Pith. sign in

REVIEW 4 major objections 4 minor 20 references

Nonparametric Estimation of Conditional Survival Function with Time-Varying Covariates Using DeepONet

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A deep operator network estimates survival curves from entire time-varying covariate histories, dropping the instant-effect assumption.

desk verdict A useful operator-learning extension of survival analysis whose simulation is fine, but the ADNI application uses internal covariates under the paper's own external-covariate assumption, so the real-data claim doesn't hold. read the letter →

arxiv 2505.22748 v1 pith:6GUZFV66 submitted 2025-05-28 stat.ME

Bingqing Hu , Bin Nan This is my paper
classification stat.ME MSC 62N0162G0568T07
keywords conditionalsurvivalfunctiontime-varyingcovariatesDeepONetoperatorlearningfulllikelihoodrightcensoringexternalnonparametricestimation
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

The paper proposes a fully nonparametric way to estimate the conditional survival function when covariates change over time, modeling the log-hazard as a nonlinear operator of the entire covariate history rather than of the covariate value at the current time. The key construction is that any function $h$ yields a valid survival function via $S(t \mid \text{history}) = \exp\{-\int_0^t e^{h(s,\text{history}(s))}\,ds\}$, so the estimator only needs to learn $h$ from censored data. It fits $h$ with a DeepONet, a branch-and-trunk neural network whose branch encodes the history on a grid and whose trunk encodes the time point, minimizing the discretized negative full log-likelihood. In simulations the method recovers the true survival curves where a Cox model with instantaneous effects is visibly biased, and on a real dementia-progression cohort it attains a lower integrated Brier score than the Cox model. The paper claims this is the first estimator to treat time-varying covariate effects this flexibly.

What carries the argument

The central object is the DeepONet, a neural operator formed by the dot product of a branch network, which encodes the covariate history evaluated at $m$ grid points, and a trunk network, which encodes the time point $s$; their inner product approximates the log-hazard $h[s,\tilde X(s)]$. Future covariate values are masked with zeros in the input, and in the CNN variant causal padding excludes them, so only the history up to $s$ contributes. The loss is the discretized negative full log-likelihood of Equation (6), derived from the full likelihood under independent censoring. The universal approximation theorem for operators justifies that such a pair of networks can approximate any continuous nonlinear operator, which is the theoretical warrant for representing unrestricted history-dependent hazards.

What would settle it

Simulate survival data from a known nonlinear functional of the covariate history, such as $h(t) = \sin(\int_0^t X(s)\,ds) + X(t)^2$, generate failure times by inversion, and compare the DeepONet estimator's survival curves with the exact curves over many replicates; if the estimates deviate systematically beyond sampling variability, the universal-approximation-based claim of flexible noninstantaneous estimation would be refuted.

Watch

Extended reading notes

Core claim

The paper claims that the conditional survival function with time-varying covariates can be estimated nonparametrically by treating the log-hazard as an operator of the full covariate history. The central identity is Equation (2), which defines a valid survival function for any $h$, so the estimator carries no proportional-hazards or instantaneous-effect structure. Estimation is based on the full likelihood for right-censored data, discretized on a time grid, and the operator is approximated by a DeepONet whose branch network reads the history and whose trunk network reads the time point. The only substantive scope condition is that covariates are external, so future values never affect current risk. The paper demonstrates the approach with simulations and a real dementia-progression dataset.

Load-bearing premise

The estimator is only well-defined when the time-varying covariates are external, meaning each subject's risk at time $t$ depends solely on covariate values observed up to $t$ and never on future values.

Editorial extensions

If this is right

  • Under external time-varying covariates, the conditional survival function can be estimated without specifying any functional form for how covariate history affects risk.
  • When covariate effects are non-instantaneous, the Cox model is misspecified and yields biased survival predictions, while the proposed estimator is designed to remain valid.
  • The full-likelihood loss naturally handles right censoring and extends the earlier nonparametric neural-network approach from current covariate values to functional historical inputs.
  • The method provides a nonparametric way to estimate individual-level survival curves and, as illustrated, treatment effects under time-varying covariates without model-based adjustment.
  • On the dementia-progression data analyzed in the paper, the estimator achieves a lower cross-validated integrated Brier score than the Cox model.

Reading between the lines

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

  • The paper does not address uncertainty quantification over the learned operator; a natural extension would be to combine the estimator with ensembles or Bayesian DeepONet variants so that confidence bands reflect model uncertainty rather than only sampling variability.
  • The same operator formulation could be carried to competing risks or recurrent-event settings by replacing the scalar survival functional with a vector-valued operator over event types.
  • If the method scales computationally, it could make it practical to model cumulative exposure and delayed treatment effects in electronic health record data, where lab values and medication histories influence risk long after they are measured.
  • One testable extension is to probe the external-covariate assumption directly by including a biomarker that is partly determined by the failure process and checking whether the estimated survival curves become unstable.
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 manuscript proposes a DeepONet-based estimator of the conditional survival function with time-varying covariates, modeling the log-hazard as a nonlinear operator of the entire covariate history and thereby relaxing the instantaneous-effect assumption. The loss function is derived from the discretized negative full log-likelihood for right-censored data. A simulation with cumulative covariate effects shows that the proposed method recovers the true conditional survival curves, while a misspecified Cox model is biased. The method is then applied to ADNI data, where it is reported to achieve a lower integrated Brier score than a Cox model.

Significance. If the external-covariate condition is satisfied, the operator-learning formulation is a natural and useful extension of Hu and Nan (2023), and the construction in Equation (2) is a valid survival function for any log-hazard function. The loss derivation from the full likelihood is transparent, and the simulation provides an out-of-sample check against known truth. However, the ADNI application relies on covariates that appear to be internal under the paper's own definition, so the real-data demonstration does not currently support the central claim. In addition, the paper provides no consistency theory, which limits the force of the 'fully nonparametric' claim.

major comments (4)
  1. [Section 2 (before Eq. 1) and Section 4] The ADNI analysis uses MMSE score and hippocampal volume as 'external' time-varying covariates, but under the paper's own definition these are internal covariates. The paper states that the conditional survival probability is not well-defined if there is an internal covariate. MMSE and hippocampal volume are measurements of the disease process itself: they are only observed while the subject is at risk, and their future trajectories are altered by progression to dementia. Forward-filling missing values on a six-month grid does not convert an internal process into an external one. Consequently, the estimand in Equation (2), the likelihood in Equation (5), and the reported integrated Brier score comparison (0.20 vs 0.28) are not well-defined for this dataset. This is a load-bearing flaw in the real-data demonstration, not a cosmetic issue.
  2. [Equation (6)] The discretized loss does not exactly evaluate the integral in Equation (5). For a subject with event time y_i in the interval (t_{j-1}, t_j), the cumulative-hazard contribution uses the full interval width (t_j - t_{j-1}) and the event term is evaluated at the left endpoint, whereas the continuous-time likelihood would use (y_i - t_{j-1}) and would evaluate the hazard at y_i. This introduces a grid-dependent approximation error that is non-negligible on the coarse six-month grid used in the ADNI analysis. The authors should either state that this is a deliberate discrete-time model with a corresponding discrete hazard interpretation, or modify the loss to use the partial interval length.
  3. [Sections 2-3 and Discussion] The paper claims to be 'fully nonparametric with the weakest assumptions,' but no consistency or convergence result is provided for the proposed estimator. There is no theorem showing that the minimizer of the empirical loss in Equation (6) converges to the true conditional survival function as the sample size n and the number of grid points m grow, nor any analysis of the approximation error of the DeepONet class. Given the strength of the claim, the authors should either add theoretical support or explicitly qualify the scope of the claim.
  4. [Section 3] The simulation evidence is based on a single data-generating mechanism: one smooth trigonometric time-varying covariate with cumulative-sum effects on the hazard. Given the paper's claim of generality for 'arbitrary effects' of covariate histories, the simulations should include at least one scenario with delayed effects and one with sparse or irregular observation times, which is the setting encountered in the ADNI application. As it stands, one favorable scenario is too narrow to support the general claim.
minor comments (4)
  1. [Section 2.1] The statement 'h[s, eXi(s)] = h[s, eXi(τ)]' is confusing, since eXi(s) and eXi(τ) are different objects; what is meant is that the hazard at time s depends only on the history up to s, and the network input can contain future values masked as zero. Please rephrase for clarity.
  2. [Notation throughout] The notation eXi(∞), eXi(τ), and eXi(s) is used somewhat interchangeably. Define the convention once and use it consistently to avoid ambiguity about which covariate history is being conditioned on.
  3. [Section 4] The ADNI comparison reports a single cross-validated mean IBS difference (0.20 vs 0.28) without any measure of uncertainty. Reporting the fold-wise IBS values or a standard error would strengthen the comparison.
  4. [Figures 2-5] The figure captions would benefit from explicitly stating that the curves are averages over simulation replications and that the shaded bands are 90% intervals across replications, not predictive uncertainty for a new individual.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the loss is derived from the censored-data full likelihood, simulations are out-of-sample checks, and the self-citation to Hu and Nan (2023) is a technical precursor rather than a load-bearing circular premise.

full rationale

The central construction is Eq. (2), S(t|history)=exp(-∫_0^t e^{h(s,history(s))} ds), which is the definitional link between a hazard and a survival function. The estimator minimizes the discretized negative full log-likelihood in Eq. (6), obtained by substituting this hazard representation into the censored-data likelihood; no fitted parameter is inserted into the derivation and then renamed as a prediction. The DeepONet approximation theorem cited from Chen and Chen (1995) and Lu et al. (2021a) is external mathematical support, not an unverified self-citation. Simulations generate data from a cumulative-exposure hazard family that is contained in the model class being estimated, which is a standard simulation design rather than circular reasoning; the fitted networks are evaluated on newly generated covariate sets and compared with ground-truth survival curves. The only self-citation, Hu and Nan (2023), supplies the row-expansion data format and an earlier FNN likelihood approach; the present operator formulation and loss do not depend on any unverified claim from that citation, so the self-citation is not load-bearing. A separate, non-circularity concern is that Section 4 labels MMSE and hippocampal volume as external time-varying covariates, although under the paper's own Section 2 definition these disease-progression measures are plausibly internal covariates; that is a validity/correctness issue for the ADNI illustration, not a reduction of the estimator to its inputs.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim depends on standard censored-data likelihood theory, the external-covariate assumption, and the operator approximation capability of DeepONets. The only tuning parameter treated as a model choice is the grid size m. No new physical or conceptual entities are introduced.

free parameters (1)
  • number of time grid points m = 250 in simulations; six months for ADNI
    m is treated as a hyperparameter in Section 2.3, chosen as the median of optimal values from 10 datasets in simulations. It controls discretization of time and covariate histories.
assumptions (4)
  • domain assumption Time-varying covariates are external: the hazard at time t depends only on covariate history up to t, so future values do not influence current risk.
    Stated in Section 2 before Equation (1). The authors note that if covariates are internal, the conditional survival probability is not well-defined.
  • domain assumption Censoring is independent of failure time given covariates.
    Assumed in Section 2 to construct the full likelihood and to justify the loss function.
  • standard math Universal approximation theorem for operators, i.e., a DeepONet with branch and trunk networks can approximate any continuous nonlinear operator.
    Invoked in Section 2.1 via Chen and Chen [1995] and Lu et al. [2021a]. No proof is given, and it is assumed the hazard operator is continuous.
  • domain assumption The discretized loss in Equation (6) accurately approximates the continuous likelihood integral in Equation (5).
    The paper replaces the integral over [0, yi] with a sum over m grid subintervals. Convergence of this approximation is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nonparametric Estimation of Conditional Survival Function with Time-Varying Covariates Using DeepONet." pith.science (2026). https://pith.science/paper/6GUZFV66

@misc{pith2026250522748,
  author       = {Pith},
  title        = {Pith review of: Nonparametric Estimation of Conditional Survival Function with Time-Varying Covariates Using DeepONet},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6GUZFV66}},
  note         = {Machine review of arXiv:2505.22748}
}
read the original abstract

Traditional survival models often rely on restrictive assumptions such as proportional hazards or instantaneous effects of time-varying covariates on the hazard function, which limit their applicability in real-world settings. We consider the nonparametric estimation of the conditional survival function, which leverages the flexibility of neural networks to capture the complex, potentially long-term non-instantaneous effects of time-varying covariates. In this work, we use Deep Operator Networks (DeepONet), a deep learning architecture designed for operator learning, to model the arbitrary effects of both time-varying and time-invariant covariates. Specifically, our method relaxes commonly used assumptions in hazard regressions by modeling the conditional hazard function as an unknown nonlinear operator of entire histories of time-varying covariates. The estimation is based on a loss function constructed from the nonparametric full likelihood for censored survival data. Simulation studies demonstrate that our method performs well, whereas the Cox model yields biased results when the assumption of instantaneous time-varying covariate effects is violated. We further illustrate its utility with the ADNI data, for which it yields a lower integrated Brier score than the Cox model.

Figures

Figures reproduced from arXiv: 2505.22748 by the authors.

Figure 1
Figure 1. The unstacked DeepONet Structure. u, y, G(u)(y) in this figure correspond to x, s, h(x)(s), respectively. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Conditional survival curves for 9 different sets of covariates estimated by the DeepONet with FNN branch net [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Conditional survival curves for 9 different sets of covariates estimated by the DeepONet with CNN branch [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Conditional survival curves for 9 different sets of covariates estimated by the Cox model with sample size [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Conditional survival curves corresponding to [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 10 canonical work pages

  1. [1]

    David R. Cox. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34 0 (2): 0 187--202, 1972

    work page 1972

  2. [2]

    L. J. Wei. The accelerated failure time model: A useful alternative to the cox regression model in survival analysis. Statistics in Medicine, 11 0 (14-15): 0 1871--1879, 1992. doi:https://doi.org/10.1002/sim.4780111409. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.4780111409

    doi:10.1002/sim.4780111409 1992

  3. [3]

    Van Der Laan

    Torsten Hothorn, Peter Bühlmann, Sandrine Dudoit, Annette Molinaro, and Mark J. Van Der Laan. Survival ensembles . Biostatistics, 7 0 (3): 0 355--373, 12 2005. ISSN 1465-4644. doi:10.1093/biostatistics/kxj011. URL https://doi.org/10.1093/biostatistics/kxj011

    doi:10.1093/biostatistics/kxj011 2005

  4. [4]

    The state of boosting

    Greg Ridgeway. The state of boosting. Computing Science and Statistics, 31: 0 172--181, 1999

    work page 1999

  5. [5]

    Kogalur, Eugene H

    Hemant Ishwaran, Udaya B. Kogalur, Eugene H. Blackstone, and Michael S. Lauer. Random survival forests . The Annals of Applied Statistics, 2 0 (3): 0 841 -- 860, 2008. doi:10.1214/08-AOAS169. URL https://doi.org/10.1214/08-AOAS169

    doi:10.1214/08-aoas169 2008

  6. [6]

    Survival trees by goodness of split

    Michael Leblanc and John Crowley. Survival trees by goodness of split. Journal of the American Statistical Association, 88 0 (422): 0 457--467, 1993. doi:10.1080/01621459.1993.10476296. URL https://www.tandfonline.com/doi/abs/10.1080/01621459.1993.10476296

    arXiv 1993

  7. [7]

    Katzman, Uri Shaham, Alexander Cloninger, Jonathan Bates, Tingting Jiang, and Yuval Kluger

    Jared L. Katzman, Uri Shaham, Alexander Cloninger, Jonathan Bates, Tingting Jiang, and Yuval Kluger. Deepsurv: personalized treatment recommender system using a cox proportional hazards deep neural network. BMC Medical Research Methodology, 18 0 (1): 0 24, 2018

    work page 2018

  8. [8]

    Rnn-surv: A deep recurrent model for survival analysis

    Eleonora Giunchiglia, Anton Nemchenko, and Mihaela van der Schaar. Rnn-surv: A deep recurrent model for survival analysis. pages 23--32, 2018

    work page 2018

Show all 20 references
  1. [9]

    Steve Cygu, Hsien Seow, Jonathan Dushoff, and Benjamin M. Bolker. Comparing machine learning approaches to incorporate time-varying covariates in predicting cancer survival time. Scientific Reports, 13 0 (1370), 2023

    2023

  2. [10]

    Conditional distribution function estimation using neural networks for censored and uncensored data

    Bingqing Hu and Bin Nan. Conditional distribution function estimation using neural networks for censored and uncensored data. Journal of Machine Learning Research, 24 0 (223): 0 1--26, 2023. URL http://jmlr.org/papers/v24/22-0657.html

    2023

  3. [11]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3: 0 218--229, 03 2021 a . doi:10.1038/s42256-021-00302-5

    2021 doi

  4. [12]

    Kalbfleisch and Ross L

    John D. Kalbfleisch and Ross L. Prentice. The Statistical Analysis of Failure Time Data (2nd ed.). Hoboken, NJ:Wiley, 2002

    2002

  5. [13]

    Feed forward neural networks for the analysis of censored survival data: a partial logistic regression approach

    Elia Biganzoli, Patrizia Boracchi, Luigi Mariani, and Ettore Marubini. Feed forward neural networks for the analysis of censored survival data: a partial logistic regression approach. Statistics in Medicine, 17 0 (10): 0 1169--1186, 1998

    1998

  6. [14]

    Case-base neural network: Survival analysis with time-varying, higher-order interactions

    Jesse Islam, Maxime Turgeon, Robert Sladek, and Sahir Bhatnagar. Case-base neural network: Survival analysis with time-varying, higher-order interactions. Machine Learning with Applications, 16: 0 100535, 2024. ISSN 2666-8270. doi:https://doi.org/10.1016/j.mlwa.2024.100535. UR...

    2024

  7. [15]

    On universal approximation and error bounds for fourier neural operators

    Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. On universal approximation and error bounds for fourier neural operators. Journal of Machine Learning Research, 22 0 (290): 0 1--76, 2021. URL http://jmlr.org/papers/v22/21-0806.html

    2021

  8. [16]

    Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its applications to dynamic systems

    Tianping Chen and Hong Chen. Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its applications to dynamic systems. Neural Networks, IEEE Transactions on, pages 911 -- 917, 08 1995. doi:10.1109/72.392253

    1995 doi

  9. [17]

    DeepXDE : A deep learning library for solving differential equations

    Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. DeepXDE : A deep learning library for solving differential equations. SIAM Review, 63 0 (1): 0 208--228, 2021 b . doi:10.1137/19M1274067

    2021 doi

  10. [18]

    Serkan Kiranyaz, Onur Avci, Osama Abdeljaber, Turker Ince, Moncef Gabbouj, and Daniel J. Inman. 1d convolutional neural networks and applications: A survey. Mechanical Systems and Signal Processing, 151: 0 107398, 2021. ISSN 0888-3270. doi:https://doi.org/10.1016/j.ymssp.2020....

    2021

  11. [19]

    Wavenet: A generative model for raw audio, 2016

    Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio, 2016. URL https://arxiv.org/abs/1609.03499

    2016 arXiv

  12. [20]

    Assessment and comparison of prognostic classification schemes for survival data

    Erika Graf, Claudia Schmoor, Willi Sauerbrei, and Martin Schumacher. Assessment and comparison of prognostic classification schemes for survival data. Statistics in medicine, 18 17-18: 0 2529--45, 1999

    1999

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.