The FDA guidance (1997b) suggests that for certain drugs, in vitro dissolution results might be sufficient to gain regulatory approval for post-marketing changes and waiver of bioequivalence requirements for lower strength dosage forms (Moore and Flanner 1996a, b). A formal similarity evaluation is a regulatory (FDA 1995, 1997a, b, c, 2000a, b, 2003; EMA 2008) for this purpose. Thus the question of demonstrating similarity or equivalence between the reference and test drug dissolution curves is of both scientific and regulatory importance.

Various methods comparing drug dissolution profiles have been proposed. In general, they can be classified into three categories: (1) model-independent approaches based on a similarity factor, (2) model-independent methods using multivariate statistical distance (MSD) test, and (3) model-dependent methods using parametric curves to describe dissolution profiles.

The profile comparison methods recommended in regulatory guidance lacks detail. Perhaps this is intentional since the scope of products regulated is wide and regulatory bodies often expect sponsors to justify approaches scientifically appropriate for a given product. Below are some details that sponsors should take into consideration when justifying a profile comparison method.

There are a number of situations in which a comparison of in vitro dissolution profiles may be required.

In vitro drug release profiles are known to be affected by formulation factors such as API particle size, excipients, presence of surfactants or polymers. Besides these formulation factors, the in vivo absorption is also complicated by interactions with physiologic parameters such as pH, transit times, presence or absence of food, and gastrointestinal physiology (Kesisoglou and Wu 2008). Characterizing the link between in vitro release and clinical outcome has been a challenging task for immediate release formulations and remains an area of intensive scientific interest. Recent papers have reported the use of dissolution media simulating physiologic conditions as an improved way to develop an IVIVC exploiting mechanistic models (Jamei et~al. 2009; Sugano 2009; Sjögren et~al. 2013) relating physical and chemical properties of the API with oral absorption and dose.

Much interest has focused on models describing in vitro release profiles over a time scale. A good review of mechanistic profile models is given by Costa and Lobo (2001). These and some empirical release models are considered by Costa et~al. (2003). Some of these models are given in Table 23.3.

The models do not consider a lag time or background. They assume that the drug release at t = 0 is zero and that drug release commences immediately. However, many of the models in the table above can be modified to accommodate lag times and/or background levels of dissolved drug. Additional parameters such as mechanistic constants can be accommodated in some models. The maximum dissolution can also be accommodated in the exponential, probit, logistic, Weibull, and Gompertz models. Of the models presented in the table, the Weibull model, which is has a mechanistic interpretation, often provides the most satisfactory fits to dissolution data.

Each parameter in these models can be treated as fixed and estimated using nonlinear least squares methods. In some cases the models can be linearized by various transformations and the parameters estimated by linear least squares. Random variation can also be incorporated into the models either as measurement variance or as variance associated with one or more of the model parameters (i.e. as a nonlinear mixed model).

Hierarchical models (see Gelman and Hill 2007) can be considered which account for random variation at various levels, such as batch to batch, run to run, or vessel to vessel. The appropriate model can be identified using available data by the usual modeling approaches (e.g., minimum AIC). However, model selection should not be based on statistics, but should consider known underlying mechanistic processes.

In order to make a judgment about the similarity of two dissolution profiles, a clear, prospective definition of similarity must be established. It is convenient to view similarity geometrically as a region in some space. If the projection of two profiles onto this space falls within a defined region, similarity can be concluded. Similarity regions can either be defined within the parameter (population) space or the observed data space. This concept is illustrated in Fig. 23.1.

Consider a comparison of a Test and Reference profile. The true (population) profiles are shown in Fig. 23.1a in blue and red, respectively. These true profiles are unknown, not based on observed data, and are therefore conceptual. A region of similarity might be defined in the p-space of the true (population) differences (’s where i = 1, 2,…,p and p = number of time points) between the profiles. Restrictions may be placed on the magnitudes of the , or some function of them, based on bioavailability and bioequivalence considerations, to define a similarity region. The resulting region of similarity is thus defined in the space of population parameters. This population based similarity definition is useful because a decision rule for similarity can then be based on statistical equivalence testing of the true population parameters (in this case the ) once observed data are available. This is the approach that was taken in defining bioequivalence (FDA 2001). This was also the approach also advocated by Eaton et~al. (2003) for similarity comparisons. Statistical modeling presumes some understanding of the underlying dissolution process and thus represents a basis for risk based decision making and continuous knowledge building. It is important to recognize that safety and efficacy risk is associated with an underlying true population state and that similarity must be defined parametrically. Parametric definitions often permit us to leverage known theory and understanding of underlying mechanisms within a statistical model.