Deriving Explicit Dose-Adjustment Instructions from Model Predictions

From Model Predictions to an Actionable Rule

A model prediction is not yet a dosing instruction. A prediction such as “the probability of response is 0.74” becomes actionable only when the decision threshold is specified: is 0.74 enough to justify treatment, and what safety risk is acceptable? A useful instruction also identifies the doses a patient can actually take, how much to change the current dose, and when to reassess.

This tutorial develops a hypothetical rule for a drug with several allowed once-daily doses. The rule uses predicted probabilities of clinical benefit and harm at each dose for the individual patient. The numbers and thresholds below are illustrative, not recommendations for any real drug. In practice, clinical evidence, the intended patient population, the seriousness of outcomes, available dosage forms, and the benefit–risk judgment must support each threshold and action.

A rule built from these predictions requires more than a fitted model. The predictions must apply to the patient and decision at hand. For this example, assume the model estimates outcomes under each candidate dose using relevant patient characteristics, that the dose and outcome definitions match the intended prescribing decision, and that the model has adequate predictive performance in the population where the rule will be used. Assume also that the required patient characteristics and safety information are available, the doses are feasible with marketed dosage forms, and no contraindication or clinical circumstance overrides the algorithm.

The threshold values are decision criteria, not quantities that the model discovers on its own. They should be chosen and justified through a benefit–risk assessment. A predicted probability of a safety event is not the same as a certainty that the event will occur in an individual patient; it summarizes risk among patients with the modeled characteristics under the modeled conditions.

Specify the Decision Before Calculating the Dose

Consider a hypothetical drug with allowed doses \(D=\{25,50,75,100,125\}\) mg once daily. Suppose the decision makers have justified these criteria for the example: predicted probability of benefit must be at least 0.80, and predicted probability of a specified serious safety event must be no more than 0.10. These cutoffs define which doses are eligible for the rule; they are not universal clinical standards.

Here \(\mathcal{E}(z)\) is the set of eligible doses for a patient with characteristics \(z\). To make the instruction unambiguous, this example selects the lowest eligible dose. That tie-breaking principle expresses a dose-minimization objective: among doses meeting both stated criteria, choose the smallest. It does not claim that the lowest dose is best under every possible clinical objective.

If the patient is currently taking \(d_{\text{cur}}\), the adjustment is \(d^*(z)-d_{\text{cur}}\) mg once daily. A positive value means an increase; a negative value means a reduction; zero means no change. If no allowed dose meets both criteria, the rule does not invent a compromise dose. It specifies that the automatic recommendation cannot be made and that the case requires clinical review or another treatment decision.

Why This Rule Selects the Lowest Eligible Dose

The calculation is a constrained decision rule. Its constraints are the prespecified benefit and safety thresholds; its objective is to minimize dose among candidates satisfying both. If a finite set of doses contains at least one eligible dose, a smallest eligible dose exists, so the rule has a unique answer under the stated tie-breaking convention. If the set is empty, there is no dose satisfying the constraints.

The adjustment amount follows directly from the selected candidate, not from rounding a model output to an arbitrary percentage. For example, if the rule selects 75 mg and the patient currently takes 100 mg, then \(75-100=-25\) mg once daily. Restricting selection to \(D\) ensures the instruction corresponds to a permitted, deliverable dose. If dose strengths, tablet splitting, or dosing frequency differ in practice, the allowed set must be revised before the rule is used.

Apply the Same Criteria to Dose Reduction

A rule should not be written only for escalation. The same paired criteria can identify a lower eligible dose when the current dose has an unacceptable predicted risk. The eligible dose is not necessarily the dose immediately below the current one; each allowed candidate must be assessed against both criteria.

When No Candidate Meets Both Criteria

A model-based rule needs an explicit outcome for conflict between benefit and safety. If all doses below the benefit threshold are safe but every dose meeting the benefit threshold exceeds the safety limit, selecting a dose anyway would silently relax one of the stated criteria. The correct result under the defined rule is that the eligible set is empty.

Reassessment Timing and Practical Checks

A dose instruction is incomplete without a reassessment time. The timing should match the model’s predictions and the clinical process: enough time for exposure to reflect the new dose, and enough time for the relevant outcome to be assessed. In this hypothetical example, assume the drug’s half-life is 36 hours, the model concerns outcomes assessed after 14 days at a stable dose, and clinical pharmacology supports reassessment 14 days after a dose change. Five half-lives are \(5\times36=180\) hours, or 7.5 days, so 14 days allows more than five half-lives for concentrations to approach steady state before the stated assessment. This calculation alone does not establish that 14 days is suitable for a real drug or endpoint.

The operational instruction for this example can therefore state: use the selected allowed dose once daily and reassess benefit, safety, and relevant patient characteristics 14 days after the change. A clinical event requiring earlier action should not wait for the scheduled reassessment. If the model’s endpoint horizon, drug kinetics, or monitoring needs do not support that timing, the reassessment interval must be changed and justified.

Before applying any patient-level rule, check that the patient’s covariates fall within the model’s supported range, the current medication and dose are known, and the predictions refer to the same outcomes and horizons used to define the thresholds. Review whether a new interaction, organ-function change, adherence problem, or acute clinical event makes the model inputs inappropriate. If essential inputs are missing or the patient is outside the intended population, the rule should not generate a falsely precise instruction.

Alternatives, Diagnostics, and Limitations

A fixed dose reduction for a broad patient category is simpler to state and may be easier to implement, but it does not use the individual patient’s predicted outcomes. Exposure-target adjustment can be useful when a validated exposure metric and target range are central to the decision, but a target exposure is not automatically equivalent to a clinical benefit–risk criterion. The approach here instead uses predicted clinical outcome probabilities at allowed doses. These approaches answer different questions; the appropriate choice depends on the evidence and the intended prescribing decision.

Common warning signs include predictions that change implausibly between adjacent doses, predicted event probabilities outside the range supported by the data, and a rule that often produces no eligible dose. Inspect dose-specific predictions and compare them with the observed evidence supporting the model. Repeated gaps between predicted and observed outcomes, or a pattern in which the rule recommends doses that cannot be delivered with available formulations, calls for review rather than mechanical use.

The finite-dose rule also has limitations. Its answer depends on the chosen thresholds, candidate doses, model predictions, and “lowest eligible dose” objective. It does not capture every patient preference or clinical consideration, and predicted probabilities do not guarantee an individual outcome. If predictions near a cutoff are sensitive to small input changes, a hard boundary may give an unstable recommendation. Such concerns require clinical and model review; they should not be hidden by adding unjustified decimal precision or changing thresholds for a single patient.