acceptodds
Under review as a conference paper at ICLR 2027

Learning the Maximum Tolerated Dose for Continuous Toxicity via Monotone Bayesian Trees

Abstract

Phase I cancer trials seek the maximum tolerated dose (MTD) while protecting patients from excessive toxicity. Dose assignments must therefore balance patient safety with learning the dose–toxicity relationship as data accrue. We model continuously measured toxicity outcomes using two forms of Bayesian additive regression trees (BART): isotonic BART projects posterior response curves onto nondecreasing functions, whereas monotone BART constrains the model. Joint curve and variance draws induce an MTD posterior that guides dose selection through escalation with overdose control (EWOC). We compare these methods with three parametric procedures across seven dose–toxicity curves in simulation. We assess dose-limiting toxicity (DLT) counts, above-MTD assignments, signed last-dose error, and relative absolute error. The tree procedures attained the lowest mean RAE on four nonlinear curves and jointly minimized mean DLT counts and above-MTD assignments on four curves. A Bayesian reinforcement learning perspective formulates these sequential decisions as a finite-horizon planning problem. Dose restrictions yield a lower bound on last-dose error; under exact posterior-predictive evaluation, an EWOC-based rollout policy has no greater expected weighted loss than its baseline. Dose Trial Lab, a desktop simulator for all five procedures, accompanies the supplementary material.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Probabilistic methods, Bayesian Dose Finding, Reinforcement Learning, Adaptive Sampling, Sequential Decision-Making, Maximum Tolerated Dose

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_isfbOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper tackles the problem setting of learning the maximum tolerated dose in phase 1 cancer trials with continuous toxicity measures. For this the authors use two versions of Bayesian additive regression trees (BART) to also allow for uncertainty estimates of the dose-response function. Further, the authors formulate the problem setting then also as a Bayesian Reinforcement Learning setup. The proposed implementations are then benchmarked against three baselines and over different dose-response curves on simulated data.

Critical strengths
  • The problem setup of learning the maximum tolerated dose in cancer doese escalation trails seems like an important application in the medical field. Using here Bayesian additive regression trees for the dose response modelling while also getting uncertainty estimates seems like a reasonable approach.
  • The paper provides Dose Trial Lab, a software for visualizing the problem setup (I did not run the software), which might help if you are unfamiliar with the field.
Critical weaknesses
  • The overall notation and problem setup is not entirely clear. E.g., Y(x) is not defined (does this imply the potential outcomes framework, a random variable, some arbitrary function?). In Eq.(1) the modelling choice of Eq. (1) e.g. the plausibility of a Gaussian distribution in this setting and implications should be explained properly. In iBART in Sec. 2.2, hj,gk,Bkh_j, g_k, B_k etc are not defined in the paper. These are examples but similar missing definitions and explanations carry throughout the whole paper.
  • The writing style in large parts of the paper, in particular sec 2 and 3 is very hard to understand, the logical flow within and between sentences and paragraphs is often very hard to grasp. Even the motivation of the connection between section 2 and 3 is not clear to me after several reads. Then, the experiments section fails directly connect to the previous sections. The long RL Sec. 3 seems to be not evaluated in the main paper and only very limitedly in the appendix. Overall, also the experiments section reads more like an automatically generated report and does not follow a clear narrative.
  • The methodological contribution seems very limited for an ICLR paper. Further, the clear benefit and necessity of the approach is not properly elaborated in the paper. This shows also in the novelty description in the introduction and in table 2 which reads more like a combination of methods and results, without showing the purpose or benefits of the choices. I think, overall this paper might be more suitable for a more tailored Biostatistics or ML for helathcare outlet, than for ICLR.
Questions for authors

What is the actual gain of the BART models over the 3PND baseline?

2 comments
AuthorsOpenReview ↗

Point-by-point responses by authors (1/2)

First of all, thank you so much for reading our paper carefully. Your question about what the trees gain over 3PND helped us see which comparison needed to be clearer.

We have uploaded the revised manuscript for your review. We hope you will take a look, and we would welcome the opportunity to discuss the new results with you.


Q. What is the actual gain of the BART models over 3PND?

A. (Table 1 and Section 4.3, lines 445–465 and 469–483 of the revised manuscript.) The clearest gain was on the piecewise-linear curve, where both trees estimated the maximum tolerated dose (MTD) more accurately and exposed fewer patients to dose-limiting toxicities (DLTs). Under the same EWOC dose rule, the mean relative absolute error (RAE) of the final dose was 9.97% with the power-law 3PND model of Lee (2023), 3.02% with iBART, and 3.00% with mBART. The paired mBART-minus-3PND difference was −6.97 percentage points (95% Monte Carlo interval [−7.00, −6.94]). The mean DLT count also fell from 3.86 per trial with 3PND to 2.02 with either tree. Both trees improved accuracy and reduced DLTs on the exponential curve as well. The gain depended on curve shape. In the original experiment, 3PND estimated the MTD more accurately than either tree on the quadratic and square-root curves.

We also doubled the residual standard deviation from 0.1 to 0.2 while keeping the true MTD at 50. Across 50 paired trials per curve, both trees had lower mean RAE than 3PND on the piecewise-linear, square-root, exponential, and logarithmic curves. All eight pointwise 95% Monte Carlo intervals for these differences lay below zero. On the exponential curve, both trees also reduced DLTs and above-MTD assignments. On the square-root and logarithmic curves, the gain in accuracy came with more DLTs (Figure 3).

The higher-variability experiment also included the linear dose finders 1PLD and 2PLD described by Lee et al. (2022). Both trees had lower mean RAE than either linear model on all six nonlinear curves. On the square-root curve, for example, mean RAE was 21.59% with 1PLD and 34.70% with 2PLD, compared with 9.99% with iBART and 9.49% with mBART. On the linear curve itself, 1PLD had the lowest mean RAE (2.28%). Figure 4 shows the paired comparisons with all three parametric models.

References

Lee, S. Y. (2023). A flexible dose-response modeling framework based on continuous toxicity outcomes in phase I cancer clinical trials. Trials, 24, 745. Article

Lee, S. Y., Munafo, A., Girard, P., & Goteti, K. (2022). Optimization of dose selection using multiple surrogates of toxicity as a continuous variable in phase I cancer trial. Contemporary Clinical Trials, 113, 106657. Article


Q. What does Y(x)Y(x) denote, why use a Gaussian likelihood, and what do the tree symbols mean?

A. (Sections 2.1–2.2, lines 92–94, 121–126, and 130–140 of the revised manuscript.) Y(x)Y(x) denotes a patient’s random, continuously measured toxicity response at assigned dose xx. The trees model how mean toxicity changes with dose. As in continuous-response BART (Chipman et al., 2010) and the continuous-toxicity dose-finding model of Lee (2023), we use Gaussian errors with a common variance across doses. This is a working assumption for approximately symmetric residuals with comparable spread across dose levels. It allows us to use both the fitted mean and the residual spread to estimate the probability that a patient’s response reaches or exceeds the dose-limiting toxicity (DLT) threshold.

Section 2.2 now defines the tree notation. The function hjh_j gives the contribution of tree jj to the mean response. For iBART, BkB_k is the kkth dose interval, and gkg_k is the value of an unconstrained ensemble draw on that interval. We project these values onto a nondecreasing sequence so that the estimated mean toxicity can stay level or increase as dose increases.

Some of the notations were originally introduced and defined in the Appendix to accommodate the conference’s strict page limit. In the revised manuscript, we have added these definitions to the main text since we have one more page allowed. We thank the reviewer for this helpful suggestion, which has improved the clarity of the paper.

References

Chipman, H. A., George, E. I., & McCulloch, R. E. (2010). BART: Bayesian additive regression trees. The Annals of Applied Statistics, 4(1), 266–298. Article

Lee, S. Y. (2023). A flexible dose-response modeling framework based on continuous toxicity outcomes in phase I cancer clinical trials. Trials, 24, 745. Article

AuthorsOpenReview ↗

Point-by-point responses by authors (2/2)

Q. How do Sections 2 and 3 connect, and what do the experiments evaluate?

A. (Sections 2–4, especially lines 209–216 and 263–265, and Appendix I, Table 4, of the revised manuscript.) Section 2 uses a lower quantile of the maximum tolerated dose (MTD) posterior to propose the next dose under escalation with overdose control (EWOC). Section 3 asks how giving that dose would affect later decisions. Two dose–toxicity curves can have the same MTD yet predict different responses for the next cohort. Rollout therefore uses the joint posterior of the curve and residual variance to simulate those responses, update the posterior, and determine what dose EWOC would propose next.

The experiments vary one component at a time. Section 4 holds the EWOC rule fixed and compares all five response models. Appendix I holds the response model fixed at 1PLD and compares EWOC with rollout. In that linear scenario, with an escalation cap of 7 dose units, rollout reduced the observed mean weighted loss by 55.8%–80.9% relative to EWOC across three loss profiles and two planning budgets (Table 4, lines 2155–2159).

We used 1PLD for this policy comparison because both its posterior update and its EWOC quantile have closed forms. Within rollout, Dose Trial Lab uses particle resample–move updates for 2PLD and 3PND and warm-started MCMC updates for iBART and mBART. Independent paths can run in parallel, but each dose decision within a path depends on the preceding cohort’s responses. Appendix I explains this computational cost in lines 2232–2243.


Q. What is the methodological contribution for a machine learning (ML) audience?

A. (Introduction, lines 76–82, Section 3, Proposition 1, and Appendix A.3, lines 866–880 of the revised manuscript.) Dose finding can be viewed as a sequential learning problem in ML because each dose we assign determines where we collect the next observation. We appreciate your point that this problem is also of interest to a biostatistics audience. Indeed, several papers on dose finding and personalized dosing have appeared at leading ML conferences. We added Appendix A.3 to discuss seven examples from ICML, AISTATS, NeurIPS, and ICLR, listed below.

Our contribution to the ML audience is to show how Bayesian trees can guide dose selection when we know only that mean toxicity is nondecreasing with dose. After each cohort, we use the continuous toxicity measurements to update the curve and its residual variation. Together, these give us a posterior distribution for the maximum tolerated dose (MTD), which we use to choose the next dose under overdose control. We also show how the escalation cap and the number of cohorts remaining limit how close the final dose can get to the MTD (Proposition 1).

References

Katz-Samuels, J., & Scott, C. (2018). Feasible arm identification. Proceedings of the 35th International Conference on Machine Learning, 80, 2535–2543. Article

Shen, C., Wang, Z., Villar, S., & van der Schaar, M. (2020). Learning for dose allocation in adaptive clinical trials with safety constraints. Proceedings of the 37th International Conference on Machine Learning, 119, 8730–8740. Article

Chien, I., Bruinsma, W. P., Gonzalez, J., & Turner, R. E. (2024). Safe exploration in dose finding clinical trials with heterogeneous participants. Proceedings of the 41st International Conference on Machine Learning, 235, 8436–8467. Article

Lee, H.-S., Shen, C., Jordon, J., & van der Schaar, M. (2020). Contextual constrained learning for dose-finding clinical trials. Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics, 108, 2645–2654. Article

Lee, H.-S., Shen, C., Zame, W., Lee, J.-W., & van der Schaar, M. (2021). SDF-Bayes: Cautious optimism in safe dose-finding clinical trials with drug combinations and heterogeneous patient groups. Proceedings of the 24th International Conference on Artificial Intelligence and Statistics, 130, 2980–2988. Article

Demirel, I., Celik, A. A., & Tekin, C. (2022). ESCADA: Efficient safety and context aware dose allocation for precision medicine. Advances in Neural Information Processing Systems, 35, 27441–27454. Article

Yao, M. S., Bastani, O., Andersson, A., Biancalani, T., BenTaieb, A., & Iriondo, C. (2026). Knowledgeable language models as black-box optimizers for personalized medicine. Proceedings of the 14th International Conference on Learning Representations. Article

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

This paper studies the dose-finding problem in early-phase oncology trials using continuously measured toxicity outcomes. The authors consider two Bayesian additive regression tree models, isotonic BART (iBART) and monotone BART (mBART), to estimate the dose–toxicity relationship. They combine these models with escalation with overdose control (EWOC) for sequential dose assignment. The authors further formulate dose finding as a finite-horizon Bayesian reinforcement learning problem and establish a policy improvement result showing that, under exact posterior-predictive evaluation, the proposed rollout policy achieves no greater expected weighted loss than the baseline EWOC policy. Simulation studies compare the BART approaches with three parametric dose-finding procedures.

Critical strengths
  1. Dose finding problem in oncology trials is an important and interesting problem.
  2. Bayesian framework enables natural ways to borrow historical information in estimating dose-toxicity relationship.
  3. The formulation of dose finding as a sequential decision-making problem is potentially interesting, as current dose assignments affect both patient safety and the information available for future decisions.
Critical weaknesses
  1. Limited sample sizes and robustness of nonparametric estimation. Early-phase oncology trials typically enroll relatively few patients, raising concerns about whether flexible BART models can reliably estimate dose–toxicity relationships. Although the authors consider monotonicity constraints, the numerical experiments primarily use a fixed sample size and residual variance. It remains unclear how the proposed methods perform under smaller sample sizes, greater outcome variability, or model and prior misspecification.

  2. Insufficient comparisons with established dose-finding designs. The numerical studies primarily compare the proposed approaches with three parametric methods. Comparisons with commonly used clinical dose-finding designs, such as 3+3 and BOIN, would help establish their practical value. Since these methods generally use discrete dose levels and binary toxicity outcomes, appropriate adjustments would be needed to ensure fair comparisons.

  3. Limited assessment of prior sensitivity. With small sample sizes, Bayesian dose-finding procedures may be particularly sensitive to prior specifications. The manuscript does not sufficiently investigate how different BART priors and induced MTD posteriors affect dose assignments, estimation accuracy, and patient safety.

  4. Clinical limitations of toxicity-only dose optimization. The maximum tolerated dose may differ from the recommended Phase II dose, especially for immunotherapies and targeted therapies, where efficacy may not increase monotonically with dose. Although the manuscript explicitly focuses on MTD estimation, the authors should discuss the applicability and limitations of the proposed framework in modern oncology trials where dose optimization often requires balancing toxicity and efficacy.

Questions for authors

Q1. Could authors compare the proposed methods with other dose finding algorithms, e.g., 3+3 and BOIN?

Q2. How sensitive are the results to the choice of BART prior distributions, including the tree-depth prior, shrinkage parameters, and residual variance prior? How do differences in the induced MTD posteriors affect the resulting allocation policies?

Q3. Could the authors discuss how their framework might be extended to incorporate efficacy outcomes, particularly in settings where the optimal biological dose or recommended Phase II dose may differ from the MTD?

5 comments
AuthorsOpenReview ↗

Point-by-point responses by authors (1/2)

First of all, thank you so much for reading our paper carefully. Your questions about 3+3 and BOIN helped us clarify why we chose the five comparators. Your question about the BART priors also led us to run a new experiment examining how much a change in the prior can affect the dose assigned to the next patient.

We have uploaded the revised manuscript for your review. We hope you will take a look, and we would welcome the opportunity to discuss the new results with you.


Q. Could the authors compare the proposed methods with other dose-finding designs, such as 3+3 and BOIN?

A. (Table 2, Appendix A.6, lines 945–952, and Figure 6 of the revised manuscript.) We agree that 3+3 and BOIN are important clinical benchmarks. Both use a binary toxicity outcome, recording whether each patient experienced a dose-limiting toxicity (DLT). Our methods use the continuous measurement itself. Figure 6 illustrates why this matters. At doses 15 and 30, both cohorts had zero DLTs among three patients. Yet the measurements at dose 30 were much closer to the DLT threshold. Our models use that information when estimating risk at the next dose.

In our simulations, all five methods observed continuous toxicity and used the same EWOC allocation rule. This allowed us to examine what changes when we change the dose–toxicity model. A direct comparison with 3+3 or BOIN would answer a further practical question. We would use the same underlying patient responses, dose grid, starting dose, enrollment limit, and true MTD. Each design would then use its own observations, allocation rule, and stopping rule. BOIN can use the prespecified target DLT probability, while conventional 3+3 follows fixed DLT-count rules. We added both designs to the discussion and clarified how such a comparison could be conducted. Their relative performance remains an empirical question.

References

Storer, B. E. (1989). Design and analysis of phase I clinical trials. Biometrics, 45(3), 925–937. Article

Liu, S., & Yuan, Y. (2015). Bayesian optimal interval designs for phase I clinical trials. Journal of the Royal Statistical Society: Series C (Applied Statistics), 64(3), 507–523. Article


Q. How sensitive are the MTD estimates and dose assignments to the tree-depth, leaf-shrinkage, and residual-variance priors?

A. (Appendix C.4, Figure 8, lines 1190–1201 and 1242–1252 of the revised manuscript.) To answer your question, we held the patients’ doses and responses fixed and changed one prior setting at a time. We used 20 trial histories for each of seven curves, with 36 patients per history. For each tree model, we halved the base splitting probability, increased leaf shrinkage from 2 to 3, or raised the prior 90th percentile of the residual standard deviation from 0.20 to 0.40. Including the baseline fits, this gave 1,120 fits.

The prior could materially affect the MTD estimate, even when it had a smaller effect on the assigned dose. On the linear curve, for example, halving the splitting probability for mBART increased the mean absolute error of the posterior median MTD by 12.74 dose units. It raised the mean EWOC proposal by 3.53 units, but the mean assigned dose rose by only 0.41 units. The reason is that EWOC uses the fifth percentile of the MTD posterior, and the escalation cap and no-decrease rule confined the next assignment to doses between 44.5 and 48.

Across the 840 comparisons with the baseline prior, the EWOC proposal changed 653 times and the assigned dose changed 213 times. Across the 42 combinations of curve, model, and prior change, the mean change in the next patient’s DLT probability ranged from −0.17 to +0.78 percentage points. These results show how the prior affected the next decision after the same 36 observations. A prior change earlier in a trial could also change which doses and responses are observed later.

AuthorsOpenReview ↗

Point-by-point responses by authors (2/2)

Q. How do the tree methods perform with fewer patients or greater outcome variability?

A. (Section 3, lines 355–357, Section 4.3, lines 469–483, and Appendices J.1 and L of the revised manuscript.) With fewer patients, the escalation cap can determine how close any method can get to the MTD. We examined the first 30 patients in the original simulated trials. Starting at dose 6, with cohorts of three and a maximum increase of 3.5 units per cohort, the highest reachable dose by patient 30 was 37.5. Both tree procedures reached 37.5 in all 700 trials per procedure. Because the true MTD was 50, even the highest permitted final dose had a relative absolute error of 25%. In this setting, additional enrollment matters because it gives the design more opportunities to reach the target.

We also ran 1,750 new trials with the residual standard deviation doubled from 0.1 to 0.2. We kept the true MTD at 50 and used 50 paired replications per curve with 45 patients per trial. Both trees had lower mean final-dose relative absolute error than 3PND on the piecewise-linear, square-root, exponential, and logarithmic curves. All eight pointwise 95% Monte Carlo intervals for these differences lay below zero. On the piecewise-linear curve, mean error was 9.30% with 3PND, 2.96% with iBART, and 3.05% with mBART. On the exponential curve, the trees also reduced DLTs and above-MTD assignments. On the square-root and logarithmic curves, the gain in accuracy came with more DLTs. This experiment addresses greater variability under a common Gaussian error model. Performance under heavy-tailed errors or dose-dependent variance remains open.


Q. How could the framework incorporate efficacy when the optimal biological dose or recommended phase II dose differs from the MTD?

A. (Section 5, lines 532–535 of the revised manuscript.) We agree that this distinction matters, particularly for immunotherapies and targeted therapies. Our present target is the MTD, defined through the probability of toxicity. If efficacy plateaus while toxicity continues to rise, a dose below the MTD may offer patients a better balance of benefit and risk.

We added a paragraph describing a phase I/II extension. We would record both efficacy and continuous toxicity, retain the monotonicity assumption for mean toxicity, and allow efficacy to plateau or decrease with dose. Among doses meeting the toxicity criterion, we would recommend the dose with the highest posterior expected benefit–risk utility. This follows the clinical reasoning of Thall and Cook (2004) and gives a concrete way to extend the present framework when efficacy is also observed.

Reference

Thall, P. F., & Cook, J. D. (2004). Dose-finding based on efficacy–toxicity trade-offs. Biometrics, 60(3), 684–693. Article

I would like to thank authors for addressing my comments. However, given the current presentation, it is difficult for me to move it to a substantially higher score; and thus, I would retain my score.

Reviewer_QLjPOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This study combines monotone Bayesian trees, Bayesian dose allocation, and rollout planning to determine the maximum tolerated dose using continuous toxicity data. Theoretically, it proves conditional MTD-posterior consistency, establishes a protocol-specific lower bound for final-dose errors, and provides an exact Bayesian policy-improvement guarantee for rollout. Additionally, the authors include comparative simulations and a desktop simulator to help users test the introduced methods.

Critical strengths
  1. This work offers a flexible modeling for the MTD problem as monotone Bayesian trees can represent nonlinear dose-toxicity relationships without being forced into a fixed shape. Moreover, this model provides a clear way to measure uncertainty by jointly inferring the response curve and the noise variance.

  2. The simulation results are promising, as the tree methods achieved the lowest mean estimation error on four of the seven response curves, while iBART successfully minimized both toxicities and unsafe allocations on four curves to clearly illustrate the trade-off between safety and accuracy.

  3. The framework explicitly handles sequential decisions by factoring in dose limits, stopping rules, and patient budgets, utilizing a reachability bound to show how early assignments restrict final accuracy while mathematically guaranteeing improvement over the baseline.

Critical weaknesses
  1. The theoretical guarantees do not fully align with the actual implementation, as the math assumes persistent sampling and exact expectations, whereas the executed policy does not guarantee that sampling and relies on approximations instead.

  2. The central combination of Bayesian trees with rollout planning is missing true experimental validation because the rollout experiments only tested a basic linear model, leaving it completely unclear if the proposed tree models are actually beneficial or computationally practical.

  3. The empirical scope is too narrow to draw general conclusions, as the tests locked in specific variables like noise level and budget while only comparing against three basic procedures, meaning broader testing against more flexible competitors is needed to see if the gains actually hold up.

Questions for authors
  1. Does the proposed allocation policy satisfy the persistent cell-sampling assumption required for posterior consistency? Can you clarify what guarantees remain valid with approximate posterior inference and finite-budget rollout?

  2. Can you provide rollout results using iBART or mBART, including computational costs, and assess robustness to different noise levels or trial settings? Comparisons with a flexible alternative would also help establish the benefits beyond the chosen parametric baselines.

If these concerns are satisfactorily addressed, I will raise my score.

5 comments
AuthorsOpenReview ↗

Point-by-point responses by authors (1/2)

First of all, thank you so much for reading our paper carefully. Your questions helped us clarify which guarantees apply to the dose rule used in our simulations and which depend on additional assumptions.

We have uploaded the revised manuscript for your review. We hope you will take a look, and we would welcome the opportunity to discuss these results with you.


Q. Does the proposed allocation policy satisfy the persistent cell-sampling assumption required for posterior consistency?

A. (Section 2 and Appendix F, Theorem 3, of the revised manuscript.) No. The theorem requires the fraction of observations in each fixed dose cell to stay bounded away from zero as enrollment grows. Under the Only Escalation rule, once the trial moves above a cell, it cannot return to that cell. The implemented policy therefore does not guarantee persistent sampling. We agree that this distinction needed to be explicit, and we have clarified it in the revision.

Theorem 3 establishes that the MTD posterior concentrates at the true MTD under the stated sampling, model, prior-support, and threshold-separation assumptions. For the implemented policy with 45 patients, we assess estimation accuracy and safety through the simulations in Section 4. One theoretical result does apply directly to that policy. Proposition 1 bounds the final doses that remain reachable from the current dose, given the escalation cap and the number of cohorts left. This bound holds for every realized dose path, including when posterior inference is approximate.


Q. What guarantee remains with approximate posterior inference and finite-budget rollout?

A. (Section 3, Proposition 2, and Appendix H.6–H.7, Proposition 6.) Proposition 2 compares rollout with EWOC under exact posterior-predictive evaluation. At each decision, the candidate set includes the dose that EWOC would assign. Selecting the candidate with the lowest exact expected continuation loss therefore gives rollout a Bayes expected weighted loss no greater than that of EWOC.

With a finite number of simulated paths, error in the estimated losses can change which dose is selected. Proposition 6 quantifies the consequence. If every candidate’s estimated continuation loss is within ϵ\epsilon of its exact value at every relevant state, and the legal candidate sets are preserved, the excess expected loss relative to EWOC is at most 2Hϵ2H\epsilon with HH decisions remaining. We do not establish a numerical value of ϵ\epsilon for our MCMC implementation. Thus, the finite-budget policy can have higher loss than EWOC, and we evaluate its performance empirically.

AuthorsOpenReview ↗

Point-by-point responses by authors (2/2)

Q. Why is rollout evaluated with 1PLD, and how expensive would evaluation be for the other models?

A. (Appendix I, Table 4 and Figure 13 of the revised manuscript.) We focused the rollout experiment on 1PLD because its posterior update and EWOC quantile have closed forms. Rollout repeatedly simulates future cohorts, updates the posterior, and evaluates subsequent dose assignments. An inexpensive underlying procedure therefore makes extensive simulation feasible. With 1PLD, we evaluated 1,000 and 2,000 simulated paths per candidate. In the linear scenario with an escalation cap of 7, rollout reduced observed mean weighted loss by 55.8%–80.9% across the three loss profiles and two planning budgets.

Our implementations of 2PLD, 3PND, iBART, and mBART require MCMC posterior updates within these simulated paths. Independent paths can run in parallel, but successive assignments within each path depend on the preceding cohort’s responses. Figure 13(a) bounds the number of simulated cohorts per trial, and Figure 13(b) gives the sampler-sweep ceiling if every update uses a full posterior refit. At 2,000 paths per candidate, the design permits up to 1.05 million simulated cohorts per trial. Using the full posterior-refitting settings at every update gives upper bounds of 142.8 billion sampler sweeps for either 2PLD or 3PND, 35.7 billion for iBART, and 10.5 billion for mBART. Extensive replicated evaluation at these settings may require large-scale high-performance computing. Instead, the desktop simulator provides rollout options for all five models at user-selected budgets, while the manuscript’s rollout experiment uses the computationally inexpensive 1PLD procedure.

The computational work of repeated rollout satisfies

Tplan=O(B∑k=1KAk(1+Hk)cm)=O(BAK2cm).T_{\mathrm{plan}} = O\left(B\sum_{k=1}^{K} A_k(1+H_k)c_m\right) = O(BAK^2c_m).

BB is the number of simulated paths per candidate, Ak≤AA_k \leq A is the number of candidate doses at stage kk, KK is the number of decision stages, and Hk≤K−kH_k \leq K-k is the number of subsequent cohorts. The quantity cmc_m bounds the cost of processing one simulated cohort with model mm, including posterior updating and dose selection.

With B=2,000B=2{,}000, A=5A=5, and K=14K=14, the number of simulated cohorts per trial is at most

BA∑h=1Kh=BAK(K+1)2=1,050,000.BA\sum_{h=1}^{K}h = \frac{BAK(K+1)}{2} = 1{,}050{,}000.

The workload ceiling is 1.05 million simulated cohorts per trial, with the model-specific computation repeated throughout these paths.

For an illustrative desktop calculation, suppose a complete simulated-cohort computation takes 1 millisecond with 1PLD and 1 second with 2PLD. At the ceiling of 1.05 million simulated cohorts, ten CPU workers with ideal parallel scaling would require approximately 1.75 minutes per 1PLD trial and 29.2 hours per 2PLD trial. For either iBART or mBART, a complete per-cohort computation taking 10 seconds would imply approximately 12.2 days per trial, increasing to 72.9 days if it takes 60 seconds. These are conditional calculations using assumed computation costs. Actual runtimes depend on the CPU, sampler settings, data, and parallelization overhead. Independent paths can run in parallel, while successive posterior updates within each path remain sequential. Extensive replicated evaluation with full MCMC refits may therefore require large-scale computing resources. The desktop simulator provides rollout for all five models at user-selected budgets, and our extensive rollout experiment uses the closed-form 1PLD procedure.


Q. How robust are the tree results to higher outcome variability, and how do they compare with other flexible models?

A. (Section 4.3, Figures 3–4, and Appendix L of the revised manuscript.) We added an experiment that doubled the residual standard deviation from 0.1 to 0.2 and recalibrated all seven curves to retain the true MTD of 50. All five models used the same EWOC rule and 45-patient budget, with 50 paired replications per curve.

On the piecewise-linear curve, mean last-dose relative absolute error fell from 9.30% with 3PND to 2.96% with iBART and 3.05% with mBART. Both trees also had lower mean error than 3PND on the square-root, exponential, and logarithmic curves. All eight pointwise 95% Monte Carlo intervals for these four curves favored the trees. On the exponential curve, both trees also reduced DLTs and above-MTD assignments. On the square-root and logarithmic curves, improved accuracy came with more DLTs. Both trees had lower mean error than the linear models, 1PLD and 2PLD, on all six nonlinear curves.

We appreciate your suggestion to include another flexible model. A comparison with a monotone Gaussian process using the same continuous toxicity measurements, EWOC rule, and patient budget would help assess the contribution of the tree representation. Our experiments establish gains over the linear and power-law dose finders in the settings described above. Performance relative to another nonparametric model remains an open question.

AuthorsOpenReview ↗

Follow-up on our rebuttal and revised manuscript

Dear Reviewer QLjP,

Thank you for your thoughtful review and constructive suggestions. We appreciate your willingness to raise your score if your concerns are satisfactorily addressed.

We would be grateful if you could review our rebuttal and revisions and consider raising your score if you find that they adequately resolve your concerns. If any issues remain, we would welcome your feedback and the opportunity to clarify them before the discussion deadline.

Thank you again for your time and consideration.

Sincerely, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Author before the end of Author- Reviewer Discussion

I appreciate anonymous reviewers, AC, SAC, and PC for the interest in the paper. It was a great lesson to learn how some of recent trends of ML/AI conference, rebuttal/response procedure, etc, are going on. Thank you a lot.

Regards, Author

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.