acceptodds
Under review as a conference paper at ICLR 2027

KAPLAN: Kolmogorov-Arnold Prognostic Learnable Activation Networks for Survival Analysis

Abstract

Survival analysis aims to model how covariates and time jointly shape the time-to-event distribution under right censoring. Classical methods such as the Cox model and generalised additive models (GAMs) require interactions and time-varying effects to be manually specified, which is increasingly impractical on rich clinical datasets. We introduce KAPLAN-HR, a B-spline Kolmogorov-Arnold Network (KAN) for nonparametric estimation of the conditional hazard as a joint function of covariates and time. A single-layer KAPLAN-HR model recovers a GAM, while deeper architectures capture interactions and time-varying effects through composition. We establish a convergence rate for the nonparametric KAN hazard estimator whose exponent depends only on the smoothness of the underlying KAN representation and not on the covariate dimension, thereby mitigating the curse of dimensionality for KAN-representable targets. In evaluations over six clinical benchmark datasets, KAPLAN-HR attains predictive performance competitive with established statistical and deep learning survival methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
survival analysis, Kolmogorov-Arnold networks, time-to-event prediction, nonparametric estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ykXHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KAPLAN-HR, a survival model based on Kolmogorov–Arnold Networks. It models event risk using patient characteristics and time, allowing interactions and effects that change over time. The authors provide a convergence analysis and evaluate the model on six clinical datasets.

Critical strengths

1- The paper provides a theoretical convergence result under clearly stated assumptions, rather than presenting only an application of an existing neural architecture.

2- The evaluation includes nine baseline methods, ten train–validation–test splits, and several prediction and calibration measures. The authors also acknowledge that the clinical results show competitive performance rather than clear superiority.

Critical weaknesses

1- The model-level novelty appears limited. The paper itself discusses earlier KAN-based survival models, including hazard regression. Although the theoretical analysis adds value, the methodological difference from the closest approaches should be explained more clearly.

2- The practical improvement is not convincing enough. KAPLAN-HR achieves the best survival prediction error and calibration score on only one of six datasets for each measure. None of the comparisons with the best-scoring baseline reaches statistical significance. These results support competitiveness, but leave the practical reason for choosing this model unclear.

3- The simulation comparison with GAM needs better control. KAPLAN-HR’s spline capacity increases with sample size, whereas the GAM uses fixed basis sizes. This could partly explain why GAM’s error stops improving in Figure 2. Allowing both models to increase their flexibility would make this comparison more informative.

4- The evidence for handling high-dimensional problems is narrow. The theoretical rate assumes that the true log-hazard has the specified smooth KAN representation. The higher-dimensional simulations keep only two informative variables and add noise variables. They therefore do not establish performance when many relevant variables interact.

Questions for authors

1- Does the advantage over GAM remain when its basis size also increases with the sample size?

2- Could the authors compare against a standard neural network using the same hazard formulation, training objective, and tuning budget to isolate the benefit of the KAN architecture?

3- How does the method perform when many covariates are informative or the true hazard does not follow the assumed KAN structure?

4 comments
AuthorsOpenReview ↗

Response (1/2)

The model-level novelty appears limited. The paper itself discusses earlier KAN-based survival models, including hazard regression. Although the theoretical analysis adds value, the methodological difference from the closest approaches should be explained more clearly.

We have made the comparison with the closest approaches explicit in the related work (lines 353–357). KAN survival models are distinguished by the structural restriction they place on the hazard. CoxKAN (Knottenbelt et al., 2025) and the model of Cheng and Hu (2025) place a KAN inside the proportional-hazards form, and KAN-AFT (Jose et al., 2025) imposes an accelerated failure time form. SurvKAN (Mastroleo et al., 2026) uses the same formulation as KAPLAN-HR, modelling the log-hazard as a KAN of the covariates and time trained on the full censored-data likelihood.

At the model level, KAPLAN-HR therefore differs from CoxKAN, the model of Cheng and Hu and KAN-AFT in the hazard structures it represents. Relative to SurvKAN, our contribution lies in the statistical analysis of the estimator. To our knowledge, no prior theoretical analysis has been conducted for any of these models.

These results support competitiveness, but leave the practical reason for choosing this model unclear.

To clarify when KAPLAN-HR is a suitable choice:

  • Prefer KAPLAN-HR when the structural form of the hazard is unknown, particularly whether effects vary with time or interact. Proportional-hazards models fix the ordering of subjects' risk over time, and additive models require interactions to be specified in advance.
  • Prefer a simpler model when its form is known to match the data.
  • Against deep, flexible baselines (DSM, DeepHit, CoxTime, SuMo-Net), KAPLAN-HR attains competitive performance. Its advantage therefore lies in the KAN parameterisation itself, which decomposes fitted effects into plottable, inspectable univariate edges (subject to a trade-off with depth) and under which we provide a convergence rate with a dimension-free exponent for a KAN-representable log-hazard.

We have added this guidance to the conclusion (lines 495–502).

The simulation comparison with GAM needs better control. Does the advantage over GAM remain when its basis size also increases with the sample size?

We have updated the simulations so the GAM's basis grows with the sample size. The spline basis of each covariate and of time now has kn=4Gn+4k_n = 4G_n + 4 functions, where GnG_n is KAPLAN-HR's number of grid intervals, so both models' capacity grows at the same rate.

Median over repetitions of the paired ISESISE_S ratio GAM ÷ KAPLAN-HR (same training set), R=20R = 20

n=512n = 512n=8,192n = 8{,}192n=65,536n = 65{,}536
DGP-10.870.540.59
DGP-21.538.7628.18
DGP-32.042.287.62
DGP-41.092.188.54

With this adaptive basis, the GAM's error on the additive DGP-1 continues to descend across all nn, where it previously plateaued, and is lower than KAPLAN-HR's at every nn. On DGPs 2 to 4, the GAM's error still plateaus (Figure 2), and its ratio to KAPLAN-HR's grows from 1.09–2.04 at n=512n = 512 to 7.62–28.18 at n=65,536n = 65{,}536.

AuthorsOpenReview ↗

Response (2/2)

Could the authors compare against a standard neural network using the same hazard formulation, training objective, and tuning budget to isolate the benefit of the KAN architecture?

We have added this comparison to both the simulations and the benchmark. MLP-HR replaces the KAN in KAPLAN-HR with a ReLU MLP on the concatenated input [x,t][x, t], keeping the same objective, training and prediction. In the simulations, each fit trains six MLP architectures (width {8,16,32}\{8, 16, 32\} × depth {2,3}\{2, 3\}) and keeps the one with the lowest validation NLL.

Median over repetitions of the paired ISESISE_S ratio MLP-HR ÷ KAPLAN-HR (same training set), n=65,536n = 65{,}536, R=20R = 20, with 95% bootstrap intervals

DGP-1DGP-2DGP-3DGP-4
4.35 [3.90, 5.28]0.61 [0.55, 0.67]0.81 [0.70, 0.99]1.32 [1.11, 1.49]

On the same training set, MLP-HR attains a median of 4.35 times KAPLAN-HR's error on the additive DGP-1 and 1.32 times on the crossing-hazards DGP-4. It attains lower error on DGP-2 and DGP-3, where covariates interact, at 0.61 and 0.81 times.

On the benchmark, MLP-HR is tuned identically to the other models, with the same budget of 30 trials per split, over the same ranges as DeepSurv.

KAPLAN-HR against MLP-HR on the benchmark (mean over splits)

KAPLAN-HR betterMLP-HR betterSignificant (Holm p<0.05p < 0.05)
C-TD4/62/6MIMIC-III, for KAPLAN-HR
IBS6/60/6FLCHAIN, for KAPLAN-HR
ICI4/62/6none

The only significant differences are in C-TD on MIMIC-III and in IBS on FLCHAIN, where KAPLAN-HR attains the better value. Replacing the KAN with an MLP therefore does not improve predictive performance on the clinical benchmark.

We have added MLP-HR to the benchmarks (Table 1, Appendix F) and reported both comparisons in a new appendix (Appendix E.6).

How does the method perform when many covariates are informative or the true hazard does not follow the assumed KAN structure?

We have revised DGP-3 in the simulations (Figure 2) to test convergence with 32 interacting covariates. In this scenario, KAPLAN-HR's error descends more slowly than the theory's predicted rate from n=29n = 2^9 to 2102^{10}, and at that rate from n=210n = 2^{10} onward.

For a true hazard outside the assumed KAN structure, we have added an experiment fitting DGP-4 at hidden width 1 (Appendix E.4, p. 31). DGP-4 satisfies (C4) at hidden width Q=3Q = 3 but not at Q=1Q = 1, since a width-1 network keeps the effect of x1x_1 relative to that of x2x_2 fixed over time, while in DGP-4 only the effect of x1x_1 reverses sign at t=1/2t = 1/2.

Median over repetitions of the paired ISESISE_S ratio width 1 ÷ width 3 on DGP-4 (same training set), R=20R = 20, with 95% bootstrap intervals

n=512n = 512n=8,192n = 8{,}192n=65,536n = 65{,}536
Width 1 ÷ width 30.90 [0.79, 1.01]1.14 [1.02, 1.34]3.26 [3.06, 3.90]

The width-1 model attains 0.90 and 1.14 times the width-3 model's error at n=512n = 512 and 8,1928{,}192. At larger nn its error flattens, reaching 3.26 times the width-3 model's error at n=65,536n = 65{,}536, consistent with a remaining approximation error, while the width-3 model continues at the theory's rate (Appendix E.4). A violation of (C4) therefore manifested in this experiment as an error curve that flattens as nn grows.

Thank you for providing additional results. However, my main concern remains the limited novelty of the work.

Thank you for providing additional results. However, my main concern remains the limited novelty of the work.

We thank the reviewer for their response. The key novelty of the paper lies in its statistical analysis of KAPLAN-HR. We clarify what was adopted and what is new.

  • Adopted: The KAN architecture and choice of B-spline edge parameterisation (Liu et al., 2025), and the observation that B-spline KANs admit dimension-free approximation for compositions of smooth univariate edges (Liu et al., 2025). The rate machinery is the M-estimator rate theorem of van der Vaart and Wellner (2023, Thm 3.4.12). The strategy of establishing convergence bounds for neural survival estimators follows Deep Extended Hazards (Zhong et al., 2021) and Deep Partially Linear Transformations (Yin et al., 2025), and the choice of Hellinger distance as the metric follows Neural Frailty Machines (Wu et al., 2023).
  • New: (i) A convergence analysis for a KAN parameterisation of the joint (x,t) log-hazard, which is the form that admits time-varying effects without a covariate-free baseline. (ii) Four proven propositions instantiating the rate theorem's hypotheses for a KAN sieve under right-censoring: an approximation bound for KANs of arbitrary depth, a bracketing entropy bound, a likelihood ratio bound, and a bracketing integral envelope of the shape the theorem requires. (iii) An extension of the rate theorem to an ηn\eta_n-near maximiser, traced through van der Vaart and Wellner (2023, Thm 3.4.1 and Lem 3.4.10). Training the non-convex objective via stochastic gradient methods does not deliver the exact sieve MLE, and the slack ηn\eta_n is required to account for this. (iv) The resulting rate, whose exponent depends only on the smoothness of the truth's KAN representation.

To our knowledge, no prior work has established a convergence rate for any KAN-based survival model.

Reviewer_EUrjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KAPLAN-HR, which models the log-hazard using a neural network with spline functions. A single layer provides an additive proportional-hazards model, and deeper networks allow interactions and time-varying effects. Convergence rates are established under a smooth compositional assumption. The method are evaluated on simulations and six clinical datasets.

Critical strengths
  1. The connection between additive spline models and deeper networks is clear.
  2. The theoretical analysis is a strength, rather than simply replacing one network architecture with another.
  3. The experiments include several relevant neural survival methods.
Critical weaknesses
  1. The reason to prefer KAPLAN-HR over existing neural survival models seems not sufficiently clear. For example, please explain whether the expected advantage is better prediction, fewer parameters, easier interpretation, or lower computational cost, and provide evidence for that advantage.
  2. The numerical results seem not a clear practical gain. KAPLAN-HR has the highest mean concordance on four of six datasets, but the best IBS and calibration error on only one dataset each. None of the comparisons with the best-scoring baseline appear to be statistical significant. This provides limited evidence that users should switch methods.
  3. The theretical results depend on the true log-hazard having a fixed-depth, fixed-width composition of smooth univariate functions. It is not a general result for high-dimensional hazards. Not sure whether it establish a better rate than competing neural estimators.
  4. Please clarify the hyperparameter tuning and preprocessing procedures, e.g. tuning using split 0 and results reported on 10 splits (observations appear in both tuning and testing?) Time normalization uses the full-sample range before splitting (should transformation applied without testing samples?)
  5. The theory discusses quadrature error, but training uses a minibatch-specific subset of grid points. Please explain how the theoretical error bound applies to the training rule and report sensitivity to both grid and batch size.
Questions for authors

please see above

5 comments
AuthorsOpenReview ↗

Response (1/2)

The reason to prefer KAPLAN-HR over existing neural survival models seems not sufficiently clear. For example, please explain whether the expected advantage is better prediction, fewer parameters, easier interpretation, or lower computational cost, and provide evidence for that advantage.

The reason to prefer KAPLAN-HR comes from its joint hazard formulation, which relaxes the restrictions of proportional-hazards and additive models, and from its KAN parameterisation, which distinguishes it from deep, flexible baselines. To clarify when KAPLAN-HR is a suitable choice:

  • Prefer KAPLAN-HR when the structural form of the hazard is unknown, particularly whether effects vary with time or interact. Proportional-hazards models fix the ordering of subjects' risk over time, and additive models require interactions to be specified in advance.
  • Prefer a simpler model when its form is known to match the data.
  • Against deep, flexible baselines (DSM, DeepHit, CoxTime, SuMo-Net), KAPLAN-HR attains competitive performance. Its advantage therefore lies in the KAN parameterisation itself, which decomposes fitted effects into plottable, inspectable univariate edges (subject to a trade-off with depth) and under which we provide a convergence rate with a dimension-free exponent for a KAN-representable log-hazard.

We have added this guidance to the conclusion (lines 495–502).

The numerical results seem not a clear practical gain.

We acknowledge that the benchmark establishes competitiveness rather than a predictive advantage. It serves three purposes: (i) comparability with prior studies, (ii) showing that the KAN parameterisation is not traded against predictive performance, and (iii) showing that performance holds on real data, where it is unknown whether the log-hazard admits the KAN representation the theory requires. KAPLAN-HR attains the best mean rank in C-TD and the second best in IBS across the six datasets, and no model ranks significantly above it on any metric.

The theoretical results depend on the true log-hazard having a fixed-depth, fixed-width composition of smooth univariate functions. It is not a general result for high-dimensional hazards.

We acknowledge that our result does not cover every high-dimensional hazard. It requires the true log-hazard to admit a KAN representation of fixed depth and width with smooth univariate components, per assumption (C4), consistent with prior KAN theory (Liu et al., 2025, Thm. 2.1; Liu, Chatzi & Lai, 2025). This class admits additive effects and, with two or more layers, interactions and time-varying effects built by composing smooth univariate functions, but not every smooth function of (x,t)(x, t). Outside it, Theorem 1 gives no rate. We have made these implications explicit where (C4) is introduced (lines 224–227).

Not sure whether it establish a better rate than competing neural estimators.

We clarify that we do not claim a faster rate than competing estimators. Our analysis establishes the convergence rate of the estimator as implemented. To our knowledge, existing convergence results do not cover a KAN sieve over the joint log-hazard or the implementation choices that map onto it. The dimension-free exponent arises from pairing the structure that (C4) imposes with a sieve that mirrors it. Other estimators that exploit the same structure may attain the same exponent. We have added this point to Section 4 (lines 190–192).

AuthorsOpenReview ↗

Response (2/2)

Please clarify the hyperparameter tuning and preprocessing procedures, e.g. tuning using split 0 and results reported on 10 splits (observations appear in both tuning and testing?) Time normalization uses the full-sample range before splitting (should transformation applied without testing samples?)

We have revised both procedures so that no split's test observations inform its tuning or preprocessing. Each tunable model is now tuned separately on each split, and times are rescaled by each split's training range, with validation and test observations beyond the largest training time censored at that time. We have re-run the entire benchmark under the revised protocol, for every model, dataset and split, and report its results throughout the revision (Table 1, Figure 3, Appendix F).

KAPLAN-HR mean rank, with the number of datasets where it attains the best value, submitted vs. revised benchmark

C-TD ↑IBS ↓ICI ↓
Submitted (of 10 models)1.7 (4/6)2.8 (1/6)3.3 (1/6)
Revised (of 11 models)2.00 (3/6)3.00 (1/6)5.50 (1/6)
Friedman pp, revised8.1×10−48.1 \times 10^{-4}6.7×10−46.7 \times 10^{-4}0.76

Under the revised protocol, KAPLAN-HR retains the best mean rank on C-TD and ranks second on IBS. Its mean rank on ICI is 5.50, against 3.3 in the submitted benchmark, however the Friedman test detects no significant difference in mean ICI rank among the eleven models (p=0.76p = 0.76). The revised results also include the corrected GAM fit and an additional baseline, MLP-HR, which replaces the KAN with an MLP.

The theory discusses quadrature error, but training uses a minibatch-specific subset of grid points. Please explain how the theoretical error bound applies to the training rule.

The reviewer is correct that the submitted training rule departed from the analysed estimator, because it integrated each mini-batch over only the time-grid nodes present in that mini-batch. KAPLAN-HR does not require this computational shortcut, so we have removed it in the revised manuscript. The updated results use the full quadrature mesh in every mini-batch.

Under this training rule, each mini-batch loss is the average, over its members, of the same per-subject terms whose full-sample average is the trapezoid criterion analysed in Section 4. A subject’s term depends neither on the other members of its mini-batch nor on the mini-batch size, and every subject enters exactly one mini-batch per epoch, so mini-batching leaves the criterion unchanged. The quadrature bound of Section 4 thus applies to the training rule as stated. Because the parameters are updated after each mini-batch, mini-batching affects the path of the optimiser and the estimator it returns. However, the theory places no restriction on that path. It requires only that the returned estimator satisfy (C5) and (C6), which we measure on the fitted simulation estimators in Appendix D.

We have clarified this in Section 4 (lines 308–314), and updated the training description in Appendix B and Algorithm 1 to the full-grid rule.

Report sensitivity to both grid and batch size.

We have repeated the simulation experiment on all four DGPs with R=5R = 5 repetitions per setting. The quadrature mesh KK is varied around the default K=256K = 256, and the batch size BB around the default B=512B = 512 (Appendix E.3).

Median over repetitions of the paired ISESISE_S ratio, setting ÷ default (B=512B = 512, K=256K = 256) on the same training set, n=65,536n = 65{,}536, R=5R = 5 per setting

SettingRange over DGP-1 to 4Cells whose 95% interval excludes 1
B∈{32,64,128,256}B \in \{32, 64, 128, 256\}0.75–1.101/16
K=32K = 320.88–4.011/4
K=64K = 640.78–1.641/4
K∈{128,512}K \in \{128, 512\}0.89–1.230/8

A coarse mesh raises the error, up to a median of 4.01 times the default's error at the coarsest K=32K = 32. From K=128K = 128 upward the ratios lie between 0.89 and 1.23 across DGPs, with no interval excluding 1. This supports that, from K=128K = 128, the quadrature error does not measurably affect the estimator's error. The batch size has no consistent effect, with ratios between 0.75 and 1.10 and 1/16 intervals excluding 1.

Thank the authors for the responses. Apologize that the "reason to prefer KAPLAN-HR over existing neural survival models" still seems not sufficiently clear to me. Rstriction of proportional-hazards and additive models can be addressed by some of the exisitng survival neural networks methods, e.g. discrete failire time model based deep learning procedures. I am not fully convinced that alternative procedures are necessary based on the current arguments. Could the authors further clarify this?

Reviewer_4TY8OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce KAPLAN-HR, a B-spline KAN for the estimation of the conditional hazard as a function of covariates and time. Theoretical results show that a single-layer KAPLAN-HR recovers B-spline GAM with proportional hazards and the convergence rate for the non-parametric KAN conditional hazard estimator, which does not depend on the dimensionality of the covariates. Experiments are presented on 6 datasets with KAPLAN -HR performing competitively with existing baselines.

Critical strengths

The theoretical results are strong, well justified, demonstrated empirically and contextualized in relation to differences in the implementation.

Critical weaknesses

It is unclear how GAM performs so well in Figure 2 relative to KAPLAN-HR, especially with moderate sample sizes, but is the worst performing model when tested on real-world datasets. Moreover, although the proposed KAPLAN-HR ranks best among the baselines considered, the differences in performance are not significant after accounting for error bars as demonstrated in Table 1, which also suggests that KAPLAN-HR will not rank significantly better than the baselines.

The proposed method enjoys well described theoretical properties, however, its practical benefit compared to existing approaches is not clear. For instance, why not presenting experiments in which the joint conditional hazard formulation demonstrates the advantages of the proposed method?

Questions for authors

Can the authors comment on the restrictiveness of Assumption C4 and how a violation will manifest in an experiment like that of Figure 2?

Consider including a significance test or showing critical differences for the ranks in Figure 3.

Can the authors provide computational cost results for KAPLAN-HR compared with the baselines?

2 comments
AuthorsOpenReview ↗

Response (1/2)

It is unclear how GAM performs so well in Figure 2 relative to KAPLAN-HR, especially with moderate sample sizes, but is the worst performing model when tested on real-world datasets.

We thank the reviewer for pointing out this inconsistency. We have identified and corrected an error in how we invoked mgcv for the GAM benchmark fit. In the revised benchmark, which uses the corrected fit, the GAM's mean rank improves on every metric.

GAM mean rank, submitted and revised benchmark, over 6 datasets

C-TD ↑IBS ↓ICI ↓
Submitted (of 10 models)9.509.177.50
Corrected (of 11 models)5.174.004.83

We have updated all benchmark results with the corrected fit (Table 1, Figure 3, Appendix F).

We have also revised the GAM in the simulations so that its basis grows with nn (kn=4Gn+4k_n = 4G_n + 4, where GnG_n is KAPLAN-HR's number of grid intervals). With this adaptive basis, the GAM attains lower error than KAPLAN-HR at every nn on the additive DGP-1, where it is correctly specified.

Consider including a significance test or showing critical differences for the ranks in Figure 3.

We have added the Friedman test for differences in mean rank across the six datasets to the revised benchmark results (Figure 3), with pairwise comparisons by the Nemenyi critical difference (Figure 11, Appendix F).

Mean rank of KAPLAN-HR among 11 models over 6 datasets, with the Friedman test and Nemenyi critical difference

C-TD ↑IBS ↓ICI ↓
KAPLAN-HR mean rank2.00 (1st)3.00 (2nd)5.50 (tied 4th)
Datasets where KAPLAN-HR attains the best value3/61/61/6
Friedman χF2\chi^2_F (pp)30.2 (8.1×10−48.1 \times 10^{-4})30.6 (6.7×10−46.7 \times 10^{-4})6.6 (0.76)
Significantly worse mean rank than KAPLAN-HR (Nemenyi)CoxPH, DeepSurvCoxPHnone

The Friedman test rejects equal mean ranks on C-TD and IBS, and not on ICI. KAPLAN-HR has the best mean rank on C-TD and the second best on IBS, behind DSM. By the Nemenyi test, KAPLAN-HR's mean rank is significantly better than those of CoxPH and DeepSurv on C-TD and of CoxPH on IBS. No model ranks significantly above KAPLAN-HR on any metric.

The differences in performance are not significant after accounting for error bars as demonstrated in Table 1, which also suggests that KAPLAN-HR will not rank significantly better than the baselines.

The added rank tests support the reviewer's expectation for 27 of the 30 comparisons between KAPLAN-HR and a baseline. The exceptions are CoxPH and DeepSurv on C-TD and CoxPH on IBS, where KAPLAN-HR's mean rank is significantly better. No baseline's mean rank is significantly better than KAPLAN-HR's on any metric (table above). This reconfirms our claim that KAPLAN-HR is competitive with the baselines in empirical performance.

The proposed method enjoys well described theoretical properties, however, its practical benefit compared to existing approaches is not clear.

The practical benefit of KAPLAN-HR comes from its joint hazard formulation, which relaxes the restrictions of proportional-hazards and additive models, and from its KAN parameterisation, which distinguishes it from deep, flexible baselines. To clarify when KAPLAN-HR is a suitable choice:

  • Prefer KAPLAN-HR when the structural form of the hazard is unknown, particularly whether effects vary with time or interact. Proportional-hazards models fix the ordering of subjects' risk over time, and additive models require interactions to be specified in advance.
  • Prefer a simpler model when its form is known to match the data.
  • Against deep, flexible baselines (DSM, DeepHit, CoxTime, SuMo-Net), KAPLAN-HR attains competitive performance. Its advantage therefore lies in the KAN parameterisation itself, which decomposes fitted effects into plottable, inspectable univariate edges (subject to a trade-off with depth) and under which we provide a convergence rate with a dimension-free exponent for a KAN-representable log-hazard.

We have added this guidance to the conclusion (lines 495–502).

AuthorsOpenReview ↗

Response (2/2)

Why not present experiments in which the joint conditional hazard formulation demonstrates the advantages of the proposed method?

Following the reviewer's suggestion, we have revised the simulations to show the structures the joint formulation is intended to represent. Three of the four DGPs now lie outside the additive class. DGP-2 and DGP-3 have two and 32 interacting covariates, and DGP-4 has crossing hazards, i.e. the log-hazard ratio between two values of x1x_1 changes sign at t=1/2t = 1/2, as would occur for example with a treatment effect that reverses during follow-up.

Median over repetitions of the paired ISESISE_S ratio GAM ÷ KAPLAN-HR (same training set), R=20R = 20

n=512n = 512n=8,192n = 8{,}192n=65,536n = 65{,}536
DGP-10.870.540.59
DGP-21.538.7628.18
DGP-32.042.287.62
DGP-41.092.188.54

KAPLAN-HR's error descends at the predicted rate on all four DGPs (on DGP-3 from n=210n = 2^{10}, Figure 2). On DGPs 2 to 4, the GAM's error relative to KAPLAN-HR's grows from 1.09–2.04 times at n=512n = 512 to 7.62–28.18 times at n=65,536n = 65{,}536. On the additive DGP-1, where it is correctly specified, the GAM attains lower error than KAPLAN-HR at every nn.

Can the authors comment on the restrictiveness of Assumption C4 and how a violation will manifest in an experiment like that of Figure 2?

(C4) requires the true log-hazard to admit a KAN representation of fixed depth and width with smooth univariate components, consistent with prior KAN theory [Liu et al., 2025, Thm. 2.1; Liu, Chatzi & Lai, 2025]. This class admits additive effects and, with two or more layers, interactions and time-varying effects built by composing smooth univariate functions, but not every smooth function of (x,t)(x, t). Outside it, Theorem 1 gives no rate.

To show how a violation manifests, we have added an experiment fitting DGP-4 at hidden width 1 (Appendix E.4, p. 31). DGP-4 satisfies (C4) at hidden width Q=3Q = 3 but not at Q=1Q = 1, since a width-1 network keeps the effect of x1x_1 relative to that of x2x_2 fixed over time, while in DGP-4 only the effect of x1x_1 reverses sign at t=1/2t = 1/2.

Median over repetitions of the paired ISESISE_S ratio width 1 ÷ width 3 on DGP-4 (same training set), R=20R = 20, with 95% bootstrap intervals

n=512n = 512n=8,192n = 8{,}192n=65,536n = 65{,}536
Width 1 ÷ width 30.90 [0.79, 1.01]1.14 [1.02, 1.34]3.26 [3.06, 3.90]

The width-1 model attains 0.90 and 1.14 times the width-3 model's error at n=512n = 512 and 8,1928{,}192. At larger nn its error flattens, reaching 3.26 times the width-3 model's error at n=65,536n = 65{,}536, consistent with a remaining approximation error, while the width-3 model continues at the theory's rate (Appendix E.4). A violation of (C4) therefore manifested in this experiment as an error curve that flattens as nn grows.

Can the authors provide computational cost results for KAPLAN-HR compared with the baselines?

We have added the fit and prediction times of KAPLAN-HR and every baseline on the real-data benchmark to Appendix F (Tables 15 and 16). A final fit of KAPLAN-HR takes 194 to 1,459 seconds, and prediction on a test split takes 0.2 to 2.7 seconds.

KAPLAN-HR final-fit time relative to each comparator, range over the 6 datasets (median over splits)

ComparatorKAPLAN-HR ÷ comparator
CoxPH1,080–2,048×
DeepSurv239–439×
CoxKAN10.1–24.0×
MLP-HR (same objective)0.9–3.0×

KAPLAN-HR is up to 2,048× slower to fit than the proportional-hazards baselines. Most of this cost is due to the joint hazard formulation, which requires evaluating the network at every pair of subject and time point. This is supported by MLP-HR, which uses the same joint hazard formulation, and relative to which KAPLAN-HR takes 0.9 to 3.0 times as long. Our implementation partially mitigates this cost by factoring the first-layer evaluation, but it remains a limitation. We have expanded on computational cost in the revised limitations section (lines 510–511).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their feedback. All three reviewers recognised the theoretical analysis as a strength, and two recognised the breadth of the evaluation. Their shared concerns are the practical reason to prefer KAPLAN-HR, the practical gain shown by the benchmark, and the implications of (C4). We address these concerns below, then summarise the changes in the revised manuscript.

Reasons to prefer KAPLAN-HR. KAPLAN-HR's joint hazard formulation relaxes the restrictions of proportional-hazards and additive models. Its KAN parameterisation distinguishes it from deep, flexible baselines.

  • Prefer KAPLAN-HR when the structural form of the hazard is unknown, particularly whether effects vary with time or interact.
  • Prefer a simpler model when its form is known to match the data.
  • Against deep, flexible baselines, KAPLAN-HR attains competitive performance, and its advantage lies in the KAN parameterisation itself. A single-layer KAPLAN-HR is an additive proportional-hazards model, and additional layers relax additivity through composition. One can therefore fit the additive model first and add depth only when it improves the fit. Further, each fitted edge is a univariate function that can be plotted and inspected, with readability trading off against depth. Under this parameterisation, we provide a convergence rate with a dimension-free exponent.

Competitiveness on clinical benchmarks. The benchmark establishes competitive rather than superior predictive performance. The revised benchmark tests differences in mean rank with the Friedman test and Nemenyi critical differences (Figure 3). No model significantly outranks KAPLAN-HR on any metric, while KAPLAN-HR significantly outranks CoxPH on C-TD and IBS and DeepSurv on C-TD. The properties of the KAN parameterisation therefore come at no detectable cost in predictive performance.

Scope of the structural assumption (C4). (C4) requires the true log-hazard to admit a KAN representation of fixed depth and width with smooth univariate components, following prior KAN theory (Liu et al., 2025). This class contains additive effects and, with two or more layers, it also contains interactions and time-varying effects built by composing smooth univariate functions. It does not contain every smooth function of (x,t)(x, t) and, outside it, Theorem 1 gives no rate. We have made the implications of (C4) explicit where it is introduced. We have also added an experiment fitting a hazard that violates (C4) (Appendix E.4). In it, the error curve flattens as nn grows, consistent with a remaining approximation error.

Novelty of the paper. In the discussion, two reviewers inquired about the novelty of the paper. The main novelty lies in the statistical analysis. To our knowledge, this is the first convergence rate for a KAN-based survival model. We instantiate the hypotheses of the sieve rate theorem for a KAN sieve under right-censoring. We also extend the theorem to an estimator whose log-likelihood falls short of the sieve approximant's by at most ηn\eta_n, since stochastic training does not return the exact maximiser. In their initial reviews, the same reviewers noted that the analysis distinguishes the paper from an application of an existing architecture.

Summary of changes (marked in red in the revised paper)

Benchmark

  • Each model is now tuned on each split, with times rescaled by the split's training range, and the full benchmark has been re-run (EUrj).
  • The GAM fit has been corrected (4TY8).
  • MLP-HR replaces the KAN with an MLP under the same objective and tuning budget. It has been added to the benchmark and the simulations (Appendix E.6) (ykXH).
  • The Friedman test and Nemenyi critical differences have been added to Figure 3 and Figure 11 (Appendix F) (4TY8).
  • Fit and prediction times have been added for every model (Appendix F, Tables 15–16) (4TY8).

Simulations

  • The DGPs now include many interacting covariates and crossing hazards, such as a treatment effect that reverses during follow-up (4TY8, ykXH).
  • The GAM now uses an adaptive basis that grows with nn at the same rate as KAPLAN-HR's grid (4TY8, ykXH).
  • An experiment that violates (C4) has been added (Appendix E.4) (4TY8, EUrj, ykXH).
  • Sensitivity to the quadrature mesh and batch size has been added (Appendix E.3) (EUrj).

Theory and text

  • Section 4 now shows that the quadrature bound applies to training, which uses the full mesh in every mini-batch (EUrj).
  • The implications of (C4) are now stated where it is introduced (4TY8, EUrj, ykXH).
  • Section 4 now attributes the dimension-free exponent to (C4) paired with a sieve that mirrors it, and claims no faster rate than competing estimators (EUrj).
  • The related work now compares KAPLAN-HR explicitly with the closest KAN survival models (ykXH).
  • The conclusion now gives guidance on when to prefer KAPLAN-HR (4TY8, EUrj, ykXH).
  • The limitations section now discusses computational cost (4TY8).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.