De-Escalation After Excessive Toxicity

Why Excessive Toxicity Changes the Next Decision

A cohort with multiple dose-limiting toxicities (DLTs) is evidence that the administered dose may be too high. It does not, by itself, establish that every lower dose is safe or determine which dose should be given next. The BLRM uses the cohort together with the prior evidence to update uncertainty about the dose-toxicity curve. The updated posterior then informs whether the current dose and lower candidates satisfy the protocol’s admissibility rule.

The decision is therefore not simply “observe several DLTs, then move down one level.” A protocol may require selecting the next lower admissible dose, or stopping if the required lower dose is not admissible. Those actions depend on the model update and the prespecified decision rule, not only on the cohort’s observed fraction with DLTs.

Model and Decision Rule

Let \(Y_j\) be the number of patients with a DLT among \(n_j\) treated at dose \(d_j\). Conditional on the BLRM parameters, the cohort is modeled as binomial, with \(p(d_j)\) denoting the probability of a DLT at dose \(d_j\). Write the dose-toxicity model as a logistic regression on log dose:

Y_j\mid \alpha,\beta \sim \operatorname{Binomial}\!\left(n_j,p(d_j)\right), \qquad \operatorname{logit}\{p(d)\}=\alpha+\beta\log(d/d_{\mathrm{ref}}).

Here \(d_{\mathrm{ref}}\) is a fixed reference dose, \(\alpha\) is the log-odds of DLT at that reference dose, and \(\beta\) is the dose slope. In the examples, \(\beta>0\), so modeled toxicity probability increases with dose. The model assumes that patients’ DLT outcomes are conditionally independent given the parameters and their dose, and that the DLT definition and assessment window are consistently applied. The binomial likelihood also assumes complete, comparable follow-up for the outcomes being counted.

For observed data \(D\), the posterior distribution is proportional to the prior density times the likelihood. If a new cohort of \(n\) patients at dose \(d\) has \(y\) DLTs, its likelihood contribution is \(\binom{n}{y}p(d)^y\{1-p(d)\}^{n-y}\). The combination coefficient is constant in \((\alpha,\beta)\), so it cancels when normalizing posterior probabilities. Multiple DLTs increase the likelihood of parameter values that predict higher toxicity at the administered dose, but they do not turn the posterior into a point estimate: uncertainty remains.

Definition: For a protocol-defined toxicity boundary \(p_{\mathrm{tox}}\), a dose \(d\) is an overdose under a BLRM parameter value if \(p(d)>p_{\mathrm{tox}}\). Its posterior overdose probability is \(P\{p(d)>p_{\mathrm{tox}}\mid D\}\). Under a rule with cutoff \(\xi\), the dose is admissible when this posterior probability is no greater than \(\xi\). The toxicity boundary and cutoff are protocol choices, not universal constants.

After updating, calculate admissibility using the updated posterior at each dose relevant to the protocol. If the current dose fails the rule, a de-escalation algorithm might select the next lower admissible dose. If the required lower dose is not admissible, the protocol may direct the team to stop rather than treat at that dose. The exact sequence, including which lower candidates may be considered, must be specified in advance.

Why the Cohort Changes the Posterior

The update can be seen directly from Bayes’ rule. Suppose the current posterior density is \(\pi(\alpha,\beta\mid D)\). After observing \(y\) DLTs among \(n\) patients at dose \(d\), the new density is proportional to the old density multiplied by that cohort’s likelihood:

\pi(\alpha,\beta\mid D,y,n,d) \propto p(d;\alpha,\beta)^y\{1-p(d;\alpha,\beta)\}^{n-y} \pi(\alpha,\beta\mid D).

A parameter value receives more relative posterior weight when the observed count was more likely under that value. For example, with \(n=3\), the likelihood contribution for two DLTs is proportional to \(p(d)^2\{1-p(d)\}\). A value predicting \(p(d)=0.60\) has contribution \(0.60^2(0.40)=0.144\), while a value predicting \(p(d)=0.20\) has contribution \(0.20^2(0.80)=0.032\). The former receives \(0.144/0.032=4.5\) times the relative likelihood contribution of the latter. This comparison updates the full parameter distribution; it is not a direct estimate of the toxicity probability at other doses.

Worked Example: Two DLTs at the Current Dose

Consider a hypothetical trial with a finite-grid approximation to the current BLRM posterior. It assigns equal prior weights to three parameter values, all with positive slope. At 40 mg, their predicted DLT probabilities are \(0.20\), \(0.40\), and \(0.60\). The first two slopes are \(\beta=1\); the third is \(\beta=4\). Suppose three patients are treated at 40 mg and two have a DLT. Let the protocol define overdose as \(p(d)>0.30\), use \(\xi=0.25\), and consider 20 mg as the next lower dose.

Update the grid weights. Before observing the cohort, each parameter value has weight \(1/3\), so the unnormalized posterior weights are proportional to the likelihoods:

\begin{aligned} L_1&=0.20^2(1-0.20)=0.032,\\ L_2&=0.40^2(1-0.40)=0.096,\\ L_3&=0.60^2(1-0.60)=0.144,\\ L_{\mathrm{total}}&=0.032+0.096+0.144=0.272. \end{aligned}

The normalized posterior weights are \(0.032/0.272=0.1176\), \(0.096/0.272=0.3529\), and \(0.144/0.272=0.5294\). At 40 mg, the second and third parameter values exceed the toxicity boundary. Thus the posterior overdose probability is \((0.096+0.144)/0.272=0.8824\), which is greater than \(0.25\). The 40 mg dose is not admissible.

Check the lower dose. With the stated logistic model, moving from 40 mg to 20 mg subtracts \(\beta\log 2\) from the log odds. The predicted probabilities at 20 mg are approximately \(0.1111\), \(0.2500\), and \(0.0857\). Each is below \(0.30\), so the posterior overdose probability at 20 mg is \(0\), and 20 mg is admissible. If the protocol calls for the next lower admissible dose, the decision is to de-escalate to 20 mg. The observed fraction with DLT at 40 mg was \(2/3\), but the decision follows from the updated posterior and the rule, not from treating that fraction as the risk at 20 mg.

How the Number of DLTs Affects the Update

For the same cohort size and dose, the likelihood depends on the DLT count. A larger count generally gives relatively more weight to parameter values that predict higher toxicity at that dose. It does not follow that every lower dose becomes inadmissible: the dose slope affects how predictions change across doses, and posterior uncertainty about the slope matters. Conversely, observing fewer DLTs than in a more severe cohort does not guarantee that the current dose passes the overdose rule.

Worked Example: One DLT Versus Two

Use the same hypothetical three-point posterior at 40 mg, with equal prior weights and predicted probabilities \(0.20\), \(0.40\), and \(0.60\). Suppose instead that one of three patients has a DLT. For \(y=1\), the likelihood contributions, omitting the common binomial coefficient, are \(p(1-p)^2\):

\begin{aligned} L_1&=0.20(0.80)^2=0.128,\\ L_2&=0.40(0.60)^2=0.144,\\ L_3&=0.60(0.40)^2=0.096,\\ L_{\mathrm{total}}&=0.128+0.144+0.096=0.368. \end{aligned}

The posterior weights are \(0.128/0.368=0.3478\), \(0.144/0.368=0.3913\), and \(0.096/0.368=0.2609\). The overdose event at 40 mg is still represented by the second and third parameter values, so its posterior probability is \((0.144+0.096)/0.368=0.6522\). Under the same illustrative cutoff \(\xi=0.25\), 40 mg is not admissible. The posterior overdose probability is lower than the \(0.8824\) obtained after two DLTs, but it still exceeds the protocol cutoff.

This calculation shows why an observed DLT fraction is not itself an admissibility test. The fraction \(1/3\) is a cohort summary; the posterior overdose probability also reflects the prior, the likelihood, and the event defined by the model at the dose.

When De-Escalation Becomes Stopping

A protocol should specify what to do if the current dose fails the overdose rule but no permitted lower dose passes it. One possible rule is to stop further enrollment at that point. This is not a mathematical necessity of the BLRM; it is a safety action defined by the protocol. A decision rule should not be improvised after seeing the cohort, since changing the candidate set or cutoff in response to the observed toxicity can alter the operating characteristics of the design.

Worked Example: Lower Dose Also Fails the Rule

Consider a second hypothetical grid of three equally weighted BLRM parameter values. At 20 mg, the predicted DLT probabilities are \(0.35\), \(0.45\), and \(0.55\); at 40 mg they are \(0.60\), \(0.75\), and \(0.90\). These pairs are consistent with positive logistic dose slopes: for each value, \(\beta=\{\operatorname{logit}(p(40))-\operatorname{logit}(p(20))\}/\log 2>0\). Suppose three patients receive 40 mg and all three have DLTs.

The likelihood contributions at 40 mg are \(0.60^3=0.216\), \(0.75^3=0.421875\), and \(0.90^3=0.729\), with total \(1.366875\). Normalizing gives posterior weights of approximately \(0.1580\), \(0.3086\), and \(0.5334\). Each parameter value predicts \(p(40)>0.30\), so the posterior overdose probability at 40 mg is \(1\). Each also predicts \(p(20)>0.30\), so the posterior overdose probability at 20 mg is likewise \(1\). With \(\xi=0.25\), neither dose is admissible. If the protocol says to stop when the required lower dose is not admissible, the appropriate action is to stop rather than treat at 20 mg. This example applies that specified rule; it does not imply that every trial should use this exact stopping criterion.

Assumptions, Diagnostics, and Alternatives

The update is meaningful only if the model and data represent the protocol’s intended quantities. In particular, DLTs must be defined consistently, outcomes must refer to a comparable assessment window, and the dose assigned to each patient must be recorded correctly. If some patients have not completed the DLT assessment window, treating them as DLT-free can distort the likelihood; the protocol’s incomplete-follow-up handling rule must be used instead. The model also assumes that the dose-response form is adequate over the candidate dose range and that the prior gives plausible weight to clinically credible curves.

Useful checks include confirming the cohort size and DLT count against the clinical data, verifying that the posterior weights normalize to one, and plotting or tabulating posterior toxicity probabilities and overdose probabilities across the doses being considered. Inspect whether the posterior is supported by the observed data or remains strongly prior-dependent, especially when the cohort is small. A surprising posterior pattern may signal a coding error, an incorrect dose transformation, an incorrectly applied threshold, or a poorly fitting model. These checks do not replace clinical review of the DLTs or the protocol’s decision rule.

A simple count-based rule might mandate de-escalation whenever a specified number of patients experience DLTs. It is transparent and can be easy to implement, but it does not directly quantify posterior risk at each candidate dose. The BLRM-based rule instead uses the modeled posterior probability of exceeding a toxicity boundary, incorporating uncertainty and information across doses. Neither approach removes the need for prespecification, clinical judgment, or monitoring; a model-based rule can be misleading if its assumptions or prior are unsuitable.

Common Pitfalls

Using the cohort fraction as the overdose probability. The observed fraction \(2/3\) at 40 mg is not the posterior probability that the toxicity probability exceeds \(0.30\). Compute the latter from the updated posterior, as in the first worked example.
Assuming that toxicity at the current dose determines risk below it. The lower-dose probabilities depend on the dose slope as well as the current-dose predictions. Recalculate admissibility at the lower candidate dose.
Moving down automatically without checking admissibility. A lower dose is not automatically acceptable. Apply the protocol’s overdose cutoff and stopping rule before recommending treatment there.
Changing the rule after seeing the data. Choosing a different cutoff or dose set only after a severe cohort compromises the prespecified decision process. Use the protocol’s rule and document the result.

Key takeaway: After excessive toxicity, update the BLRM with the cohort likelihood, calculate posterior overdose probabilities at the current and protocol-permitted lower doses, and apply the prespecified de-escalation or stopping rule. A DLT count is evidence for the update, not a substitute for it.

Check Your Understanding

Answer each question using the definitions and decision logic above. Solutions follow the questions.

In a cohort of four patients, three have DLTs. What likelihood contribution, apart from the binomial coefficient, applies to a parameter value predicting DLT probability \(p\) at the cohort dose?
At one candidate dose, posterior parameter weights are \(0.20\), \(0.30\), and \(0.50\). The first parameter value predicts \(p(d)>p_{\mathrm{tox}}\), while the other two do not. What is the posterior overdose probability?
A dose has posterior overdose probability \(0.22\), and the protocol cutoff is \(\xi=0.20\). Is it admissible under the stated rule?
Why can a cohort with multiple DLTs justify reassessing a lower dose without proving that the lower dose is admissible?

Solutions. (1) The likelihood contribution is \(p^3(1-p)\), because \(y=3\) and \(n-y=1\). (2) The posterior overdose probability is the weight on parameter values satisfying the overdose event, \(0.20\). (3) No. The rule requires the probability to be no greater than \(0.20\), and \(0.22>0.20\). (4) The model’s lower-dose toxicity probabilities depend on the posterior over both the intercept and slope. The lower dose must be evaluated against the overdose boundary and cutoff; the observed count at the higher dose does not perform that calculation.