Pith. sign in

REVIEW 4 major objections 5 minor 20 references

Patient-Specific Models of Treatment Effects Explain Heterogeneity in Tuberculosis

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Contextualized logistic regression on over 3,000 TB patients attributes heterogeneous drug effects to anemia, HIV, and age of onset, and reaches 83.7% overall survival-prediction accuracy.

desk verdict A legitimate extension of the authors' contextualized ML framework to TB Portals, with a genuinely new bootstrap significance test, but the 'treatment effect' language overreaches what an observational association model supports. read the letter →

arxiv 2411.10645 v1 pith:7G4KKOAV submitted 2024-11-16 cs.LG stat.ML

Ethan Wu , Caleb Ellington , Ben Lengerich , Eric P. Xing This is my paper · ORCID
classification cs.LGstat.ML
keywords tuberculosiscontextualizedmachinelearningheterogeneoustreatmenteffectspatient-specificmodelingTBPortalsco-morbidityinteractionsmortalitypredictionpersonalizedmedicine
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 tuberculosis outcomes should be modeled patient-by-patient rather than through broad predefined subgroups, and it uses contextualized logistic regression to do so on the TB Portals dataset of over 3,000 patients. The central claim is that three context factors—anemia, age of onset, and HIV status—systematically change how well specific anti-TB drugs work, not just how sick patients are. On prediction, the contextualized model reaches 83.7% overall survival accuracy and lifts mortality-class accuracy from 19.3% for no-context logistic regression to 38.6%, showing that context carries signal that a flat model discards. If these interaction claims hold, anemia in particular would deserve monitoring as a modifier of drug efficacy, and age- and HIV-specific regimens would warrant clinical attention.

What carries the argument

The central object is the contextualized logistic regression decomposition $$\log \mathrm{odds}(Y|X,C)=X\$\beta$ + X\$\beta$(C) + \mu(C),$$ where $X$ is treatment, $C$ is patient context, $\beta$ is the population-level treatment effect, $\beta(C)$ is the context-dependent treatment effect, and $\mu(C)$ is the direct effect of context on mortality. A deterministic context encoder $f(C)$ maps multi-modal context into sample-specific coefficients, and significance is assessed by many bootstrap fits: one-sided $t$-tests for the homogeneous terms and, because the true $\beta(C)$ is assumed monotonic, a high-versus-low comparison across each context. This machinery is what turns the dataset into named drug-context interactions instead of a single average treatment effect.

What would settle it

Re-fit the contextualized model on the same TB Portals data without the monotonicity restriction and check whether the age-response curves for Kanamycin and Fluoroquinolones are U-shaped; if a middle-age peak or trough appears where the high-low test reports no effect, the reported interaction claims would not survive without the monotonicity assumption.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a single learned decomposition of the log odds of mortality—population treatment effect plus context-dependent treatment effect plus direct context effect—recovers structured interactions in observational TB data. The bootstrap significance tests identify anemia as altering the efficacy of aminoglycosides, clarithromycin, and pyrazinamide; age of onset as altering kanamycin and fluoroquinolones; and HIV as altering terizidone. The paper also reports that these personalized models are not less accurate than black-box alternatives: 83.7% overall accuracy, with mortality-class accuracy more than double the no-context logistic baseline. The authors take this as evidence that heterogeneity and multi-modality should be modeled jointly, and that patient context can be encoded into sample-specific coefficients rather than averaged away.

Load-bearing premise

The paper's significance tests assume that each drug's true treatment effect changes steadily in one direction as a patient context such as age increases, so an effect that rises then falls across age could be missed or mischaracterized by comparing only low and high values.

Editorial extensions

If this is right

  • TB outcome models that ignore context leave roughly half of the mortality signal on the table; the contextualized model's 0.386 mortality-class accuracy versus 0.193 for no-context logistic regression implies context is predictive, not noise.
  • Anemia should be tracked as a treatment-effect modifier in TB, not only as a prognostic marker, because the model links it to the efficacy of aminoglycosides, clarithromycin, and pyrazinamide.
  • Age at onset and HIV status identify drugs—kanamycin, fluoroquinolones, terizidone—whose dosing or choice may need to be personalized.
  • The same decomposition can generate patient-specific treatment recommendations directly from the per-patient coefficients $\beta(C)$.
  • Contextualized modeling offers a way to study heterogeneous treatment effects without exhaustive pairwise subgroup tests, which is especially valuable when many co-morbidities and drugs interact.

Reading between the lines

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

  • A natural next test is external validation: a retrospective cohort with hemoglobin measurements could check whether pyrazinamide- or aminoglycoside-containing regimens indeed fail more often in anemic patients, which the paper does not provide.
  • Because the significance test assumes monotonic context effects, U-shaped responses (for example, a drug working best in middle age) would be missed; a nonparametric test or a quadratic extension could reveal more interactions than this analysis reports.
  • The application likely transfers to other infectious diseases with co-morbidities and multi-modal records, such as sepsis or malaria, where context-dependent drug response is suspected but subgroup analyses are underpowered.
  • The accuracy gains in this observational dataset do not by themselves establish causal treatment effects; confounding by indication could create the same patterns, so the clinical claims need a causal or prospective follow-up.
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

4 major / 5 minor

Summary. The manuscript applies contextualized logistic regression, logodds(Y|X,C)=Xβ+Xβ(C)+µ(C), to the TB Portals dataset to decompose treatment-outcome relationships into homogeneous and context-dependent components. It reports improved mortality prediction relative to logistic regression and GBDT baselines (Table 1) and uses bootstrap-based significance tests to identify anemia, age of onset, and HIV as contexts that modify treatment-outcome associations. The authors interpret these findings as heterogeneous treatment effects and draw clinical conclusions about personalized tuberculosis regimens.

Significance. If the causal interpretation were supported, the paper would offer a clinically relevant demonstration of contextualized modeling in a heterogeneous observational cohort. Strengths include public code and data availability, a transparent model decomposition, and benchmarking against simple and black-box baselines. However, the central clinical claims currently exceed what the identification strategy supports: Eq. (1) is a conditional association model, and the significance tests rest on an untested monotonicity assumption. As it stands, the contribution is best viewed as a predictive modeling and association-mining case study rather than an estimate of causal treatment efficacy.

major comments (4)
  1. [Abstract; §2.3; §3; Eq. (1)] The parameter β(C) in Eq. (1) is the coefficient of treatment X in a logistic regression of the outcome on X and C; it is a conditional association, not a causal treatment effect. The manuscript repeatedly uses causal language, calling β(C) a 'treatment effect' and concluding that anemia, age of onset, and HIV 'alter efficacy' and should guide 'personalized treatment plans.' No causal identification assumptions (e.g., unconfoundedness, positivity, no interference) are stated, and TB Portals is an observational dataset in which treatment assignment is plausibly driven by resistance profiles, disease severity, and co-morbidities. The bootstrap tests in Section 1 validate sign consistency of the fitted estimator under the model, but they cannot validate a causal interpretation. The authors should either reframe the claims as heterogeneous associations or effect modification in the predictive sense, or add a causal identification argument with sensitivity analyses. This is load-bearing because the headline findings are efficacy claims.
  2. [§1 (significance of heterogeneous effects)] The proposed test for β(C) assumes 'the true β(C) is monotonic over C' and then tests the high versus low end of the observed context range. This assumption is stated but not justified, and non-monotonic relationships (e.g., U-shaped age effects) would be missed or mischaracterized. The manuscript also does not specify how C is scaled, how multiple contexts are handled jointly, how many bootstrap resamples are used, or how the one-sided t-test is constructed. At minimum, the authors should test the monotonicity assumption with a flexible model, provide sensitivity analyses, or restrict the heterogeneous-effect claims to contexts where the assumption is plausible or empirically checked.
  3. [§2.1; Table 1] The accuracy results in Table 1 are point estimates with no confidence intervals, number of cross-validation folds, or repeated splits. The mortality class accuracy is only 0.386, and the claim that contextualization 'particularly impro[ves] the prediction of mortality' needs supporting uncertainty quantification, such as confidence intervals, AUROC or precision-recall curves, and a statistical test of the difference from the GBDT baseline (0.837 versus 0.834). Without these, the empirical superiority of the proposed model over the context-augmented GBDT is not established.
  4. [§2 (unspecified methods details)] The main empirical claims require several modeling choices that are not described in the paper: the exact cohort size after filtering, the definition of the mortality outcome and its time horizon, the encoding of treatments (binary drug flags versus doses), the architecture and training procedure of the context encoder, the train/test split, and the bootstrap protocol. The public code mitigates this concern, but the manuscript should summarize these choices (or cite a methods appendix) so that the reported 83.7% accuracy and the significance tests can be assessed independently of re-running the code.
minor comments (5)
  1. [§2.2; §2.3; Figure labels] Several figure references are incorrect: 'Fig. 2.3' appears in §2.3 where Figure 4 is intended, and 'Figures 3, 2.3' in §2.2 is confusingly formatted.
  2. [§2.2] The sentence 'socioeconomic and social factors such an unemployment, disabled, and student identity' contains a typo: 'such an' should be 'such as.'
  3. [Eq. (1) and surrounding text] The notation 'Xβ + Xβ(C)' is ambiguous when X is a vector of treatments; it should be written as X^T β + X^T β(C) or with explicit indexing to clarify that β(C) is a vector of treatment-context interaction coefficients.
  4. [§2.3] The phrase 'treatment dose-response' overstates what is modeled if treatments are binary drug-presence indicators; consider using 'treatment association' or 'drug-presence effect' unless actual doses are available and used.
  5. [Introduction, HIV statistics] The claim that 'HIV patients are 18 times more likely to develop active TB, with TB responsible for one-third of HIV-related deaths globally' is cited to Haiminen et al. (2022), but that reference appears to be about metagenomic classification pipelines and does not support this statement; a proper epidemiological citation is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is fit to data, benchmarked against baselines, and the heterogeneity tests are empirical checks on fitted parameters, not inputs recycled as outputs.

full rationale

The paper's derivation chain is self-contained. Equation (1) posits a contextualized logistic model (logodds(Y|X,C)=Xβ+Xβ(C)+μ(C)); this is a modeling assumption, not a result derived from the data. The coefficients are estimated from TB Portals data, and Table 1 benchmarks the contextualized model against context-free logistic regression, GBDT, and concatenated-context baselines. The headline findings about anemia, age of onset, and HIV being influential for treatment efficacy come from bootstrap-based significance tests on the fitted β(C) function (Section 1, Figures 3 and 4). These tests are not definitional: the model could in principle learn a constant or zero β(C), and the tests are designed to detect variation in the estimated function under stated assumptions. The monotonicity assumption for the high-vs-low test is an explicit statistical assumption, not a tautology. The cited contextualized-ML papers (Lengerich et al., Ellington et al., Al-Shedivat et al.) are used as methodological provenance and software references, not as a uniqueness theorem or as the evidentiary basis for the TB-specific empirical findings. Concerns about causal identification of 'treatment effects' from observational data, and about the absence of external held-out validation, are validity or generalizability issues, not circularity, and are not counted in this analysis.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central findings depend on the contextualized model's fitted encoder, the untested monotonicity assumption for the significance test, and an implicit causal interpretation of observational associations. No new physical entities are introduced.

free parameters (2)
  • Context encoder neural network weights = Not reported
    The contextualized model's predictions and inferred β(C) depend on the trained encoder, whose architecture, learning rate, and regularization are not specified in the paper.
  • Bootstrap test parameters = Not reported
    Number of bootstrap samples, confidence level, and centering procedure are not specified, affecting the significance results.
assumptions (4)
  • domain assumption All patients' outcome distributions belong to the same parametric family, specifically logistic regression with linear effects
    Stated in Section 1: 'An enabling assumption is that all P_i belong to the same parametric family, i.e., Y_i ∼ P(Y_i | X_i, θ_i)'.
  • ad hoc to paper The true β(C) is monotonic over the context variable C
    Stated in Section 1: 'we assume the true β(C) is monotonic over C and instead test the difference between the high and low end of the observed range of C'.
  • domain assumption Bootstrap estimates of the model parameters are consistent and the one-sided t-test on signs is valid
    Used to assess significance in Section 1; no proof or simulation is provided.
  • domain assumption Observed treatment-context-outcome associations can be interpreted as treatment effects
    The paper repeatedly uses 'treatment effects' language (e.g., abstract, Section 2.3) without discussing unmeasured confounding or causal identification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Patient-Specific Models of Treatment Effects Explain Heterogeneity in Tuberculosis." pith.science (2026). https://pith.science/paper/7G4KKOAV

@misc{pith2026241110645,
  author       = {Pith},
  title        = {Pith review of: Patient-Specific Models of Treatment Effects Explain Heterogeneity in Tuberculosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7G4KKOAV}},
  note         = {Machine review of arXiv:2411.10645}
}
read the original abstract

Tuberculosis (TB) is a major global health challenge, and is compounded by co-morbidities such as HIV, diabetes, and anemia, which complicate treatment outcomes and contribute to heterogeneous patient responses. Traditional models of TB often overlook this heterogeneity by focusing on broad, pre-defined patient groups, thereby missing the nuanced effects of individual patient contexts. We propose moving beyond coarse subgroup analyses by using contextualized modeling, a multi-task learning approach that encodes patient context into personalized models of treatment effects, revealing patient-specific treatment benefits. Applied to the TB Portals dataset with multi-modal measurements for over 3,000 TB patients, our model reveals structured interactions between co-morbidities, treatments, and patient outcomes, identifying anemia, age of onset, and HIV as influential for treatment efficacy. By enhancing predictive accuracy in heterogeneous populations and providing patient-specific insights, contextualized models promise to enable new approaches to personalized treatment.

Figures

Figures reproduced from arXiv: 2411.10645 by the authors.

Figure 1
Figure 1. Contextualized Modeling Diagram In this study, we use contextualized machine learn￾ing to analyze heterogeneous treatment effects in the TB Portals dataset by inferring patient-specific treat￾ment models. For context C, treatments X, and mor￾tality Y , contextualized logistic regression decom￾poses treatment effect models into three components logodds(Y |X, C) = Xβ + Xβ(C) + µ(C) (1) describing the context-invariant… view at source ↗
Figure 2
Figure 2. Homogeneous Effects of Context [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Heterogeneous Treatment Effects. The sig￾nificance of the change in treatment effi￾cacy over context, visualized as −log(p-val) 2.2. Contextualization Reveals Direct Context and Treatment Effects Exploring the direct effects of context µ(C) ( [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Heterogeneous Effects of Treatments 2.3. Contextualized Models Reveal Critical Context-Treatment Interactions Our analysis of treatment heterogeneity (β(C)) across patient contexts uncovers key interactions that have clinical implications. Anemia significantly al￾ters …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 16 canonical work pages

  1. [1]

    Maruan Al-Shedivat, Avinava Dubey, and Eric P. Xing. Contextual Explanation Networks , September 2020. URL http://arxiv.org/abs/1705.10301. arXiv:1705.10301 [cs, stat]

    arXiv 2020

  2. [2]

    Lino Ferreira da Silva Barros, G

    M.H. Lino Ferreira da Silva Barros, G. Oliveira Alves, L. Morais Florêncio Souza, E. da Silva Rocha, J.F. Lorenzato de Oliveira, T. Lynn, V. Sampaio, and P.T. Endo. Benchmarking machine learning models to assist in the prognosis of tuberculosis. Informatics, 8: 0 27, 2021. doi:10.3390/informatics8020027. URL https://www.mdpi.com/2227-9709/8/2/27

    work page doi:10.3390/informatics8020027 2021

  3. [3]

    Chakaya, M

    J. Chakaya, M. Khan, F. Ntoumi, E. Aklillu, R. Fatima, P. Mwaba, N. Kapata, S. Mfinanga, S.E. Hasnain, P.D.M.C. Katoto, et al. Global tuberculosis report 2020 - reflections on the global tb burden, treatment and prevention efforts. Int. J. Infect. Dis., 113: 0 S7--S12, 2021. doi:10.1016/j.ijid.2021.02.107. URL https://pubmed.ncbi.nlm.nih.gov/33716195/

    work page arXiv 2020

  4. [4]

    Dartois and E.J

    V.A. Dartois and E.J. Rubin. Anti-tuberculosis treatment strategies and drug development: challenges and priorities. Nat. Rev. Microbiol., 20: 0 685--701, 2022. doi:10.1038/s41579-022-00721-4. URL https://pubmed.ncbi.nlm.nih.gov/35478222/

    work page arXiv 2022

  5. [5]

    Contextualized Policy Recovery: Modeling and Interpreting Medical Decisions with Adaptive Imitation Learning

    Jannik Deuschel, Caleb N. Ellington, Benjamin J. Lengerich, Yingtao Luo, Pascal Friederich, and Eric P. Xing. Contextualized Policy Recovery : Modeling and Interpreting Medical Decisions with Adaptive Imitation Learning , October 2023. URL http://arxiv.org/abs/2310.07918. arXiv:2310.07918 [cs, stat]

    work page Pith review arXiv 2023

  6. [6]

    Ellington, Benjamin J

    Caleb N. Ellington, Benjamin J. Lengerich, Wesley Lo, Aaron Alvarez, Andrea Rubbi, Manolis Kellis, and Eric P. Xing. Contextualized: Heterogeneous Modeling Toolbox . Journal of Open Source Software, 9 0 (97): 0 6469, May 2024. ISSN 2475-9066. doi:10.21105/joss.06469. URL https://joss.theoj.org/papers/10.21105/joss.06469

    work page doi:10.21105/joss.06469 2024

  7. [7]

    W.S. Fox, N. Strydom, M.Z. Imperial, L. Jarlsberg, and R.M. Savic. Examining nonadherence in the treatment of tuberculosis: The patterns that lead to failure. Br. J. Clin. Pharmacol., 89: 0 1965--1977, 2023. URL https://pubmed.ncbi.nlm.nih.gov/36036095/

    work page arXiv 1965

  8. [8]

    Impact of metagenomic classification on the computational efficiency and clinical utility of microbiome analysis pipelines

    Niina Haiminen, Paul Greco, Allison Davis, Ekaterina Smirnova, Sudip Mukherjee, Rob Knight, James Wolff, and Scott Kelley. Impact of metagenomic classification on the computational efficiency and clinical utility of microbiome analysis pipelines. Scientific Reports, 12 0 (1): 0 18832, 2022. doi:10.1038/s41598-022-23316-4. URL https://www.nature.com/articl...

    work page doi:10.1038/s41598-022-23316-4 2022

Show all 20 references
  1. [9]

    Varying- Coefficient Models

    Trevor Hastie and Robert Tibshirani. Varying- Coefficient Models . Journal of the Royal Statistical Society: Series B (Methodological), 55 0 (4): 0 757--779, 1993. ISSN 2517-6161. doi:10.1111/j.2517-6161.1993.tb01939.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.251...

    1993

  2. [10]

    Bosch, Walter C

    Sheila Isanaka, Said Aboud, Ferdinand Mugusi, Ronald J. Bosch, Walter C. Willett, and Wafaie W. Fawzi. Iron deficiency and anemia predict mortality in patients with tuberculosis. The Journal of Nutrition, 142 0 (2): 0 350--357, 2012. doi:10.3945/jn.111.147181. URL https://doi....

    2012 doi

  3. [11]

    Benjamin Lengerich, Bryon Aragam, and Eric P. Xing. Learning Sample - Specific Models with Low - Rank Personalized Regression , October 2019. URL http://arxiv.org/abs/1910.06939. arXiv:1910.06939 [cs, stat]

    2019 arXiv

  4. [12]

    Ellington, Andrea Rubbi, Manolis Kellis, and Eric P

    Benjamin Lengerich, Caleb N. Ellington, Andrea Rubbi, Manolis Kellis, and Eric P. Xing. Contextualized Machine Learning , October 2023. URL http://arxiv.org/abs/2310.11340. arXiv:2310.11340 [cs, stat]

    2023 arXiv

  5. [13]

    Lengerich, Bryon Aragam, and Eric P

    Benjamin J. Lengerich, Bryon Aragam, and Eric P. Xing. Personalized regression enables sample-specific pan-cancer analysis. Bioinformatics (Oxford, England), 34 0 (13): 0 i178--i186, July 2018. ISSN 1367-4811. doi:10.1093/bioinformatics/bty250

    2018 doi

  6. [14]

    Lengerich, Maruan Al-Shedivat, Amir Alavi, Jennifer Williams, Sami Labbaki, and Eric P

    Benjamin J. Lengerich, Maruan Al-Shedivat, Amir Alavi, Jennifer Williams, Sami Labbaki, and Eric P. Xing. Discriminative Subtyping of Lung Cancers from Histopathology Images via Contextual Deep Learning , November 2022 a . URL https://www.medrxiv.org/content/10.1101/2020.06.25...

    2022 doi

  7. [15]

    Automated Interpretable Discovery of Heterogeneous Treatment Effectiveness : A COVID -19 Case Study

    Benjamin J Lengerich, Mark E Nunnally, Yin Aphinyanaphongs, Caleb Ellington, and Rich Caruana. Automated Interpretable Discovery of Heterogeneous Treatment Effectiveness : A COVID -19 Case Study . J. Biomed. Inform., page 104086, April 2022 b . ISSN 1532-0464. doi:10.1016/j.jb...

    2022

  8. [16]

    Mdluli, T

    K. Mdluli, T. Kaneko, and A. Upton. The tuberculosis drug discovery and development pipeline and emerging drug targets. Cold Spring Harb. Perspect. Med., 5: 0 a021154, 2015. doi:10.1101/cshperspect.a021154. URL https://pubmed.ncbi.nlm.nih.gov/25635061/

    2015

  9. [17]

    Nagu, Said Aboud, Rose S

    Tumaini J. Nagu, Said Aboud, Rose S. Mwiru, Mecky I. Matee, Madhu Rao, Wafaie W. Fawzi, and Ferdinand M. Mugusi. Anaemia in patients with tuberculosis is associated with increased mortality. The International Journal of Tuberculosis and Lung Disease, 18 0 (1): 0 161--167, 2014...

    2014 doi

  10. [18]

    Rosenthal, A

    A. Rosenthal, A. Gabrielian, E. Engle, D.E. Hurt, S. Alexandru, V. Crudu, et al. The tb portals: an open-access, web-based platform for global drug-resistant-tuberculosis data sharing and analysis. J. Clin. Microbiol., 55 0 (11): 0 3267--3282, 2017. doi:10.1128/JCM.01013-17. U...

    2017 doi

  11. [19]

    Kaseb, Hesham M

    Zeya Wang, Ahmed O. Kaseb, Hesham M. Amin, Manal M. Hassan, Wenyi Wang, and Jeffrey S. Morris. Bayesian Edge Regression in Undirected Graphical Models to Characterize Interpatient Heterogeneity in Cancer . Journal of the American Statistical Association, 117 0 (538): 0 533--54...

    2022

  12. [20]

    Zumla, J

    A. Zumla, J. Chakaya, R. Centis, L. D’Ambrosio, P. Mwaba, M. Bates, N. Kapata, T. Nyirenda, D. Chanda, S. Mfinanga, et al. Tuberculosis treatment and management—an update on treatment regimens, trials, new drugs, and adjunct therapies. Lancet Respir. Med., 3: 0 220--234, 2015....

    2015

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.