Why a Model Needs an Applicability Domain
A model prediction is not supported merely because software can calculate it. A model may return a dose-response estimate for a population, regimen, endpoint, or clinical setting that is unlike the evidence used to develop or assess it. The result can look precise while relying on assumptions that have not been tested in the relevant setting. Defining the model’s applicability domain makes explicit where its predictions are supported for a particular use, and where additional evidence or a different model is needed.
The intended use and decision rule provide the starting point: a domain is not an abstract boundary around a model that applies equally to every question. It is a statement about the settings in which the model can support a specified prediction or decision. The earlier material on validation and predictive checks helps assess evidence against that use; the domain organizes the limits revealed by that evidence.
A useful domain statement names four dimensions. Population includes relevant characteristics such as age, disease severity, organ function, or prior treatment. Dose includes amount, schedule, route, and other regimen features that affect exposure. Outcome includes what is predicted, how it is measured, and when it is assessed. Conditions include the clinical, operational, and measurement circumstances under which the model is intended to be used, such as concomitant therapies, assay methods, or care setting.
These dimensions are considered jointly. Evidence that covers adults and evidence that covers a high dose do not by themselves establish support for high-dose use in adults if those features were never represented together or the model’s behavior under that combination was not otherwise justified. Likewise, a model validated for one endpoint or assessment time is not automatically supported for another.
State the Domain as a Use-Specific Evidence Boundary
A practical domain description should be specific enough to distinguish supported from unsupported cases. It can include categorical limits, numerical ranges, and conditions that must hold. A range alone is often inadequate: age from 18 to 75 years does not say whether the evidence includes patients at the youngest age with severe organ impairment, or whether high dose was studied in that subgroup. Domain boundaries should therefore record important combinations, not only separate marginal ranges.
“Supported” is an evidence-based judgment, not a claim that the model is correct for every individual in the domain. The judgment should match the intended decision and its evidence standard. It normally draws on the development data, validation evidence relevant to the use, and justified scientific knowledge. A range seen in the development data is informative, but it does not establish adequate prediction performance throughout that range. Conversely, a model may have support beyond the exact observed values if external evidence or a well-justified model structure supports that use; such justification should be stated rather than assumed.
Name the target population, regimen, outcome, and time point, and connect them to the intended decision.
Identify which combinations of population characteristics, doses, outcomes, and conditions are represented in development and relevant validation evidence.
Check categorical features, numerical ranges, combinations of features, and whether the proposed outcome and measurement conditions match.
State what is supported, what is outside or uncertain, and which evidence or scientific rationale supports each boundary.
Worked Applications
Worked Example: Dose Range Is Not the Whole Domain
Hypothetical use. A pharmacokinetic model is intended to predict steady-state exposure to support selection of a dose regimen for adults with a chronic condition. Development data include adults aged 20 to 70 years, doses of 10, 20, and 40 mg once daily, and patients with normal or mildly impaired renal function. Validation checks cover the 10- and 20-mg regimens in the same population. The proposed use is 40 mg once daily in adults with severe renal impairment.
Step 1: Compare dose. The proposed dose, 40 mg once daily, is within the dose values represented in development data. This fact alone does not establish support for the proposed use.
Step 2: Compare population and conditions. Severe renal impairment is not represented; the data include only normal or mildly impaired renal function. The proposed population therefore differs on a feature that may affect drug clearance and exposure.
Step 3: State the conclusion. The proposed use is outside the directly evidenced domain, even though its dose is within the observed dose range. Validation at 10 and 20 mg also does not directly validate prediction at 40 mg in severe impairment. The model can produce a numerical prediction, but the evidence described here does not establish that prediction as supported for this combination. A separate rationale or additional relevant evidence would be needed before treating it as decision-supporting.
Worked Example: Joint Support Across Population Features
Hypothetical use. A model predicts the probability of a clinical response at week 12 for patients receiving 100 mg every two weeks. The development and validation evidence includes: (a) patients aged 18 to 65 years with moderate disease receiving 100 mg every two weeks; and (b) patients aged 66 to 75 years with mild disease receiving the same regimen. The proposed use is in patients aged 66 to 75 years with moderate disease.
The age interval and disease categories each appear somewhere in the evidence. But the proposed combination, older age with moderate disease, appears in neither group. Separate coverage of the two features does not establish their joint coverage. If age and disease severity each modify response, or if their combination affects the model’s prediction, the model could behave differently in the proposed subgroup.
The relevant applicability-domain statement should therefore distinguish the observed combinations: support is evidenced for ages 18 to 65 with moderate disease and ages 66 to 75 with mild disease, under the stated regimen and week-12 outcome. The older, moderate-disease subgroup is not directly supported by these data. This is a domain limitation even though the proposed age is within the overall observed age range and moderate disease occurs in the data.
Worked Example: Outcome and Measurement Conditions
Hypothetical use. A model was developed and validated to predict change from baseline in a laboratory biomarker at week 8, using a specified central laboratory assay. A team proposes using its predictions for the same biomarker at week 24, measured with a local assay. Assume the treatment, dose, and patient population otherwise match the evidence.
The outcome dimension does not match in time: week 24 is not week 8. The measurement condition also differs because the local assay is not the assay used for development or validation. Even if both assays report the biomarker in the same units, comparable units do not demonstrate comparable measurement performance, and week-8 validation does not establish prediction performance at week 24.
Thus the proposed use falls outside the directly supported domain on two dimensions. A valid extension might be possible if evidence establishes comparability of the assays and supports the longer-term prediction, but neither follows from the stated facts. The domain should explicitly name the week-8 outcome and central assay rather than saying only “biomarker prediction.”
Why Joint Support Matters
The logic can be stated simply. Let a prediction case be represented by \(z=(P,D,Y,C)\), where \(P\) denotes population features, \(D\) dose regimen, \(Y\) outcome definition, and \(C\) conditions. Let \(E\) denote the development and relevant validation evidence, and let \(U\) denote the intended use and its evidence standard. The supported domain is the subset of cases for which \(E\), interpreted in light of \(U\), provides adequate support.
This expression is a definition, not a universal calculation that turns evidence into a single automatic boundary. Its value is to make the scope of the claim explicit. In particular, suppose \(E\) contains cases with population feature \(P_1\) and cases with dose \(D_1\), but never contains their combination. From those facts alone, one cannot infer that \((P_1,D_1,Y,C)\) belongs to \(\mathcal{A}(U,E)\). That conclusion would require evidence or a stated, defensible rationale for transferring support across the combination.
This distinction explains why a simple minimum-to-maximum range can overstate support. Ranges summarize each variable separately, while predictions depend on the joint setting. A range table is useful as a first screen, but it should be supplemented by combinations and conditions that matter to the model and decision.
Diagnostics, Alternatives, and Limitations
Domain review is going wrong when the statement is so broad that nearly any proposed use seems included, or so vague that it cannot identify a case outside support. Warning signs include listing only the overall age and dose ranges; omitting regimen schedule, outcome timing, or assay; treating all subgroups within a range as interchangeable; and assuming that a successful predictive check in one setting validates a different population or endpoint. A prediction that is unusually dependent on a sparsely represented subgroup is also a reason to examine the evidence more closely, not a reason to expand the domain by assertion.
The simplest alternative is a range-only description, such as “adults, doses from 10 to 40 mg.” It is easy to communicate but can hide unsupported combinations and mismatched outcomes. A more informative alternative is a joint evidence description that records which important combinations were represented and which predictions were assessed. A third approach is a mechanistic or scientific extension rationale: it may justify use outside the directly observed data, but it should identify the assumptions and evidence that support the extension. None of these descriptions should be confused with a guarantee of accuracy.
The domain also has limitations. Evidence is rarely balanced across all combinations of population and regimen, and clinical knowledge may not supply a sharp boundary. Therefore, some cases are clearly supported, some clearly outside, and some require a qualified judgment. State that distinction plainly. The domain can be model- and decision-specific: evidence adequate for a descriptive summary may not be adequate for a high-consequence dosing decision. Defining limits does not itself quantify how much uncertainty matters to the decision; that is a separate question.
Common Mistake
Correction: Check the joint population–dose setting, outcome, and conditions. Marginal ranges do not establish support for every combination, and an observed range does not by itself demonstrate adequate predictive evidence throughout it.
Check Your Understanding
For each item, decide what can be concluded from the stated evidence. Solutions follow each question.
- A model was assessed for 5- and 10-mg daily doses in adults with normal renal function. A proposed prediction is for 10 mg daily in adults with moderate renal impairment. Is the dose being extrapolated? Is the full use supported by the stated evidence?
Solution: The dose is not outside the stated dose range, but the population condition differs. The full use is not established as supported by the evidence given.
- A development dataset includes adolescents receiving a low dose and adults receiving a high dose. Can this alone support a prediction for adolescents receiving the high dose?
Solution: No. The proposed age–dose combination is not represented. Support would require relevant additional evidence or a stated rationale that justifies transferring evidence to that combination.
- A model predicts a week-4 outcome measured with assay A. What two domain features must be checked before using it for a week-12 outcome measured with assay B?
Solution: Check whether the outcome assessment time and measurement method match the supported evidence. Neither week-12 performance nor assay-B comparability is established merely by week-4 validation with assay A.
- Why is “ages 18 to 75 and doses 5 to 40 mg” an incomplete domain statement?
Solution: It gives marginal ranges but not the joint combinations represented, the outcome definition and timing, or the conditions of use. It may therefore include unsupported combinations.
- Does a prediction outside the stated domain prove that the model is wrong?
Solution: No. It means the stated evidence does not support that prediction to the same standard. The prediction could be accurate, but further evidence or justification is needed to rely on it for the intended use.