Why schedule constraints change the optimization
A dose amount cannot be optimized in isolation when a regimen must also obey limits on total daily dose and administration frequency. Increasing the number of administrations may improve a model-predicted outcome, but it shortens the interval and may violate a minimum-interval rule. Increasing the amount per administration may improve the same outcome, but it can exceed the permitted daily total. The optimization must account for all of these constraints at once.
Here the schedule is an exact repeating daily schedule: the same number of administrations is given in each 24-hour day, with administrations equally spaced. The optimization is over the dose amount per administration and the number of administrations per day. This definition matters: a schedule every 30 hours, for example, is not an exact schedule with a fixed integer number of administrations in every 24-hour day.
The objective is not determined by these constraints. It must be specified separately, using an appropriate model-based quantity, such as a predicted response, probability of meeting a target, or a utility that also accounts for administration burden. The earlier material on representing dose schedules as optimization variables provides the general optimization framing; here the emphasis is on constructing and searching the feasible set created by these three limits.
Deriving the feasible set
For an exact daily schedule, the interval constraint gives \(24/n\geq\tau_{\min}\). Since \(n\) is positive, this is equivalent to \(n\leq24/\tau_{\min}\). The frequency cap separately requires \(n\leq f_{\max}\). Combining these with the requirement that \(n\) be an integer gives the allowable frequencies:
This expression gives a largest allowable frequency only when \(n_{\max}\geq1\). If it evaluates to zero, there is no feasible exact daily schedule: even one administration per day would violate at least one of the interval or frequency limits. For each allowable \(n\), the daily-dose cap gives \(d\leq D_{\max}/n\). Thus the continuous feasible dose range at that frequency is \(0\leq d\leq D_{\max}/n\), unless the practical regimen also restricts doses to a specified set of available amounts.
The restriction to integer \(n\) is substantive, not a numerical detail. A continuous optimization over \(\tau\) might return an interval that does not divide 24 hours into a whole number of equal intervals. Rounding that result can alter the daily dose, violate the minimum interval, or change the predicted outcome. The safe approach is to optimize over feasible integer schedules directly, or to verify every proposed rounded schedule against all constraints and recompute its outcome.
Worked Example: Finding Feasible Frequencies and Dose Limits
Consider a hypothetical regimen with \(D_{\max}=600\) mg/day, \(\tau_{\min}=8\) hours, and \(f_{\max}=3\) administrations/day. Find the feasible frequencies, intervals, and largest allowed dose per administration.
The maximum frequency from the interval limit is \(24/8=3\) administrations/day. The frequency cap is also 3, so
The allowed values are \(n=1,2,3\). Their intervals are \(24/n\): 24, 12, and 8 hours, respectively. The dose cap at each frequency is \(d\leq600/n\) mg per administration:
| Administrations/day \(n\) | Interval \(24/n\) | Maximum dose per administration \(D_{\max}/n\) |
|---|---|---|
| 1 | 24 hours | 600 mg |
| 2 | 12 hours | 300 mg |
| 3 | 8 hours | 200 mg |
For instance, 3 administrations of 200 mg give \(3(200)=600\) mg/day and an 8-hour interval, so both the daily-dose and interval limits are met. A 3-administration schedule of 210 mg each would give 630 mg/day and is not feasible, even though its frequency and interval are allowed.
The feasibility check also handles an important edge case. If instead \(\tau_{\min}=30\) hours and \(f_{\max}=2\), then \(n_{\max}=\lfloor\min(24/30,2)\rfloor=\lfloor0.8\rfloor=0\). Because a daily schedule requires \(n\geq1\), this does not mean that a zero-administration schedule is the answer; it means there is no feasible exact daily schedule under those limits.
Choosing an objective and comparing feasible schedules
Constraints identify which regimens may be considered; they do not identify the best one. A model-informed optimization needs an objective with a clear interpretation. For example, \(J(d,n)\) might be a predicted probability of attaining a clinical target, with the goal of maximizing that probability. Alternatively, the goal could be to minimize daily dose while meeting a specified target probability. These are different optimization problems and can select different regimens.
For every candidate, first apply the constraints, then evaluate the objective using the stated model and assumptions. If doses are restricted to available tablet strengths or other practical increments, enumerate only those doses. If a continuous dose range is permitted, optimize within \(0\leq d\leq D_{\max}/n\) for each allowable \(n\), using a method appropriate to the objective. The model's predicted value must be recomputed at the actual candidate dose and interval; an objective evaluated at an infeasible continuous solution does not establish the value for a rounded regimen.
Worked Example: Maximizing Predicted Target Attainment
Suppose the limits are \(D_{\max}=600\) mg/day, \(\tau_{\min}=8\) hours, and \(f_{\max}=3\) administrations/day. A hypothetical model has been evaluated for the following candidate regimens. Its outputs are predicted probabilities of meeting a prespecified target; they are illustrative predictions, not data from a real study.
| \(n\) (administrations/day) | \(d\) (mg/administration) | \(nd\) (mg/day) | Predicted target probability |
|---|---|---|---|
| 1 | 400 | 400 | 0.42 |
| 1 | 600 | 600 | 0.55 |
| 2 | 200 | 400 | 0.62 |
| 2 | 300 | 600 | 0.78 |
| 3 | 150 | 450 | 0.75 |
| 3 | 200 | 600 | 0.86 |
The feasible-frequency calculation gives \(n_{\max}=\lfloor\min(24/8,3)\rfloor=3\), so all three listed frequencies are allowed. Check the daily-dose limit for each candidate: the totals are 400, 600, 400, 600, 450, and 600 mg/day, all at or below 600 mg/day. Therefore all six candidates are feasible. Under the stated objective, the largest predicted probability is 0.86, achieved by \(d=200\) mg and \(n=3\), with an 8-hour interval and 600 mg/day.
The conclusion is conditional on the objective and on the model predictions. It is not a general claim that three administrations per day are preferable: a different model, target, or burden penalty could yield a different ranking. In particular, if the objective also penalized the number of administrations, that penalty would have to be included explicitly before comparing schedules.
Worked Example: Minimizing Daily Dose While Meeting a Target
Use the same hypothetical constraints and model predictions, but now require a predicted target probability of at least 0.80 and minimize total daily dose. The candidate regimens are those in the preceding example, and the search is over this stated candidate set.
The one-administration candidates have probabilities 0.42 and 0.55, both below 0.80. The two-administration candidates have probabilities 0.62 and 0.78, also below 0.80. The three-administration candidate with 150 mg per dose has probability 0.75, below the target. The candidate with 200 mg per dose has probability 0.86 and daily dose
Thus, among the specified candidates, the 200 mg every 8 hours regimen is the only one that meets the probability requirement and is therefore the minimum-daily-dose feasible choice in this candidate set. This conclusion would require additional searching if other dose amounts were allowed. A rigorous continuous-dose optimization would need to evaluate the model throughout each feasible dose range, rather than infer that no unlisted dose can meet the target.
Justification, diagnostics, and common pitfalls
The integer enumeration follows directly from the definition of an exact daily schedule. With equal spacing, \(n\) intervals fit into 24 hours, so each interval is \(24/n\). Requiring \(24/n\geq\tau_{\min}\) and multiplying by positive \(n\) gives \(n\leq24/\tau_{\min}\). Intersecting that upper bound with \(n\leq f_{\max}\), then retaining positive integers, yields the set from 1 through the floor of the smaller bound. At any retained \(n\), \(nd\leq D_{\max}\); division by positive \(n\) gives \(d\leq D_{\max}/n\). These steps establish both the integer frequency rule and the per-administration dose ceiling.
A practical diagnostic is to report the active constraint at the selected regimen. If \(nd=D_{\max}\), the daily-dose cap is binding; if \(n=f_{\max}\), the frequency cap is binding; if \(24/n=\tau_{\min}\), the interval limit is binding. A selected regimen can have more than one binding constraint. If none is binding, a better objective value may still be available elsewhere, so feasibility alone is not evidence of optimality.
Common warning signs include a computed \(n_{\max}=0\), a selected total daily dose above \(D_{\max}\), or an interval shorter than \(\tau_{\min}\). Another warning sign is an “optimal” interval that does not correspond to an integer number of administrations per day under the stated schedule definition. When a model-based objective is non-smooth, non-monotone, or has multiple local optima, a coarse dose grid can miss the best regimen; refine the search or use a suitable optimization method, while verifying the final candidate directly. If no candidate satisfies a required target, report that the target is infeasible within the specified limits and search domain rather than labeling the closest candidate as target-achieving.
The main alternatives answer different questions. A continuous optimization over interval and dose can be useful for exploration, but any resulting regimen must be mapped to permitted integer frequencies and re-evaluated. Splitting a fixed daily dose among administrations is appropriate when the total daily amount is fixed; here, by contrast, the daily amount may vary up to its cap and is part of the decision. Direct enumeration is especially transparent when only a few frequencies are feasible. It also makes it straightforward to show which constraints exclude a candidate and to compare model predictions on the same basis.
These calculations have limits. They assume a repeating, equally spaced 24-hour schedule and a daily-dose cap defined on that same 24-hour basis. They do not resolve schedules with irregular administration times, constraints tied to specific clock times, or a dose interval that shifts across calendar days. They also do not account automatically for patient-specific uncertainty, adherence, or operational burden; such factors require an appropriate model, objective, or additional constraints. The administration-time feasibility issues are distinct from the frequency and daily-dose calculations here.
Check Your Understanding
For each question, determine feasibility or the requested optimum from the information given. Solutions are included below the questions.
- With \(\tau_{\min}=6\) hours and \(f_{\max}=3\), what frequencies and intervals are allowed?
- If \(D_{\max}=480\) mg/day and \(n=3\), what is the maximum dose per administration?
- With \(\tau_{\min}=30\) hours and \(f_{\max}=2\), does an exact daily schedule exist? Explain.
- A model predicts target probabilities 0.70 for 250 mg twice daily and 0.82 for 150 mg three times daily. If both are feasible and the objective is to maximize target probability, which is selected?
- Why is a continuous optimizer's interval not sufficient to establish the best exact daily schedule?
Solutions. (1) \(n_{\max}=\lfloor\min(24/6,3)\rfloor=\lfloor3\rfloor=3\). The allowable frequencies are 1, 2, and 3 per day, with intervals 24, 12, and 8 hours. (2) The daily-dose condition is \(3d\leq480\) mg/day, so \(d\leq480/3=160\) mg per administration. (3) \(n_{\max}=\lfloor\min(24/30,2)\rfloor=\lfloor0.8\rfloor=0\); since \(n\) must be a positive integer, there is no feasible exact daily schedule. (4) Select 150 mg three times daily because 0.82 is greater than 0.70, provided both candidates satisfy all stated limits. (5) The continuous result may not correspond to an integer number of equally spaced administrations in 24 hours. Candidate integer schedules must be checked against the limits and their model outcomes evaluated directly.