Why a Safety Limit Is Different From a Preference
A dosing policy may balance expected benefit against expected harm, but some clinical requirements are not preferences to be balanced. If a dose must not exceed a specified toxicity-risk criterion, a recommendation that violates that criterion is unacceptable under the rule, even if it has greater expected utility. The central design choice is therefore not only which dose has the greatest utility? It is which dose has the greatest utility among those that satisfy every non-negotiable requirement?
Earlier material on expected utility and asymmetric loss provides ways to summarize clinical consequences and compare tradeoffs. A hard safety constraint plays a different role: it defines which actions are allowable before utility is optimized. The numerical utility assigned to an action cannot compensate for its failure to meet a hard constraint.
The inequality notation is general. For example, a maximum acceptable predicted toxicity risk \(r_{\max}\) gives the constraint \(R(d\mid x)-r_{\max}\le 0\), where \(R(d\mid x)\) is the model-based probability of toxicity for action \(d\) and patient information \(x\). A minimum safety-related requirement could instead be written with the sign reversed. The clinical meaning and direction of each constraint must be stated explicitly.
Assumptions Before Optimizing
The constrained rule is straightforward mathematically, but its clinical interpretation depends on how it is specified. At minimum, the following assumptions should be made explicit:
- The action set is defined. The candidate actions are the dose choices being compared, and each has a well-defined utility and safety assessment for the same patient context.
- The limits are clinically justified. The constraint and its cutoff come from an applicable clinical or regulatory requirement, not from choosing a convenient value after seeing which dose wins.
- The safety quantity is clearly defined. For example, a predicted toxicity probability needs a defined event, time horizon, and patient context. A probability over one time horizon cannot silently stand in for risk over another.
- The predictive model is relevant to the patient and decision. A constraint based on model predictions is only as credible as those predictions, their calibration, and the data and assumptions supporting them.
- All mandatory conditions apply together. Feasibility means satisfying every specified hard constraint, not most of them or a favorable average across them.
A risk threshold applied to a model prediction is not a guarantee that an individual will avoid toxicity. It is a rule about the model-based risk quantity. If the model is misspecified, poorly calibrated for the patient group, or evaluated over the wrong time horizon, a recommendation can satisfy the mathematical constraint while failing to provide the intended clinical protection.
Worked Example: Choosing the Best Feasible Dose
Suppose, hypothetically, that three candidate doses have the following expected utilities and predicted probabilities of a defined toxicity event over the same treatment period. The clinical rule requires predicted toxicity risk no greater than 0.08.
| Dose (mg) | Expected utility (points) | Predicted toxicity risk |
|---|---|---|
| 25 | 6.2 | 0.02 |
| 50 | 8.4 | 0.06 |
| 75 | 9.1 | 0.14 |
Write the safety condition as \(R(d\mid x)\le 0.08\). At 25 mg, \(0.02\le0.08\), so the dose is feasible. At 50 mg, \(0.06\le0.08\), so it is also feasible. At 75 mg, \(0.14>0.08\), so that action is infeasible.
The feasible utilities are 6.2 and 8.4 points. Therefore, $$ \max_{d\in\mathcal{F}(x)}U(d\mid x)=\max(6.2,8.4)=8.4 \text{ points}, $$ and the constrained recommendation is 50 mg. The unconstrained utility maximum would be 75 mg, with utility 9.1 points, but it violates the specified safety criterion. Its higher utility does not make it an allowable recommendation under this rule.
Why the Constrained Rule Works
For a finite set of candidate actions, the justification follows directly from the definition of constrained choice. First remove every action that violates at least one hard constraint. The remaining actions form \(\mathcal{F}(x)\). Comparing their utilities then identifies the best action that meets all requirements. If an action outside \(\mathcal{F}(x)\) has greater utility, it still cannot be the constrained optimum because it is not an eligible candidate.
This also clarifies a crucial edge case. If no action satisfies every condition, then \(\mathcal{F}(x)=\varnothing\), and the constrained maximization has no solution. It is incorrect to describe the least-violating dose as “the constrained optimum”: that dose is not feasible. Infeasibility is information about the decision problem or its inputs, not permission to ignore one of the limits without authorization.
Worked Example: Two Mandatory Criteria Leave No Feasible Action
Consider a hypothetical set of three options. The clinical requirements are a predicted toxicity risk no greater than 0.03 and a predicted probability of achieving a specified minimum treatment target of at least 0.70. The second requirement is included here as a mandatory treatment criterion; it is not a substitute for the toxicity constraint.
| Dose (mg) | Predicted toxicity risk | Predicted target-attainment probability | Utility (points) |
|---|---|---|---|
| 25 | 0.01 | 0.45 | 5.0 |
| 50 | 0.04 | 0.78 | 7.5 |
| 75 | 0.10 | 0.90 | 8.8 |
For 25 mg, the risk criterion is met because \(0.01\le0.03\), but the target criterion fails because \(0.45<0.70\). For 50 mg, the target criterion is met, but the risk criterion fails because \(0.04>0.03\). At 75 mg, target attainment is adequate, but \(0.10>0.03\), so the risk criterion also fails.
Thus, no row satisfies both conditions: $$ \mathcal{F}(x)=\varnothing. $$ The utilities do not change this conclusion. In particular, selecting 50 mg because it has the highest utility among the options that meet the target criterion would silently discard the toxicity requirement. A sound decision system should report that these stated requirements yield no feasible candidate, so the situation can be handled through an authorized clinical process rather than presented as a valid constrained recommendation.
Hard Constraints Versus Soft Penalties
A soft penalty subtracts a cost for violating a requirement and then maximizes the resulting score. One possible form is \(U(d\mid x)-\lambda V(d,x)\), where \(V(d,x)\) measures the amount of violation and \(\lambda\) converts that amount into utility units. This can be useful when the underlying concern is genuinely a tradeoff. It is not equivalent by definition to a hard constraint: with a finite penalty, an action that violates the limit can still win.
A short calculation shows why. Suppose a safe action has utility 7 points, while a higher-utility action has utility 10 points and exceeds a risk limit by 0.03. Under a penalty rule, the higher-utility action is selected whenever \(10-0.03\lambda>7\), or \(\lambda<100\). The selection therefore depends on the chosen penalty scale. Under a hard constraint, the violating action is excluded regardless of its utility or the size of its advantage.
Worked Example: A Penalty Can Preserve an Unacceptable Action
Suppose, hypothetically, the maximum acceptable predicted risk is 0.05. A dose with predicted risk 0.08 has utility 10 points; a dose with risk 0.04 has utility 7 points. Define violation as the amount by which risk exceeds the limit, \(V=\max(0,R-0.05)\). The higher-utility dose has \(V=0.08-0.05=0.03\).
With \(\lambda=50\) points per unit of risk, its penalized score is $$ 10-(50)(0.03)=10-1.5=8.5 \text{ points}. $$ Since \(8.5>7\), the penalty rule selects the dose with risk 0.08, even though it violates the stated limit. With \(\lambda=120\), its score is \(10-(120)(0.03)=6.4\) points, and the other dose wins with 7 points. The change in decision comes from changing the penalty weight, not from any change in the patient or clinical limit.
Under the hard rule \(R\le0.05\), the dose with risk 0.08 is simply infeasible, and the 0.04-risk dose is selected from the feasible options. This is the appropriate mathematical representation when the requirement is genuinely non-negotiable.
Diagnostics, Limitations, and Common Mistakes
A constrained recommendation should be checked for both mathematical and clinical failure modes. Useful diagnostics include verifying the units and time horizon of each safety quantity; confirming that every reported recommendation satisfies every constraint; inspecting how close recommendations lie to a limit; and checking whether the feasible set is often empty or contains only one action. Frequent boundary selections can signal a real clinical tradeoff, but they can also expose an overly restrictive or poorly calibrated model. These checks do not justify changing a limit after observing inconvenient results; any revision requires clinical justification and transparent governance.
Model error is a particularly important limitation. If a model underpredicts toxicity in a subgroup, a rule based on its predicted risk may label unsafe actions feasible. Calibration of predicted toxicity is relevant to whether the risk scale is trustworthy, but calibration alone does not establish that a chosen cutoff is clinically appropriate or that every individual outcome is safe. The model’s target population, prediction horizon, and uncertainty should therefore be considered when the constraint is defined and evaluated.
Common mistake: “The higher dose violates the safety limit, but its utility gain is large enough to justify it.” That reasoning describes a soft tradeoff, not a hard constraint. The correction is to exclude the violating action from the feasible set. If no feasible action remains, report infeasibility rather than covertly replacing the rule with a penalty or choosing the smallest violation.
Another mistake is treating “predicted risk below the cutoff” as proof that toxicity cannot occur. A probability criterion limits a model-based risk quantity; it does not prevent stochastic events, remove model uncertainty, or certify performance for patients unlike those used to develop the model. A hard rule is only as meaningful as its clinical definition and its supporting evidence.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A dose has expected utility 12 points but violates one hard safety constraint. Another feasible dose has utility 9 points. Which dose is selected by the constrained rule, and why?
- A predicted risk is 0.045 and the specified maximum is 0.05. Is the action feasible with respect to that constraint? What does feasibility not establish?
- Two mandatory constraints leave no candidate action that satisfies both. Is the highest-utility candidate among those satisfying only one constraint a constrained optimum?
- In a penalty rule, a violating action has utility 11 points, violation \(V=0.02\), and \(\lambda=100\) points per unit of violation. A feasible alternative has utility 9 points. Which action does the penalty rule select, and what does this show?
Solutions. 1. Select the feasible dose with utility 9 points. The 12-point action is excluded by the hard constraint, so its utility cannot offset the violation. 2. Yes, because \(0.045\le0.05\). This establishes only that the model-based risk meets the stated criterion; it does not guarantee that the patient will avoid toxicity or that the model is correct. 3. No. The feasible set is empty, so there is no constrained optimum; selecting that candidate would disregard a mandatory condition. 4. The penalized score is \(11-(100)(0.02)=9\) points, equal to the alternative’s 9 points. The penalty rule produces a tie, which requires a prespecified tie rule; it has not automatically excluded the violating action. That contrasts with a hard constraint, under which the violating action is ineligible.