Learning the Maximum Tolerated Dose for Continuous Toxicity via Monotone Bayesian Trees
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Learning the Maximum Tolerated Dose for Continuous Toxicity via Monotone Bayesian Trees".
Jane: Here's the gist of the paper. Phase I cancer trials aim to determine the maximum tolerated dose (MTD) while minimizing patient toxicity risks, utilizing continuous toxicity measurements.
Tom: The study employs two forms of Bayesian additive regression trees: isotonic BART (iBART), which projects posterior response curves onto nondecreasing functions, and monotone BART (mBART), which constrains the model itself.
Jane: These methods are used to model continuously measured toxicity outcomes within a Bayesian framework to guide dose selection through escalation with overdose control (EWOC).
Tom: The analysis compares three parametric procedures: one-parameter linear dose finder (1PLD), two-parameter linear dose finder (2PLD), and three-parameter nonlinear dose finder (3PND).
Paper summary: Tom: So we're looking at a paper called "Learning the Maximum Tolerated Dose for Continuous Toxicity via Monotone Bayesian Trees." This study looks at how to find the maximum tolerated dose while trying to keep patient toxicity risks low as more data comes in.
Jane: Exactly. The main idea is that they use these Bayesian additive regression trees, specifically two versions of them, called isotonic BART and monotone BART, to model those continuously measured toxicity outcomes. They're trying to figure out the MTD as the data accumulates.
Lu: What's really interesting is that they don't just pick one method; they compare these tree procedures against three different ways of finding a dose across seven different kinds of dose-toxicity curves in simulation, like linear or exponential ones.
Meng: So, what does this mean for someone who is actually trying to run a trial? Is this just theoretical math that won't work when the data gets messy?
Lalam: Well, the paper shows that these tree procedures do something pretty strong; isotonic BART did the best job minimizing the observed mean dose-limiting toxicity counts and above-MTD assignments on four of those seven curves.
Tom: That's a big win for iBART, but I want to know about precision too. How accurate are we when we actually pick a dose? Does it just give us a good guess or is the error small enough for real patient decisions?
Jane: They also looked at precision using signed last-dose error and relative absolute error, and they found that the one-parameter linear dose finder performed best on the linear curve, and three parameter nonlinear dose finder did well on the quadratic and square root curves.
Lu: The paper also explores a way to make decisions sequentially under a Bayesian reinforcement learning view, which sets a lower bound on how much error you can expect for that last dose estimate.
Meng: A lower bound is good, but what about the actual trial rollout? If you're actually running this in the field, how does this change your daily decision-making compared to just using an escalation with overdose control strategy?
Lalam: They found that under exact posterior predictive evaluation, a rollout policy based on an overdose control strategy performed no better than the baseline strategy for expected weighted loss.
Tom: So it sounds like the theoretical setup is solid, and even when you look at sequential decision-making under those specific conditions, the rollout policy isn't worse than what we already do. What about how much improvement we actually see in practice?
Jane: They did show that rollout attained lower observed mean weighted loss than overdose control across every single loss profile and planning budget, with reductions ranging from four point one percent to twenty-five point nine percent.
Paper summary: Lu: That range of reduction sounds significant when you think about how much you can shave off the cost or risk in a real trial setting. The paper also concludes that flexible MTD estimation works even in small-sample settings without needing to pre-specify the shape of the dose-toxicity curve.
Meng: I'm skeptical about that claim about small sample sizes; my models break when you have real traffic noise, not just clean simulated data. What happens if the true toxicity curve is really complex and we only have a few patients?
Tom: That's where the paper gets its nuance; it says they established MTD posterior consistency under the condition of a fixed-cell Gaussian model and prior support at the truth. They are making claims based on specific assumptions about what those models look like.
Jane: And they did have a separate comparison using an O6-benzylguanine example, which used a pharmacodynamic response instead of clinical toxicity, to explore different models like 1PLD and two tree models <ref:2609.30190#pg1>.
Lu: That comparison showed that while those methods assign terminal doses in the eighty to one hundred milligrams per square meter region consistent with earlier analyses, they can't compute metrics like above-MTD assignments or relative absolute error against a known target because the true boundary is unknown.
Meng: So, even when using different response types, if you don't know the actual toxicity limit beforehand, you can't really measure how far off your estimate is from the truth. That’s a big limitation for deployment.
Tom: It seems like these methods are very useful for setting up the initial dose selection process and guiding escalation with overdose control, even if we have to be careful about what we assume about the underlying curve shape.
Jane: The title of this paper, "Learning the Maximum Tolerated Dose for Continuous Toxicity via Monotone Bayesian Trees," really captures that balance between learning a limit and managing risk over time.
Lu: It's less about finding one perfect number and more about building a flexible framework that adapts as you gather more data on continuous toxicity outcomes.
Meng: For production deployment, the focus has to be on what breaks when things aren't perfectly Gaussian or when the true boundary is just not known. We need methods that can handle that uncertainty without requiring us to know the exact shape of everything upfront.
Lalam: The paper suggests that flexible MTD estimation remains practical in small sample settings because it doesn't require prespecification of the curve shape, which is a key part of how these tree procedures operate.
Tom: So, while iBART looks strong on the simulated curves for minimizing toxicity counts and above-MTD assignments, we have to keep an eye on those precision metrics like signed last-dose error.
Jane: Right. And remember that the analysis is done under specific model assumptions, so when translating this to a real trial with unknown dynamics, we have to consider those constraints carefully.
Conclusion: Tom: So we've been looking at how these monotone trees help us find the maximum tolerated dose when toxicity measurements are continuous and keep tracking them over time.
Jane: Right, so they’re using these Bayesian trees to model those toxicity outcomes—like how bad a side effect is—as the data comes in, trying to figure out the safest dose.
Lu: Basically, they're using this method to guide dose escalation with overdose control while minimizing patient risk by looking at where the uncertainty falls.
Meng: So what’s the actual result? Does this mean we can just pick one number and be done with it for a trial?
Lalam: The main thing is that these tree procedures do a strong job of finding the maximum tolerated dose across different types of curves, even when you don't know the exact shape of that curve beforehand.
Tom: It sounds like they’ve built a flexible way to handle uncertainty in dosing decisions.
Jane: They did compare different methods, like linear finders and these tree models, on seven totally different kinds of toxicity shapes.
Lu: The paper shows that the monotone BART method tied with isotonic BART for one of those piecewise linear curves, while other methods were better for other curve types.
Meng: I’m still thinking about how this works in a real trial setting where you have to make a quick decision about the next dose.
Tom: Well, they also looked at sequential decision making under a Bayesian view, and they found that the rollout policy actually performed as well as the standard overdose control strategy in terms of expected weighted loss.
Jane: That's interesting because it suggests this isn't just some theoretical setup; it’s practical for how we manage patients during a trial rollout.
Lu: They conclude that flexible maximum tolerated dose estimation is practical even when you don't know the exact shape of the toxicity curve from day one.
Tom: It seems like this opens up a new way to think about how we find safety limits in complex dosing scenarios.
stat.AP
Submitted: 2026-09-24
Updated: 2026-11-13
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Phase I cancer trials aim to determine the maximum tolerated dose (MTD) while minimizing patient toxicity risks, utilizing continuous toxicity measurements.
Key concepts
- Bayesian Additive Regression Trees (BART)
- These are statistical models used to predict continuous outcomes, like toxicity levels. They work by building a collection of decision trees that combine predictions from individual trees. In this study, two specialized versions were used: isotonic BART (iBART) and monotone BART (mBART), which help control the shape of the predicted results.
- Maximum Tolerated Dose (MTD)
- The MTD is the highest dose a patient can safely receive without experiencing unacceptable toxicity. The study aimed to find this value using continuous measurements of toxicity. They used these models to guide dose escalation, ensuring patients received a safe and effective dose while minimizing harm.
- Isotonic BART (iBART) and Monotone BART (mBART)
- These are two specific types of Bayesian trees used for modeling toxicity curves. iBART projects predictions onto nondecreasing functions, meaning the predicted toxicity levels will always increase as the dose increases. mBART constrains the model itself to ensure it remains monotone, which is crucial when modeling dose-response relationships.
- Dose-Limiting Toxicity (DLT) and EWOC
- DLT refers to toxicity that limits a patient's ability to continue treatment. The study used a cumulative DLT stopping rule: enrollment ends if the DLT count exceeds four or 45 patients are observed. This decision-making process, combined with an escalation strategy called Overdose Control with Escalation (EWOC), helps determine the final safe dose.
Terminology
Summary
Phase I cancer trials aim to determine the maximum tolerated dose (MTD) while minimizing patient toxicity risks, utilizing continuous toxicity measurements. p. 0 LEARNING THE MAXIMUM TOLERATED DOSE
FOR CONTINUOUS TOXICITY VIA
MONOTONE BAYESIAN TREES
Se Yoon Lee
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. 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. 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). Given the small sample sizes of phase I cancer trials, the dose-finding literature has typically
relied on parametric models. Our comparators are the one-parameter linear dose finder (1PLD),
which estimates a slope; the two-parameter linear dose finder (2PLD), which also estimates residual
variance; and the three-parameter nonlinear dose finder (3PND), which estimates scale, variance,
and a curvature exponent (Eichhorn & Zacks, 1973; Lee et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The comparison involved seven different dose–toxicity curves: linear, piecewise linear, quadratic (convex), square root (concave), sigmoid, exponential (convex), and logarithmic (concave). p. 7 4 EXPERIMENTS
4. 1 DESIGN AND PERFORMANCE CRITERIA
The study compares 1PLD, 2PLD, 3PND, iBART, and mBART across seven dose–toxicity curves:
linear, piecewise linear, quadratic, square root, sigmoid, exponential, and logarithmic. MTD posterior consistency. Appendix F proves consistency of the MTD posterior under a fixedcell Gaussian model, prior support at the truth, persistent cell sampling, and a strict risk-threshold
margin. NPD and NPO assess safety performance. The precision criteria BTM and RAE describe MTD
estimation error: BTM retains its direction, whereas RAE measures its relative magnitude without
cancellation.Where the paper says it
Where the paper says it
Where the paper says it
The tree procedures demonstrated strong performance: iBART jointly minimized the observed mean NPD and NPO on four of the seven curves, while mBART tied with iBART for the piecewise linear curve. p. 8 16
mBARTp 0. 19
For safety, Table 1 shows that iBART jointly minimized the observed mean NPD and NPO on four of
seven curves, including a tie with mBART for the piecewise linear curve. The sigmoid curve instead favored 3PND on safety. For precision, the lowest observed mean RAE was attained by 1PLD on the linear curve, 3PND on
the quadratic and square-root curves, iBART on the exponential curve, and mBART on the piecewise
linear, sigmoid, and logarithmic curves.Where the paper says it
Where the paper says it
In terms of sequential decision-making under a Bayesian reinforcement learning perspective, dose restrictions provide a lower bound for last-dose error. p. 0 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. 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. The final cohort is fully observed and its posterior update is completed before
checking the stopping rule. Under the numerical-study protocol, N = 45, c = 3, and KN = 4:
five or more cumulative DLTs trigger the count rule after the cohort, and enrollment also ends at 45
patients. The following finite-horizon Bellman recursion holds
when measurable ε-optimal selectors exist for the admissible action sets:
V (s) = R(b, x), D > KN or n = N,
V (s) = inf
a∈A(s)
wOm ob(a) + Z
wDd(y) + V (s
′
)
Pb(dy | a)
, otherwise. (3)
Its terminal estimate is the dose actually administered to the last cohort, including after a DLT stop. 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. Our
design uses the available structure: dose is the sole predictor, both Bayesian tree models enforce
monotonicity, and controlled EWOC limits escalation. Across seven dose–toxicity curves with 45
patients, these models balance safety and MTD precision without prespecifying curve shape, showing
that flexible MTD estimation remains practical in this small-sample setting.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
A separate comparison using the O6-benzylguanine example (Lee, 2023) explored five models, including 1PLD and two tree models, using a pharmacodynamic response rather than clinical toxicity. p. 8 Across seven dose–toxicity curves with 45
patients, these models balance safety and MTD precision without prespecifying curve shape, showing
that flexible MTD estimation remains practical in this small-sample setting. We also applied all five
methods to the O6
-benzylguanine example to estimate an AGT-based dose target. We also applied all five
methods to the O6
-benzylguanine example to estimate an AGT-based dose target. Their final doses
of 80–100 mg/m2
are broadly consistent with the 100 mg/m2 biochemical modulatory dose reported
by Friedman et al.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Bayesian learning of dose-toxicity relationships can be improved by using monotone Bayesian regression trees (BART) to model continuously measured toxicity outcomes, ensuring that
the sum f is then nondecreasing
for mBART, which enforces a clinical requirement thattoxicity not decrease with dose.
-
Dose allocation can be refined by employing a Bayesian reinforcement learning formulation where the state includes
the joint posterior, last administered dose, enrollment, and cumulative DLT count,
allowing for planning decisions based on sequential learning under protocol constraints. -
Sequential decision-making can be optimized by using an EWOC-based rollout policy that minimizes the
posterior expected asymmetric dose loss
or a more comprehensive loss function combiningDLT counts, above-MTD assignments, signed last-dose error, and relative absolute error.
-
The system can achieve MTD posterior consistency by ensuring that the posterior distribution of the MTD is consistent at its true value under prior support, persistent cell sampling, and a strict risk-threshold margin.
-
The AI system can be designed to handle uncertainty in both dose response shape and residual variability by using iBART, which projects
posterior draws onto nondecreasing functions,
while retaining the paired residual standard deviation for MTD inference. -
The system can incorporate protocol-specific constraints like
the upward cap and Only Escalation restriction
into the planning action set to ensure that the proposed dose respects these limits, resulting in a policy whereeven when the posterior alpha-quantile suggests a substantially higher dose, the increase for the next patient or cohort cannot exceed delta max.
-
The system can perform exact expected-loss rollout improvement by evaluating candidate doses through hypothetical trials and updating its posterior after every simulated cohort until stopping, as described in
Algorithm 3,
which providesexact posterior primitives and independent continuations.
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.
Sources
- Approximate Dynamic Programming and Its Applications to the Design of Phase I Cancer Trials
- Incorporating Individual and Collective Ethics into Phase I Cancer Trial Designs
- Dose Finding with Escalation with Overdose Control (EWOC) in Cancer Clinical Trials
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data