From a Proposed Use to a Bounded Prediction
A QSP model can produce many outputs for many virtual patients, dosing regimens, and time points. A proposed use such as “compare two regimens” does not yet say which patients are being compared, what outcome matters, or when it is measured. Without those details, neither the model prediction nor the scope of the conclusion is well defined.
An intended-use claim states what a model is intended to predict, for whom, under which interventions, for which outputs, and over what horizon. It is a statement about the model’s target use, not a statement that the model has already been shown to be reliable for that use. The introductory material on model qualification provides the broader context; here the focus is making the proposed use precise enough to identify its prediction target.
The claim should be narrow enough to be evaluated and broad enough to represent the actual decision. “Predicts efficacy” is too broad: it does not identify an efficacy output, population, intervention, time point, or type of prediction. “Predicts mean change in biomarker concentration at week 24 for adults meeting specified entry criteria assigned to regimen A or regimen B” is much more informative. It still needs operational definitions for the population, regimens, and summary, but its prediction target is visible.
Specify the Prediction Target
A practical way to formulate a claim is to write down five elements before discussing whether evidence supports it. The target population includes characteristics that define the intended patients, not merely the patients convenient for a simulation. Interventions should identify dose, schedule, route, duration, and relevant co-interventions when these can affect the output. The output should name the variable and its units. The horizon identifies when predictions are made, including whether treatment continues through that time. Finally, the summary says whether the target is an individual response, a population mean, a probability, a contrast, or another defined quantity.
State the clinically relevant inclusion characteristics and any exclusions that define who the claim covers.
Specify each regimen and comparator, along with background treatment or implementation conditions that affect predictions.
Define the measured quantity, units, and measurement time or time interval.
State whether the claim concerns a patient-level outcome, a population summary, a probability, or a comparison.
Make assumptions such as adherence or background therapy explicit, and avoid extending the claim beyond the stated population, intervention, output, or horizon.
These elements are related, not interchangeable. A population mean at week 24 is not the same target as the outcome for a typical individual, and neither is equivalent to the probability that an individual crosses a response threshold. Similarly, predicting a biomarker under one regimen does not by itself define a comparative treatment effect. A contrast requires both regimens and a common target population and summary.
Worked Formulations
Worked Example: Comparing Two Regimens on a Biomarker
Hypothetical proposal. A development team wants to use a QSP model to compare two dosing regimens for adults with a specified inflammatory disease. Both regimens are added to the same background therapy. The biomarker is measured at baseline and week 24, in units per milliliter. The intended summary is the difference in population mean change from baseline.
Let \(Y_{24}\) be the week-24 biomarker concentration and \(Y_0\) the baseline concentration. For regimen \(r\), define the individual change as \(\Delta_r=Y_{24,r}-Y_0\). The target contrast is
where \(\mathcal{P}\) denotes the specified target population. The intended-use claim is: “For adults meeting the stated disease and entry criteria who receive the specified background therapy, the model predicts the difference between regimens A and B in mean change from baseline in the biomarker at week 24, in units per milliliter, under the stated dosing and adherence conditions.”
Suppose, hypothetically, the simulated mean change is \(-2.4\) units per milliliter for regimen A and \(-3.1\) units per milliliter for regimen B. The predicted contrast is
Under this sign convention, the model predicts that mean change is 0.7 units per milliliter higher under A than under B. Because both changes are negative, B has the larger mean decrease. The sign and units are interpretable only because the output, comparator order, and direction of subtraction have been stated. This calculation is a description of the hypothetical model prediction, not evidence that the regimen difference is clinically meaningful or that the model is qualified for this use.
Worked Example: Predicting a Threshold Response Probability
Hypothetical proposal. A team proposes to predict the probability that an adult in a defined target population achieves at least a 30% reduction in a biomarker by week 12 while receiving a specified regimen. The population, regimen, background therapy, and adherence conditions are fixed in the claim.
For an individual, define response as \(R=1\) if the week-12 reduction from baseline is at least 30%, and \(R=0\) otherwise. The population target is
Suppose 10,000 simulated patients from the stated virtual population are evaluated, and 6,800 meet the response definition. The estimated response probability is
The corresponding claim is that the model predicts a 68% response probability for that population under those conditions at week 12. It is not a claim that 68% of all patients will respond in every setting, nor does the number alone describe uncertainty from estimated parameters, model structure, or population representation. If a reported interval is intended to reflect uncertainty in the predicted probability, its source and interpretation must be specified; it should not be silently treated as an interval for individual outcomes.
The response definition is essential. If “response” were instead defined as a 30% reduction at any time through week 12, or as a reduction sustained at week 12, the target would change. The output and horizon must encode which definition is intended.
Worked Example: Defining a Safety Prediction Horizon
Hypothetical proposal. A team plans to use a model to predict the proportion of adults in a specified population whose biomarker exceeds a prespecified safety threshold at least once during 26 weeks of treatment with regimen C. The claim assumes the stated dosing schedule, background therapy, and adherence implementation.
Let \(S(t)\) be the safety biomarker concentration at time \(t\), and let \(c\) be the stated threshold in the biomarker’s units. For an individual, define an event indicator
The target is \(P(I=1\mid\mathcal{P},C)\): the probability of at least one threshold exceedance during the 26-week treatment period. If, hypothetically, 1,250 of 10,000 simulated patients cross the threshold at least once, then
The model therefore predicts a 12.5% probability of at least one exceedance over that horizon, for the defined population and conditions. This is different from the probability of exceeding the threshold at week 26 alone. It is also different from the expected number of exceedances or the duration above threshold. A claim that omits “at least once during 26 weeks” leaves those distinct targets unresolved.
Why the Formalization Matters
The value of making a claim explicit can be shown by writing the prediction as a function of the relevant conditions. Let \(Y(t;z,a)\) denote a model output at time \(t\) for patient characteristics \(z\) under intervention \(a\). A population-level target requires a distribution of characteristics \(Z\) representing the target population, as well as a specified summary. For a mean output at time \(T\), one possible target is
A comparison of two interventions in that same population is
The common population and the time \(T\) are not decorative details: changing either changes the target quantity. If the claim instead concerns a threshold event, the target has a different form:
These expressions provide a short justification for the claim checklist. A prediction is a function of at least the patient distribution, intervention, output, horizon, and summary. Leaving one unspecified can leave the mathematical target non-unique. For example, two virtual populations with different baseline severity distributions can yield different mean outcomes under the same regimen, even if the individual-level model equations are unchanged.
The expressions also clarify what a claim does not establish. A precise target does not show that the model accurately predicts it. Nor does a predicted contrast alone establish a causal effect in clinical practice. Causal interpretation requires appropriate assumptions about the intervention comparison and the representation of the target population, in addition to a well-defined model prediction. Qualification evidence is a separate question: it must be relevant to the claim’s particular population, interventions, outputs, and horizon.
Diagnostics, Alternatives, and Scope Limits
A useful diagnostic is to ask whether two analysts could read the claim and independently identify the same simulated population, regimen implementation, output calculation, time point, and summary. If not, the wording is not yet operational. Other warning signs include an undefined response threshold, a horizon stated as “long term,” a comparison that does not identify its reference regimen, or an output named only as “efficacy” or “safety.”
A common mistake is to write, “The model is qualified to predict treatment response in patients,” and then treat that as a single, universal use. The reasoning fails because “patients” could refer to different populations, “response” could mean several outputs, and no intervention or horizon is stated. The correction is to make a bounded claim, such as predicting a specified population’s probability of meeting a defined biomarker-response threshold by a specified time under a specified regimen. Even then, qualification should be limited to the use and conditions actually supported by relevant evidence.
An alternative to one broad claim is to state several narrower claims: for example, one for a week-12 response probability and another for a week-24 mean biomarker change. This can make the targets clearer, but it does not make them interchangeable. A single integrated claim may be useful when a decision depends on joint efficacy and safety outcomes, provided the joint event or summary is explicitly defined. A list of outputs without a decision-relevant summary is not a substitute for a claim.
Claims also have practical limits. A model may be asked to predict a dose outside the range represented in its evidence, a population with materially different baseline characteristics, or a horizon beyond the available observations. Naming those extensions makes the scope visible; it does not remove their uncertainty. Likewise, a predicted quantity can be precisely defined but still be difficult to estimate or clinically interpret. The claim should describe what the model is asked to predict, while evidence and decision criteria address whether that prediction is adequate for its intended use.
Check Your Understanding
For each item, identify the issue or write the specified target. Suggested solutions follow the questions.
- A proposal says, “Use the model to predict efficacy in adults.” Name three details that must be added to make this an operational claim.
- A team predicts mean biomarker concentration at week 16 under regimen A. What additional elements are needed to state a comparison against regimen B?
- In a specified population, a simulation estimates that 540 of 1,000 patients meet a defined response threshold at week 8. State the estimated probability and its interpretation.
- Explain why “probability of a safety threshold exceedance at week 20” and “probability of any exceedance through week 20” are different claims.
- True or false: A well-specified intended-use claim proves that the model is accurate for that use. Explain.
Solutions. (1) At minimum, name an efficacy output and its units or definition, the intervention and any comparator, the target population, and the prediction time or horizon; also state the intended summary. (2) Define regimen B under comparable conditions, use a common target population and week-16 output, and specify the contrast order and summary, such as the difference in population mean concentration. (3) \(\hat{p}=540/1{,}000=0.54\), or 54%; the model estimates that 54% of the defined population meet the stated response criterion by week 8 under the stated conditions. (4) The first concerns the biomarker’s status at week 20 only; the second concerns whether it crossed the threshold at any time from treatment start through week 20, so a transient earlier exceedance counts only for the latter. (5) False. Precise wording defines the prediction target and scope; evidence is still needed to assess whether the model predicts that target adequately.