-ECG correlations provide real-patient targets for validating cardiac electrophysiology simulations: simulated ECGs from size-varying meshes should reproduce the observed signs and rough magnitudes of association.","The released model and 100 synthetic meshes with fibers and anatomical labels allow large, privacy-preserving virtual cohorts for in silico trials and training data-driven calibration models.","Because size is the dominant mode, accurately representing heart size in a digital twin is the highest-leverage anatomical choice for reproducing clinically measured ECG features."],"supporting_citations":[{"why":"Defines the universal ventricular coordinates that the paper uses as the correspondence mechanism.","marker":"[29]"},{"why":"Provides the digital-twin mesh generation and UVC computation pipeline that the anatomical models are built with.","marker":"[11]"},{"why":"Supplies the large biventricular atlas baseline for comparing variance explained, reconstruction error, and dataset-size trends.","marker":"[16]"},{"why":"Offers the four-chamber SSM and simulated-function comparison that supports the anatomical-mode interpretations and the QT-interval correlation.","marker":"[18]"},{"why":"Establishes the bi-atrial SSM and the identical LOOCV reconstruction-error metric used to benchmark the present model.","marker":"[19]"},{"why":"Provides the prospective athlete-and-control cohort data used for the male subjects in the study.","marker":"[32]"},{"why":"Supplies simulation-based evidence that increasing heart size raises R-wave amplitude, against which the observed ECG correlations are compared.","marker":"[66]"},{"why":"Reports similar correlations between ventricular size and heart rate and height, supporting the demographic and ECG associations found here.","marker":"[20]"}],"fun_headline_variants":["Size, elongation, and chamber position drive heart's ECG links","Three heart shape modes predict ECG features from 271 CT scans","Ventricular shape axes reveal anatomy's role in ECG: CT study","Biventricular shape model links 3 anatomical modes to ECG in 271 healthy","Heart anatomy's top 3 shape components predict ECG signals in healthy CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that sampling the same grid of universal ventricular coordinates on every heart yields the same anatomical points on each heart; if that correspondence is distorted, the principal components and all reported correlations with ECGs and demographics would be misleading.","fun_headline_variants_meta":{"raw":{"variants":["Size, elongation, and chamber position drive heart's ECG links","Three heart shape modes predict ECG features from 271 CT scans","Ventricular shape axes reveal anatomy's role in ECG: CT study","Biventricular shape model links 3 anatomical modes to ECG in 271 healthy","Heart anatomy's top 3 shape components predict ECG signals in healthy CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000599,"raw_usage":{"total_tokens":2824,"prompt_tokens":993,"completion_tokens":1831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1737}},"tokens_in":609,"tokens_out":1831,"duration_ms":14322,"temperature":1.0,"reasoning_tokens":1737,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:30:57.669492+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rebuild the SSM from the same 271 CT scans using a conventional nonlinear registration method for correspondence instead of UVC sampling, then compare the top three PCs and the PC1-RR-interval correlation: if the leading modes or the sign and significance of the reported correlations shift materially, the UVC correspondence assumption is the cause.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the large biventricular atlas baseline for comparing variance explained, reconstruction error, and dataset-size trends."},{"cited_title":"Van De Heyning, Paul Van Herck, Bernard Paelinck, Haroun El Addouli, André La Gerche, Lieven Herbots, Hein Heidbuchel, Rik Willems, and Guido Claessen","cited_arxiv_id":null,"evidence_quote":"Provides the prospective athlete-and-control cohort data used for the male subjects in the study."},{"cited_title":"Young, Nay Aung, Luis R","cited_arxiv_id":null,"evidence_quote":"Reports similar correlations between ventricular size and heart rate and height, supporting the demographic and ECG associations found here."}],"review_version":1}