Decision Objectives and Constraints for Regimen Optimization

From a Dose Choice to a Decision Problem

Selecting a regimen is not just identifying the dose with the largest estimated efficacy. A clinical decision also has boundaries: a regimen may need to achieve a minimum level of benefit, keep an unacceptable safety risk below a specified limit, and be deliverable in practice. Among regimens that meet those requirements, the decision-maker still needs a way to express which tradeoffs are preferable.

A decision framework makes these elements explicit. It states what can be chosen, what makes a choice acceptable, and how to compare acceptable choices. This does not determine clinical preferences by itself. It makes the preferences and requirements visible enough to scrutinize, communicate, and revise when evidence or the decision context changes.

In the simplest form, the decision is to choose a regimen with the greatest utility among feasible options:

Here, \(g_j(x)\leq 0\) represents the \(j\)th required condition. If multiple candidates achieve the same maximum utility, the result is a set of tied optimal candidates rather than a unique choice. If \(\mathcal{F}\) is empty, no candidate meets all stated requirements; the utility function cannot make an unacceptable candidate feasible.

Separate Requirements From Preferences

An efficacy constraint might require a predicted response measure to exceed a minimum, or the probability of meaningful benefit to exceed a threshold. A safety constraint might limit the probability of a serious adverse event or a clinically important toxicity burden. Feasibility constraints can describe practical requirements, such as an acceptable administration schedule or availability in the intended setting. Each condition should specify the clinical measure, the population and time horizon to which it applies, and the threshold that makes it acceptable.

A utility function serves a different role. It ranks consequences rather than simply marking them acceptable or unacceptable. It can encode the value placed on efficacy relative to treatment burden or adverse outcomes. Its scale may be clinical, preference-based, or an explicitly defined score; the units and interpretation should be stated. A utility score is not automatically a probability, and a larger numerical value is meaningful only under the chosen scoring rule.

A useful distinction is that a safety limit can act as a hard boundary while the utility function also reflects preference for lower risk within the acceptable region. That is not inherently contradictory: the boundary excludes regimens considered unacceptable, while the utility distinguishes among those that remain. But decision-makers should check that they have not inadvertently counted the same consequence twice or imposed a penalty so large that the stated constraint has no practical role.

A Short Derivation: Expected Utility Across Outcomes

When a regimen can lead to several mutually exclusive clinical outcomes, expected utility gives a simple way to combine their values. Suppose the possible outcomes are \(k=1,\ldots,K\), with probability \(p_k(x)\) under regimen \(x\), and utility \(u_k\) assigned to outcome \(k\). The probabilities sum to one. The expected utility follows by weighting each outcome's utility by its probability:

This is the expectation of the outcome-utility variable: each possible utility \(u_k\) occurs with probability \(p_k(x)\). It provides a single preference-based summary when the utilities are defined on a common scale. It does not imply that the outcomes are equally important, nor does it determine the utilities. Those values must reflect the decision context and be justified independently of the arithmetic.

In an application, probabilities may be estimated or predicted from evidence and a specified model. The resulting decision therefore depends on both the preference specification and the evidence used to characterize outcomes. A calculated optimum should not be presented as certain if those inputs are uncertain.

Worked Example: Expected Utility Among Acceptable Regimens

Consider two hypothetical regimens with four possible outcomes: response without a serious event, response with a serious event, no response without a serious event, and no response with a serious event. The outcome utilities are 100, 20, 30, and 0 points, respectively. The decision requirements are response probability of at least 0.60 and serious-event probability no greater than 0.10.

For regimen A, the four outcome probabilities are 0.60, 0.05, 0.30, and 0.05. They sum to \(0.60+0.05+0.30+0.05=1.00\). Response probability is \(0.60+0.05=0.65\), and serious-event probability is \(0.05+0.05=0.10\), so A meets both requirements. Its expected utility is:

For regimen B, the probabilities are 0.68, 0.04, 0.23, and 0.05. They also sum to 1.00. Its response probability is \(0.68+0.04=0.72\), and its serious-event probability is \(0.04+0.05=0.09\). B meets both requirements. Its expected utility is:

Both regimens are feasible, and B has the higher expected utility under the assigned outcome values, so B is preferred by this framework. The event limit remains an explicit acceptability requirement; the outcome utilities also express the relative value assigned to the possible clinical outcomes. The numerical ranking depends on these hypothetical values and is not a universal clinical preference.

Worked Example: When Practical Requirements Change the Choice

Suppose a hypothetical program has three candidate regimens. The intended setting can support no more than one administration per day. The team requires a meaningful-benefit probability of at least 0.70 and a serious-toxicity probability no greater than 0.08. A prespecified utility score, already calculated using the program's stated preferences, is shown below. The numbers are illustrative and do not represent a real study.

Candidate A meets the benefit threshold, toxicity limit, and administration limit, so it is feasible. B meets the clinical thresholds but requires two administrations per day, exceeding the setting's limit; it is infeasible. C meets the benefit and administration requirements, but \(0.09>0.08\), so it is infeasible on safety grounds. Consequently, A is selected, despite B having the largest utility score in the table. If the practical administration limit or safety threshold were changed, the feasible set could change and the decision would need to be recalculated.

Choosing and Checking the Decision Framework

The structure of the decision should match the actual clinical question. An efficacy-only rule ranks regimens by benefit but can select one that violates a safety or delivery requirement. A single weighted score can represent tradeoffs, but may conceal which outcomes drove the ranking and can allow a very high score on one dimension to compensate for a clinically unacceptable value on another. A constrained utility approach makes minimum requirements explicit and uses preferences only to rank the candidates that meet them.

If stakeholders cannot defend a common utility scale, it can be more transparent to state separate criteria and a prespecified priority order rather than invent precise weights. Different approaches answer different questions: a utility function asks which feasible outcome profile is preferred under stated values; constraints ask whether a regimen crosses an acceptability boundary. Neither supplies those values or boundaries automatically.

Before interpreting a recommendation, check that the decision variable and candidate set are clearly defined; each constraint uses the right measure, population, and time horizon; and the utility scale has a stated meaning. Confirm that all candidate values use compatible definitions and that the feasible set is not empty. A threshold that is unsupported or a utility weight that is arbitrary can create a numerically precise but clinically unconvincing answer.

Also examine what happens near a constraint boundary. If small plausible changes in a predicted risk move the selected regimen from feasible to infeasible, the decision is sensitive to that input. Report that sensitivity rather than describing the point estimate as a guarantee. Conversely, if several candidates comfortably satisfy constraints but their utility scores are nearly tied, the evidence or preference information may not support a strong preference for one over another.

A common mistake is to select the regimen with the largest utility score first and treat safety and practicality as secondary considerations. That reasoning confuses ranking with eligibility. The correction is to define the constraints, form \(\mathcal{F}\), and compare utility only among candidates in \(\mathcal{F}\). Another mistake is to treat an arbitrary utility scale as if its numerical units were self-evident. State how scores were assigned and what preferences they represent.