Pith. sign in

REVIEW 3 major objections 6 minor 23 references

Guiding Treatment Strategies: The Role of Adjuvant Anti-Her2 Neu Therapy and Skin/Nipple Involvement in Local Recurrence-Free Survival in Breast Cancer Patients

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Applying LiNGAM causal discovery to a 922-patient breast cancer dataset, this paper estimates that adjuvant anti-Her2 therapy lengthens the last local recurrence-free assessment by 169 days on average, while skin/nipple involvement…

desk verdict A clear-eyed application of LiNGAM to a breast cancer dataset that collapses on a load-bearing flaw: the outcome is a censored follow-up time, not an uncensored linear target, and the backdoor analysis is run on a graph the authors themselves admit contains latent confounders. read the letter →

arxiv 2501.04724 v1 pith:D2EFUU3A submitted 2025-01-04 stat.AP cs.LG

Joe Omatoi , Abdul M Mohammed , Dennis Trujillo This is my paper · ORCID
classification stat.APcs.LG
keywords breastcancercausalinferenceLiNGAMDirectlocalrecurrence-freesurvivaladjuvantanti-Her2therapyskin-nippleinvolvementobservationaldata
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

This paper argues that causal effects can be extracted from routine observational breast-cancer records, not just from randomized controlled trials. Using the Linear Non-Gaussian Acyclic Model (LiNGAM) on more than 40 features from 922 patients, it reports that adjuvant anti-Her2 therapy adds 169 days on average to the time until the last local recurrence-free assessment, while skin/nipple involvement removes 351 days. The authors present these as direct causal effects, and they claim the estimates survive refutation tests that add random common causes, use data subsets, and replace the treatment with a placebo. If the paper is right, clinicians gain quantitative, individualized guidance for which therapies matter and which anatomical presentations signal higher recurrence risk.

What carries the argument

The key machinery is DirectLiNGAM, a causal-discovery algorithm for linear non-Gaussian acyclic structural equation models. Each variable $x_i$ is written as a linear combination of earlier variables $x_j$ plus an independent non-Gaussian error $e_i$; the algorithm finds the causal order by exploiting the fact that regressing an effect on its cause leaves a residual independent of the cause, while regressing in the reverse direction leaves a dependent residual unless errors are Gaussian. Once the ordering is fixed, adaptive Lasso removes weak connections. The paper also uses the LiNGAM RCD extension to allow hidden common causes, and checks the resulting causal effects with the backdoor criterion and three refutation techniques.

What would settle it

Run a time-to-event analysis on the same 922-patient cohort, treating recurrence or death as the event and the last recurrence-free assessment as censoring, and compare Cox or competing-risk estimates for adjuvant anti-Her2 therapy and skin/nipple involvement; if the adjusted hazard ratio for the therapy is not clearly below 1, the claimed 169-day benefit is an artifact of treating censoring time as an outcome. A randomized comparison of Her2-positive patients with and without the therapy that fails to reproduce a roughly 169-day difference would also falsify the central claim.

Watch

Extended reading notes

Core claim

The central claim is that the causal graph learned from the data contains a direct edge from Adjuvant Anti-Her2 Neu Therapy to the outcome 'Days to last local recurrence free assessment (from the date of diagnosis),' with an effect of +169 days, and a direct edge from Skin/Nipple Involvement with an effect of −351 days. The authors state these estimates in terms of the encoded variables: no skin/nipple involvement increases the interval by 351.68 days, and not receiving adjuvant anti-Her2 therapy decreases it by 169.64 days. They interpret the magnitude as clinically meaningful for Her2-positive patients and for patients with skin or nipple involvement, and they describe the refutation results as showing robustness to confounders and sampling variation.

Load-bearing premise

The load-bearing premise is that 'Days to last local recurrence free assessment (from the date of diagnosis)' is an uncensored continuous outcome that enters a linear non-Gaussian causal model directly; if the real endpoint is censored time-to-recurrence or responds nonlinearly, the 169-day and 351-day causal estimates are biased and the paper's causal claims collapse.

Editorial extensions

If this is right

  • For Her2-positive patients, adjuvant anti-Her2 therapy should be a priority in treatment planning, since an average gain of 169 days to the last recurrence-free assessment is a concrete benefit.
  • Skin/nipple involvement is a prognostic marker for early recurrence: the 351-day average reduction points to patients who need intensified treatment or closer follow-up.
  • Causal discovery on observational data can complement RCTs, widening the evidence base for treatment decisions beyond narrowly selected trial populations.
  • The methodology, if correct, transfers to other observational oncology datasets and can estimate causal effects where trials are impractical or too expensive.

Reading between the lines

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

  • The paper leaves implicit that its outcome is the last assessment date rather than a confirmed recurrence time; a survival-analysis companion would either confirm the 169-day estimate or reveal it as a censoring artifact.
  • The same causal-discovery pipeline could be applied to the twenty other breast cancer datasets cited in the paper to test whether these effect sizes replicate across institutions and imaging protocols.
  • Because the model estimates average effects, a natural next step is estimating conditional effects by receptor status, stage, and age, which could tell clinicians which Her2-positive patients gain the full 169 days and which high-risk patients need more than standard monitoring.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript applies the LiNGAM family of causal discovery methods (DirectLiNGAM and LiNGAM RCD) to the Duke MRI Breast Cancer dataset to estimate causal effects of treatment and clinical features on the time to the last recurrence-free assessment. The authors report that Adjuvant Anti-Her2 Neu Therapy increases this time by 169 days and that Skin/Nipple involvement decreases it by 351 days. They attempt to validate these effects using the backdoor criterion and three refutation techniques implemented in DoWhy. The central methodological choices are to treat the outcome as an uncensored continuous variable in a linear non-Gaussian structural equation model, and to combine a graph learned with methods that allow hidden confounders with a backdoor adjustment procedure designed for DAGs.

Significance. If the results were valid, they would provide actionable causal estimates from observational data and could illustrate a useful workflow for other clinical datasets. The paper has several strengths: it uses a publicly available dataset, explicitly documents preprocessing steps, applies established causal discovery algorithms, and includes a suite of refutation checks. However, the central effect estimates are not credible because the outcome is a censored follow-up time that is modeled as an uncensored continuous variable, and because the backdoor analysis is applied to a graph that the authors themselves state contains bidirected edges representing latent common causes. These problems directly undermine the headline claims, so the paper does not currently contribute a trustworthy causal finding.

major comments (3)
  1. [Section 2.3, Table 1] The outcome variable 'Days to last local recurrence free assessment (from the date of diagnosis)' is a censored time-to-event variable. For patients who experience a local recurrence, this is the time of the last recurrence-free assessment before the event; for patients who do not, it is the time of the last follow-up assessment (administrative censoring). The linear SEM in Section 2.3 treats this variable as an uncensored continuous outcome with no censoring indicator and no survival model. Censoring induces a selection mechanism: longer follow-up is associated with treatment choices and with event-free status, so the regression residuals are not independent of treatment even under the null. The LiNGAM coefficients reported in Table 1 (-169.638 and 351.6813) are therefore biased and cannot be interpreted as causal effects on local recurrence-free survival. The paper provides no justification for ignoring censoring and no sensitivity analysis that accounts for it.
  2. [Sections 2.3 and 2.4] The authors state that the absence-of-hidden-common-causes assumption is violated in many pairs and therefore switch from DirectLiNGAM to LiNGAM RCD, whose output contains bidirected edges representing latent common causes. In Section 2.4, they nonetheless apply the backdoor criterion and estimate effects with DoWhy 'with the causal structure from LiNGAM' as though the graph were a DAG. The backdoor criterion is defined only for directed acyclic graphs without bidirected edges; adjustment sets chosen from a graph with bidirected edges do not block paths through unobserved common causes. The reported backdoor estimates and the refutation results therefore do not support the claim that hidden confounding has been controlled, and the causal claims in Section 3 rest on an identification procedure that is invalid under the authors' own stated model.
  3. [Sections 2.2 and 3] Feature selection is performed with Lasso on the same outcome variable that is later used in the causal model, and no correction is made for the selection step. The selected features and the causal coefficients are estimated from the same data, so the reported effects are in-sample quantities with no clear uncertainty quantification. The 'Use a subset of data' refutation does not address this problem, because it refits the same selected model on a subset rather than repeating feature selection. As a result, the reported effect sizes may be substantially overfit and do not provide evidence of out-of-sample or replicable causal effects.
minor comments (6)
  1. [Section 2.4 and references] The text cites 'Javidian and Valtorta (2018)', 'Hitchcock and Pearl (2001)', and 'Sharma and Kiciman (2020)', but none of these works appear in the reference list; full citations should be added.
  2. [Section 2.3] The statement 'We already know from Lasso regression that the outcome variable is a linear combination of the extracted features' does not justify linearity, because Lasso assumes linearity rather than testing it; the correlation-matrix evidence is also weak. Linearity should be stated as an assumption or checked with a more direct diagnostic.
  3. [Table 1 and Section 3] The variable names include a 'No' suffix (e.g., 'Adjuvant Anti-Her2 Neu TherapyNo'), which makes the sign interpretation potentially confusing. The text says the therapy increased days by 169, while the table reports -169.638 for the 'No' level; please clarify the coding so that readers can verify the direction of each effect.
  4. [Section 2.2] The description of the feature-selection threshold is vague ('importance seems to be similar to Random 5'); the exact rule used to exclude features should be specified.
  5. [Section 2.1] Mean/mode imputation for missing clinical variables is known to be problematic for linear models; even if this is not a primary concern, the paper should discuss the potential bias introduced by single imputation.
  6. [Abstract and Section 3] The text uses 'local recurrence-free survival' for what is actually 'days to last local recurrence-free assessment'; this terminology overstates the clinical endpoint and should be corrected or justified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported causal effects are fitted LiNGAM coefficients, not predictions derived from their own inputs; the paper's validity problems are not circularity.

full rationale

The manuscript's derivation chain is: preprocess the Duke data; select features with Lasso regressions against the outcome; run DirectLiNGAM/LiNGAM RCD on the selected features and the outcome; read causal effects from the fitted structural equation model; and check stability with DoWhy backdoor adjustment and refutation tests. None of these steps defines its output in terms of its own output. The reported 169-day and 351-day effects are the estimated linear coefficients of the fitted SEM (Table 1), i.e., the fitted parameters themselves. The paper does not fit a parameter on one subset and then 'predict' a closely related quantity on another; it reports the fitted model. The stated assumptions (acyclicity, linearity, non-Gaussianity, absence of hidden common causes) are premises of the estimator, not conclusions derived from the data. When the paper finds the hidden-common-cause assumption violated, it switches to LiNGAM RCD, which relaxes that assumption; the later use of the backdoor criterion and DoWhy is an internal validity inconsistency, not a circular reduction. There are no author self-citations used as load-bearing evidence; all method citations are to external authors (Shimizu et al., Maeda & Shimizu, Sharma & Kiciman). The linearity justification ('We already know from Lasso regression that the outcome variable is a linear combination of the extracted features') is an empirical motivation, not a circular derivation. Consequently, no circular step meets the quoted-evidence bar. The skeptic's concerns about censoring, outcome definition, and latent confounding are substantive threats to validity, but they are not circularity because they do not show that a claimed result is equivalent by construction to its inputs.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claims depend on a set of strong modeling assumptions, including linearity, non-Gaussianity, temporal ordering, and the treatment of a censored survival outcome as an uncensored linear target. The effect sizes are fitted coefficients from the same data used for feature selection, with no external validation.

free parameters (6)
  • Lasso regularization alpha = not reported (optimal alpha via 5-fold CV, Appendix A)
    Used in feature selection to rank features; the choice of alpha affects which features are retained.
  • Feature selection importance threshold = importance of Random 5 variable
    Features below the importance of the fifth random Gaussian variable were dropped; this cutoff is data-dependent and can bias variable selection.
  • Number of random Gaussian variables = 5
    Ad hoc choice for the Lasso feature selection stability check.
  • Causal effect coefficient (Adjuvant Anti-Her2 Neu TherapyNo to outcome) = -169.638 days
    Primary claimed effect; estimated from the LiNGAM model on the same data used for feature selection.
  • Causal effect coefficient (Skin/Nipple InvolvementNo to outcome) = 351.6813 days
    Primary claimed effect; estimated from the LiNGAM model.
  • Mean/mode imputation values = not specified
    Missing values were imputed with mean or mode; the imputed values affect all downstream estimates.
assumptions (6)
  • domain assumption The true causal structure is acyclic and variables follow a temporal order where demographics precede conditions which precede treatments.
    Invoked in Section 2.3 to justify the acyclicity assumption of LiNGAM; if temporal order is misspecified, the causal directions can be wrong.
  • domain assumption All error terms are non-Gaussian.
    Tested via Shapiro-Wilk and reported as p=0.00 for all features; this is a key assumption for LiNGAM identifiability.
  • domain assumption Relationships between variables are linear.
    Justified by Lasso and the correlation matrix, but many clinical relationships are nonlinear; linearity is a strong assumption in LiNGAM.
  • ad hoc to paper The backdoor criterion can be applied to the LiNGAM RCD output that contains bi-directed edges.
    The paper detected hidden common causes and used LiNGAM RCD, but then applies backdoor adjustment as if the graph were a DAG, which is a conflict.
  • ad hoc to paper The outcome 'Days to last local recurrence free assessment' is an uncensored continuous variable suitable for linear modeling.
    This is implicit; the analysis does not account for censoring or the event indicator, which is a critical flaw.
  • domain assumption Mean/mode imputation preserves the true distribution and does not bias causal estimates.
    Stated in Section 2.1; no missingness rates or sensitivity analysis are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Guiding Treatment Strategies: The Role of Adjuvant Anti-Her2 Neu Therapy and Skin/Nipple Involvement in Local Recurrence-Free Survival in Breast Cancer Patients." pith.science (2026). https://pith.science/paper/D2EFUU3A

@misc{pith2026250104724,
  author       = {Pith},
  title        = {Pith review of: Guiding Treatment Strategies: The Role of Adjuvant Anti-Her2 Neu Therapy and Skin/Nipple Involvement in Local Recurrence-Free Survival in Breast Cancer Patients},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2EFUU3A}},
  note         = {Machine review of arXiv:2501.04724}
}
read the original abstract

This study explores how causal inference models, specifically the Linear Non-Gaussian Acyclic Model (LiNGAM), can extract causal relationships between demographic factors, treatments, conditions, and outcomes from observational patient data, enabling insights beyond correlation. Unlike traditional randomized controlled trials (RCTs), which establish causal relationships within narrowly defined populations, our method leverages broader observational data, improving generalizability. Using over 40 features in the Duke MRI Breast Cancer dataset, we found that Adjuvant Anti-Her2 Neu Therapy increased local recurrence-free survival by 169 days, while Skin/Nipple involvement reduced it by 351 days. These findings highlight the therapy's importance for Her2-positive patients and the need for targeted interventions for high-risk cases, informing personalized treatment strategies.

Figures

Figures reproduced from arXiv: 2501.04724 by the authors.

Figure 1
Figure 1. Feature importance plot showing the absolute coefficients of various clinical and treatment variables in predicting breast cancer [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustration of three possible causal structures under [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Correlation matrix showing the relationships between different clinical variables in a medical study. The values range from -1 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: P-value matrix displaying the statistical significance of relationships between medical variables. The values range from 0 to 1, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Backdoor criterion example With the causal structure from LiNGAM, we estimated the causal effect and applied the refutation techniques for each variable of interest respectively using DoWhy, “an open-source Python library” for causal inference provided by Sharma and Ki…
Figure 6
Figure 6. Figure 6: LiNGAM (Linear Non-Gaussian Acyclic Model) causal graph depicting the structural relationships between breast cancer [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages

  1. [1]

    J. R. Bosdriesz, V . S. Stel, M. Van Diepen, Y . Meuleman, F. W. Dekker, C. Zoccali, and K. J. Jager. Evidence-based medicine—when ob- servational studies are better than randomized con- trolled trials. Nephrology, 25(10):737–743, 2020. doi:10.1111/nep.13742

    work page doi:10.1111/nep.13742 2020

  2. [2]

    F. Bray, M. Laversanne, H. Sung, J. Ferlay, R. L. Siegel, I. Soerjomataram, and A. Jemal. Global can- cer statistics 2022: Globocan estimates of incidence and mortality worldwide for 36 cancers in 185 coun- tries. CA Cancer J Clin , 74(3):229–263, May–June

    work page 2022

  3. [3]

    De La Cruz, A

    L. De La Cruz, A. M. Moody, E. E. Tappy, S. A. Blankenship, and E. M. Hecht. Overall survival, disease-free survival, local recurrence, and nipple– areolar recurrence in the setting of nipple-sparing mas- tectomy: A meta-analysis and systematic review. An- nals of Surgical Oncology , 22(10):3241–3249, 2015. doi:10.1245/s10434-015-4739-1

    work page doi:10.1245/s10434-015-4739-1 2015

  4. [4]

    A. R. Donders, G. J. van der Heijden, T. Stij- nen, and K. G. Moons. Review: a gentle in- troduction to imputation of missing values. J Clin Epidemiol , 59(10):1087–91, October 2006. doi:10.1016/j.jclinepi.2006.01.014. PMID: 16980149. Epub 2006 Jul 11

    doi:10.1016/j.jclinepi.2006.01.014 2006

  5. [5]

    Dong and C

    Y . Dong and C. Y . Peng. Principled missing data methods for researchers. Springerplus, 2(1):222, May 2013. doi:10.1186/2193-1801-2-222. PMID: 23853744. PMCID: PMC3701793

    work page doi:10.1186/2193-1801-2-222 2013

  6. [6]

    A. S. Gershon, P. K. Lindenauer, K. C. Wilson, L. Rose, A. J. Walkey, M. Sadatsafavi, K. J. Anstrom, D. H. Au, B. G. Bender, M. A. Brookhart, R. A. Dweik, M. K. Han, M. J. Joo, V . Lavergne, A. B. Mehta, M. Miravitlles, R. A. Mularski, N. Roche, E. Oren, K. A. Riekert, N. C. Schoenberg, T. A. Stukel, C. H. Weiss, H. Wunsch, J. J. Africk, and J. A. Krishna...

    work page 2021

  7. [7]

    A. Guha, M. G. Fradley, S. F. Dent, N. L. Wein- traub, M. B. Lustberg, A. Alonso, and D. Addi- son. Incidence, risk factors, and mortality of atrial fibrillation in breast cancer: a seer-medicare analy- sis. European Heart Journal , 43(4):300–312, 2021. doi:10.1093/eurheartj/ehab745

    work page doi:10.1093/eurheartj/ehab745 2021

  8. [8]

    An intro- duction of variable and feature selection

    Isabelle Guyon and Andr ´e Elisseeff. An intro- duction of variable and feature selection. J. Ma- chine Learning Research Special , Issue on Vari- able and Feature Selection. 3:1157–1182, 2003. doi:10.1162/153244303322753616

    doi:10.1162/153244303322753616 2003

Show all 23 references
  1. [9]

    Harbeck, F

    N. Harbeck, F. Penault-Llorca, J. Cortes, et al. Breast cancer. Nat Rev Dis Primers , 5:66, 2019. doi:10.1038/s41572-019-0111-2

    2019 doi

  2. [10]

    F. E. Harrell. Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis . Springer-Verlag, New York,

  3. [11]

    M. S. Iqbal, W. Ahmad, R. Alizadehsani, S. Hussain, and R. Rehman. Breast cancer dataset, classifica- tion and detection using deep learning. Healthcare, 10(12):2395, 2022. doi:10.3390/healthcare10122395

    2022 doi

  4. [12]

    Msaouel, J

    P. Msaouel, J. Lee, J. A. Karam, and P. F. Thall. A causal framework for making individualized treatment decisions in oncology. Cancers, 14(16):3923, 2022. doi:10.3390/cancers14163923

    2022 doi

  5. [13]

    Olier, Y

    I. Olier, Y . Zhan, X. Liang, et al. Causal inference and observational data. BMC Med Res Methodol, 23:227,

  6. [14]

    A. Saha, M. R. Harowicz, L. J. Grimm, C. E. Kim, S. V . Ghate, R. Walsh, and M. A. Mazurowski. A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 dce-mri fea- tures. British journal of cancer , 119(4):508–516, 2018

    2018

  7. [15]

    J. L. Schafer and J. W. Graham. Missing data: our view of the state of the art. Psychol Methods , 7(2):147–77, June 2002. PMID: 12090408

    2002

  8. [16]

    Shi and B

    J. Shi and B. Norgeot. Learning causal effects from observational data in healthcare: A review and summary. front med (lausanne). 9:864882, July 2022. doi:10.3389/fmed.2022.864882. PMID: 35872797. PMCID: PMC9300826

    2022

  9. [17]

    Shimizu, P

    S. Shimizu, P. Hoyer, A. Hyv ¨arinen, and A. Kerminen. A linear non-gaussian acyclic model for causal discovery. Journal of Ma- chine Learning Research , 7:2003–2030, 2006. https://jmlr.org/papers/volume7/shimizu06a/shimizu06a.pdf

    2003

  10. [18]

    Di- rectlingam: A direct method for learning a linear non- gaussian structural equation model

    Shohei Shimizu, Takanori Inazumi, Yasuhiro Sogawa, Aapo Hyvarinen, Yoshinobu Kawahara, Takashi Washio, Patrik Hoyer, and Kenneth Bollen. Di- rectlingam: A direct method for learning a linear non- gaussian structural equation model. Journal of Ma- chine Learning Research, 12, 2011

    2011

  11. [19]

    Spirtes, C

    P. Spirtes, C. Glymour, and R. Scheines. Causa- tion, prediction, and search. MIT Press. Tibshirani, R, 58(1):267–288, 2001. 1996). Regression shrinkage and selection via the lasso. Journal of the Royal Sta- tistical Society: Series B (Methodological

    2001

  12. [20]

    E. Zhu, L. Zhang, Y . Liu, T. Ji, J. Dai, R. Tang, J. Wang, C. Hu, K. Chen, Q. Yu, Q. Lu, and Z. Ai.De- termining individual suitability for neoadjuvant sys- temic therapy in breast cancer patients through deep learning. Clinical & Translational Oncology, 2024. doi:10.1007/s12...

    2024 doi

  13. [2001]

    doi:10.1007/978-1-4757-3462-1

    doi

  14. [2023]

    doi:10.1186/s12874-023-02058-5

    doi

  15. [2024]

    PMID: 38572751

    doi:10.3322/caac.21834. PMID: 38572751. Epub 2024 Apr 4

    2024 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.