Enforcing Dose Bounds and Rate Limits

Why Dose Bounds and Rate Limits Need Explicit Enforcement

A model can produce a numerical dose recommendation that is useful as a starting point but not eligible to prescribe as written. A minimum or maximum dose may restrict the permissible range, and a rate limit may restrict how far the dose can change at one decision time. These rules answer different questions: absolute bounds define which doses are permitted overall, while rate limits define which changes are permitted from the current dose.

The distinction matters when the unconstrained recommendation lies outside either range. It also matters when a dose must be selected from a discrete set, such as multiples of 5 mg. Merely rounding a model output, or clipping it to one limit without checking the others, can produce a dose that still violates a rule. The reliable approach is to define the full feasible set first, choose a dose from that set, and verify the final value against every constraint.

This formulation assumes that all doses use the same quantity and time basis—for example, mg per day throughout—and that the bounds and rate limits are in force at this decision time. Intervals are taken as inclusive: a dose equal to a specified minimum, maximum, or permitted change is allowed. Upward and downward limits need not be equal. They may represent absolute dose changes or be derived from a stated percentage rule; in either case, calculate their dose-unit equivalents before intersecting intervals.

Finding and Using the Feasible Interval

Write each rule as an interval before altering the model output. The absolute range is \([L,U]\). A permitted decrease no greater than \(r_{\downarrow}\) means \(d\geq d_0-r_{\downarrow}\); a permitted increase no greater than \(r_{\uparrow}\) means \(d\leq d_0+r_{\uparrow}\). Thus the combined limits can be written compactly as

If \(d_{\min,\mathrm{feas}}\leq d_{\max,\mathrm{feas}}\), the feasible interval is \([d_{\min,\mathrm{feas}},d_{\max,\mathrm{feas}}]\). If the lower endpoint exceeds the upper endpoint, no dose satisfies all rules at once. That is a constraint conflict, not a reason to quietly relax a limit or return an out-of-range value.

When the model's recommendation is intended to be the closest feasible dose on the same scale, the usual rule is projection onto the feasible interval: keep \(d_{\mathrm{raw}}\) if it is inside the interval; otherwise select the nearest endpoint. This is the minimizer of squared distance to the raw recommendation. It is not automatically the right rule for every decision model: if the recommendation came from a clinical utility or loss function, the principled constrained decision is to optimize that original objective over the feasible set.

Dose Grids, Rounding, and Boundary Checks

A mathematically feasible real-valued dose may not be an available dose. For instance, a formulation or dosing protocol may permit only multiples of a specified increment. In that case, first intersect the feasible interval with the set of allowed doses, then choose among those doses according to the decision rule. Rounding the raw recommendation first is not generally equivalent: rounding can move the dose outside a rate limit or absolute bound, and subsequent clipping can create a value that is not on the permitted grid.

If the objective is to stay as close as possible to \(d_{\mathrm{raw}}\), select the grid value in \(\mathcal{F}\cap\mathcal{G}\) that minimizes the distance to \(d_{\mathrm{raw}}\). If two grid values are equally close, use a prespecified tie rule; do not let software defaults decide inconsistently. The example below uses multiples of 5 mg/day and has no tie.

The order of operations deserves care. For a continuous interval and a squared-distance objective, intersecting all intervals and then projecting gives a well-defined result. In contrast, clipping to the absolute range and then applying a rate limit can obscure an empty intersection or produce a different answer if the implementation rounds between steps. A robust implementation computes the final lower and upper endpoints from all applicable limits, then applies the dose-grid rule, then checks the result against each original rule.

When the Constraints Conflict

A nonempty absolute range does not guarantee a feasible dose. The current dose may be outside the absolute range, or a strict rate limit may prevent reaching any dose within that range at the current decision. In those cases, do not describe the result as a clipped recommendation: there is no value satisfying all stated requirements. The conflict should be made explicit for resolution under the applicable dosing process.

Diagnostics, Alternatives, and Common Errors

A small set of explicit checks makes errors easier to detect. After computing a final dose \(d_{\mathrm{final}}\), verify all inequalities in the original units:

If there is a dose grid, also verify that the final dose belongs to it. Log the raw recommendation, current dose, each bound, the calculated feasible interval, the selected dose, and whether the feasible set was empty. These values allow a reviewer to locate whether a discrepancy arose in the model output, constraint calculation, or rounding step.

• Unit mismatch: combining a limit in mg per dose with a current dose in mg/day does not define a valid interval. Convert to a common dose basis before calculation.
• Sign error: a permitted decrease of \(r_{\downarrow}\) gives the lower rate endpoint \(d_0-r_{\downarrow}\), not \(d_0+r_{\downarrow}\).
• Boundary error: using a strict comparison where the rule allows an exact maximum can wrongly exclude a boundary dose. Conversely, accepting a value just beyond a limit because it is “close” can violate the rule.
• Premature rounding: rounding before applying all constraints can create an infeasible dose. Apply the stated grid rule only after establishing the continuous feasible interval.
• Empty-set concealment: forcing the lower endpoint down to the upper endpoint when they cross invents a feasible value. Report that the intersection is empty instead.

Floating-point calculations can represent decimal values with tiny numerical discrepancies. If the implementation uses a tolerance, it should be specified in dose units, justified by the representation and dose precision, and applied consistently. A tolerance must not silently expand a clinical limit. The displayed and transmitted dose should be checked after formatting as well as before it, since rounding for display can make the communicated value differ from the value that was checked.

The main alternatives are not interchangeable. Sequential clipping is simple but can hide conflicts and becomes fragile when rounding is inserted between steps. Projection onto the intersection is transparent when the goal is the nearest dose on a continuous scale. For discrete doses, direct selection from the feasible grid is safer than rounding and clipping. Finally, if the underlying recommendation maximizes an expected utility or minimizes a clinically defined loss, the constrained optimum should be found by evaluating that objective over feasible doses, rather than assuming closeness to \(d_{\mathrm{raw}}\) is clinically equivalent.