Dissolution Windows for Postapproval Formulation Changes

Why a Dissolution Window Is More Than a Set of Limits

A postapproval formulation change can alter how quickly, or in what pattern, a product dissolves. When an established in vitro-in vivo correlation (IVIVC) is applicable, it can help assess whether the proposed dissolution change is expected to alter in vivo performance. The key question is not merely whether every measured dissolution value falls between two numbers. It is whether the whole proposed profile, under relevant test conditions, lies within the range of profile behavior for which the IVIVC has support.

A dissolution window is therefore best understood as a model-supported region of profile shapes and rates, together with the relevant product, test, and use conditions. It is not automatically a regulatory acceptance criterion, a guarantee of bioequivalence, or permission to change any formulation whose profile passes a set of time-point limits. The assessment should be tied to a specific proposed change and to the evidence and decision criteria agreed for that assessment.

Conditions for Using an Established IVIVC

Before applying the model, confirm what evidence establishes its domain. The IVIVC should have been developed and evaluated using formulations that span relevant release behavior, and its predictive performance should be adequate for the intended use. The earlier material on IVIVC validation distinguishes evaluation within the development data from evaluation using external data; both the validation evidence and the limits it supports matter when considering a new profile.

The proposed assessment also assumes that the IVIVC's essential conditions remain relevant. These include the product type and dose, the dissolution method and test conditions, the mapping from dissolution to absorption, and the disposition assumptions used to predict exposure. For example, a prediction based on one dissolution medium does not automatically apply to another. A model developed for one dose or release mechanism should not be presumed to cover a materially different one. If the proposed change could alter absorption or disposition in a way not represented in the model, dissolution input alone may not support a dependable prediction.

The model's supported region should be stated in practical terms before evaluating a candidate. It might be defined by slow and fast reference profiles, a parameter range in a time-scaling model, or a more complex set of profile-shape constraints. The definition should identify sampling times, acceptable measurement procedures, and whether support is limited to interpolation among established formulations. It should also state the intended in vivo outputs and the prespecified criteria for judging their predictions. A commonly used bioequivalence interval should not be treated as an automatic rule for every IVIVC-based change assessment.

Profile Membership and Interpolation

A useful simple case is a model supported by a slow and a fast reference profile, with intermediate profiles represented by a common interpolation parameter \(\lambda\). At time \(t_j\), let \(D_S(t_j)\) and \(D_F(t_j)\) be the percent dissolved for the slow and fast references. A candidate profile \(D_C\) follows their interpolation if the same \(\lambda\) accounts for every sampling time:

This condition is stronger than requiring each candidate value to fall between the slow and fast values at that time. Independent time-point bounds can admit profiles that change shape in a way none of the supporting formulations represents. When the established IVIVC supports a more complex or nonlinear domain, membership must be checked using that domain's actual structure rather than forcing this simple interpolation.

The reason for requiring a common \(\lambda\) is direct. Rearranging the interpolation equation gives \(\lambda_j=(D_C(t_j)-D_S(t_j))/(D_F(t_j)-D_S(t_j))\), whenever the two references differ at \(t_j\). For a profile on the supported interpolation path, each \(\lambda_j\) equals the same value. Substantially different values indicate that the candidate is not simply an intermediate profile on that path, even if all its individual observations lie between the anchor values.

From Profile to Decision: Predictions and Their Uncertainty

An IVIVC-based assessment uses the candidate profile as an input to the established model, then evaluates predicted in vivo performance against criteria specified for the assessment. Depending on the model, relevant outputs may include AUC, \(C_{\max}\), or a time-dependent concentration or absorption profile. The prediction should be reported with its uncertainty, not just as a point estimate. Uncertainty can reflect parameter estimation, residual error, variability in the dissolution input, and other components represented in the model's prediction procedure.

Uncertainty matters because a point estimate inside a decision range does not establish that the prediction is precise enough to support the decision. Conversely, an interval that extends beyond a decision boundary is a reason to examine the uncertainty and its source; it does not prove that the true in vivo effect lies outside the boundary. The interpretation depends on the prespecified decision rule. In particular, a prediction interval describes uncertainty for a future quantity under its stated model assumptions; it is not a confidence statement that the fixed, unknown parameter has a specified probability of lying in the interval.

Worked Example: Values Within Pointwise Bounds but Outside the Profile Domain

Suppose the supported path runs between slow and fast profiles with dissolution values of 10%, 25%, 50%, and 70% for the slow formulation, and 30%, 55%, 85%, and 95% for the fast formulation, at 1, 2, 4, and 6 hours. A proposed profile is 25%, 40%, 60%, and 90% at those times.

Each candidate value lies between the two reference values at its time. But the interpolation fractions are different: at 1 hour, \(\lambda=(25-10)/(30-10)=0.750\); at 2 hours, \(\lambda=(40-25)/(55-25)=0.500\); at 4 hours, \(\lambda=(60-50)/(85-50)=0.286\); and at 6 hours, \(\lambda=(90-70)/(95-70)=0.800\). There is no common \(\lambda\).

The candidate is therefore outside the stated one-parameter interpolation path. The fact that its values satisfy pointwise bounds does not justify using that path's IVIVC predictions. The next step is to determine whether the established model has other support for this profile shape. If it does not, the profile is outside the model's supported range, and an in vivo assessment or an appropriately developed and validated extension may be needed. One should not obtain a number by extrapolating the model and present it as a supported prediction.

Diagnostics, Alternatives, and Limitations

A defensible assessment checks both the input profile and the model prediction. Review the dissolution observations for data quality, adequate sampling of the release phase, monotonicity where appropriate, and consistency with the specified test procedure. Examine whether the proposed profile's shape, rate, and any lag behavior are represented in the supported domain. Then inspect prediction uncertainty and any available diagnostics for the established model. Warning signs include a candidate outside the development range, a new curve shape, sensitivity to small changes in dissolution input, and prediction uncertainty too large to resolve the prespecified decision.

Dissolution profile similarity measures and IVIVC prediction answer different questions. A similarity measure such as \(f_2\) summarizes closeness between in vitro profiles under its applicable conditions; it does not itself establish the in vivo consequence of a difference. A dissolution specification controls product performance against defined in vitro limits; passing it does not automatically demonstrate that every profile within it is covered by an IVIVC. A direct in vivo comparison measures performance in people but requires an appropriate study and its own design and analysis. An IVIVC can provide a mechanistic, model-based bridge when it is applicable, but it does not eliminate the need for in vivo evidence when the candidate or uncertainty falls outside its support.

The following sequence keeps the decision tied to evidence:

The method has limits even when each step is followed. The model predicts only under its assumptions and the conditions represented by its evidence. A profile outside its supported domain may require additional in vivo evidence, but the need and design depend on the specific change and applicable scientific and regulatory expectations. A model extension is not supported merely because it produces plausible predictions: its performance must be evaluated for the new intended range. These limitations make a transparent statement of the domain as important as the predicted ratios themselves.