Formulation and Dose-Rounding Constraints

From a Model-Based Dose to a Regimen That Can Be Given

An optimization may identify a dose that performs well according to an efficacy–toxicity objective, yet that dose may not be deliverable using the available products. The recommended amount might fall between tablet strengths, require more units per administration than is practical, exceed a maximum dose, or violate a required dose increment. These are not cosmetic details: they determine which regimens can actually be considered.

The task here is to include formulation and administration constraints in the choice itself. The introductory material on multiobjective optimization establishes why a regimen may be assessed using a utility that balances several outcomes. Here, write that utility as \(U(r)\), where \(r\) is a regimen, and restrict the decision to regimens that can be made and administered under the specified constraints. The utility’s definition and the clinical constraints must be specified before comparing candidates.

For example, if a single administration may contain \(k_j\) units of strength \(s_j\), its dose is \(d=\sum_j k_j s_j\), where each \(k_j\) is a nonnegative integer. A unit limit, a maximum dose, or a rule against splitting a tablet further restricts the allowed combinations. If there are multiple administrations, the constraints must describe the amount at each administration as well as any limit on the total daily dose or number of administrations. A dose expressed in milligrams is not, by itself, a complete description of an implementable regimen.

The constraints should reflect the actual product and intended use. In particular, do not assume that a tablet can be split, that two formulations can be combined, or that an available strength can be used by every route. Those permissions need to be established for the product and setting under consideration. If a proposed dose increment is required, the feasible doses must satisfy it; the presence of a maximum dose alone does not establish which intermediate doses are permitted.

When Nearest-Dose Rounding Is Justified

A common approach is to find an unconstrained optimal dose \(d^*\) and choose the feasible dose closest to it. This is justified under a specific shape of the objective, not as a general rule. Suppose the dose utility has the form \(U(d)=C-a(d-d^*)^2\), where \(C\) is constant and \(a>0\), and suppose all other aspects of the regimen are fixed. Maximizing utility over feasible doses is then equivalent to choosing the feasible dose with the smallest squared distance from \(d^*\).

The equivalence follows because \(C\) is the same for every candidate and \(a\) is positive: a larger squared distance makes \(U(d)\) smaller. If there is a tie in distance, this quadratic utility assigns equal utility to the tied doses. A tie-breaking rule would then need to come from another stated consideration, not from this utility.

Outside this special case, distance from a continuous optimum need not rank feasible options correctly. Utility may be asymmetric around the optimum, or it may include a penalty that changes with the number of units or administrations. The safer general procedure is to generate feasible candidates and evaluate the specified objective for each one. For a small, finite set, direct enumeration is transparent; for a larger set, the same constraints can be encoded in an appropriate discrete optimization method.

Why the Nearest Feasible Dose Can Be Wrong

The derivation above requires a symmetric quadratic loss around the continuous optimum, with no other dose-dependent features that change the ranking. A dose may be closer numerically but still have lower utility if, for example, the cost of exceeding an efficacy target is steeper than the cost of falling short. The relevant comparison is between utilities for feasible regimens, not simply between their dose values.

Representing Multiple Implementation Limits

When several restrictions apply at once, write them down before scoring doses. This makes it possible to distinguish a clinically unattractive candidate from an impossible one. A dose above a hard maximum is excluded; it is not merely assigned a modest utility penalty unless the decision model explicitly treats that limit as a soft preference. Likewise, an administration limit is a constraint on how the regimen is delivered, not just a preference for fewer units.

Diagnostics, Alternatives, and Limitations

A recommendation can appear feasible in a mathematical model and still fail an implementation check. Before reporting it, verify that the strengths correspond to the intended product, the quantities are integers where required, and the proposed combination is permitted. Check the dose per administration and total dose on the same basis as the stated limits. For example, a maximum daily dose must be compared with the daily total, not mistakenly with each individual administration.

Also check whether the dose increment applies to each administration or to a total dose, and whether the increment is compatible with the available strengths. If a candidate depends on tablet splitting, combining formulations, or using multiple administrations, verify that the assumed practice is allowed. A constraint table or a direct enumeration of combinations can expose errors such as an omitted candidate, an impermissible unit count, or an incorrectly applied cap.

Three approaches are worth distinguishing. Nearest-dose rounding is simple and appropriate when utility is quadratic around the continuous optimum and the feasible set is correctly specified. Relax-and-round first solves a continuous problem and then rounds; it can be a useful way to generate candidates, but the rounded result still needs a feasibility check and direct utility comparison. Direct discrete optimization evaluates or optimizes over feasible regimens themselves, so it can accommodate asymmetric utility and unit-dependent constraints without relying on a nearest-dose assumption. For small candidate sets, enumeration makes the reasoning easy to audit; more complex sets may require an optimization algorithm, but the model and constraints still need explicit verification.

Formulation-constrained optimization is only as reliable as its inputs. Availability, permitted combinations, and administration limits must be correct for the relevant setting. The chosen utility also represents a modeling judgment: an optimal score does not by itself prove that differences are clinically important or that the model captures every relevant outcome. When several candidates have similar utility, report that fact rather than implying a decisive distinction based on a tiny numerical difference. The central safeguard is to keep feasibility and utility separate: first exclude regimens that cannot be implemented, then compare the remaining regimens under the stated objective.