Release Kinetics: Zero-Order, First-Order, and Weibull Models

Why model a dissolution profile?

A dissolution profile records the amount or proportion of drug released as time passes. A model can summarize the profile with a small number of parameters, provide predictions between observed sampling times, and make comparisons across formulations more interpretable. The model is a mathematical description of the observed release pattern; its usefulness depends on whether its assumptions and fit are adequate for the intended purpose.

This tutorial considers three common empirical or mechanistic-form candidates: zero-order, first-order, and Weibull models. The introductory material on dissolution profile variability and unit-level analysis distinguishes measurements from individual dosage units from summaries across units. That distinction matters here: fitting a curve to mean percentages does not make the variability among dosage units disappear, and a curve fitted to means alone does not characterize that variability.

Model forms and parameter meanings

For a finite dose with a plateau, one simple zero-order model is a straight line from zero, capped at \(F_{\infty}\). Here \(k_0\) is the release rate in percentage points per unit time. The cap represents completion at the modeled plateau; the model does not imply that a formulation must physically release at a constant rate under all conditions.

Before the plateau, the profile is linear and each equal time interval adds the same percentage-point amount. The time to reach the plateau is \(F_{\infty}/k_0\) when \(k_0>0\). If \(k_0=0\) and \(F_{\infty}>0\), the model remains at zero and does not reach the plateau. If the observed times do not approach the plateau, \(F_{\infty}\) may not be estimable from those data; a straight-line fit over the observed range does not establish the eventual plateau.

In the first-order model, the instantaneous release rate is proportional to the amount not yet released. If \(R(t)=F_{\infty}-F(t)\) is that remaining amount on the same percentage scale, the rate assumption is \(dF/dt=k_1R(t)\), where \(k_1\) has units of inverse time. With \(F(0)=0\), the solution is:

The curve rises rapidly at first and then approaches \(F_{\infty}\), with a decreasing release rate. The half-release time relative to the plateau is \(\ln(2)/k_1\). Unlike the capped zero-order expression, this first-order curve approaches the plateau asymptotically and does not reach it at a finite time.

The Weibull model is a flexible cumulative curve. Its parameters are \(F_{\infty}\), the plateau percentage; \(\alpha\), a time-scale parameter with units of time; and \(\beta\), a dimensionless shape parameter.

At \(t=\alpha\), the released fraction of the plateau is \(1-e^{-1}\), approximately 0.632. Thus, \(\alpha\) is not generally the time to 50% release. The shape parameter changes the curve’s shape: \(\beta=1\) gives the first-order exponential form with \(k_1=1/\alpha\); values above or below 1 produce different patterns of curvature. These descriptions are properties of the fitted equation, not proof of a particular physical process.

Fitting and comparing the equations

A common fitting method is nonlinear least squares. For observations \(y_i\) at times \(t_i\), residuals are \(e_i=y_i-F(t_i)\), and the residual sum of squares is \(SSE=\sum_i e_i^2\). The fitted parameters minimize this sum over the chosen parameter space. For the uncapped part of the zero-order model, the expression is linear in \(k_0\); the first-order and Weibull expressions are generally fitted by numerical optimization.

The equal-variance assumption can be unsuitable when measurement variability changes with the percentage dissolved. In that case, a weighted fit or an explicit variance model may be more appropriate, but its weights or variance structure need justification. A numerical optimizer also needs sensible starting values and parameter bounds: a converged result is not sufficient evidence that the optimum is meaningful or unique.

The following examples use hypothetical data. Percentages are on the original scale, so residuals are in percentage points and \(SSE\) is in percentage-points squared.

Comparing fits without over-interpreting them

A smaller \(SSE\) means a closer fit to the observed points on the scale used, but it cannot by itself fairly compare models with different numbers of parameters. Adding parameters can reduce \(SSE\) even when the additional flexibility is not useful. One common summary for models fitted by likelihood is the Akaike information criterion, \(AIC=-2\ell+2p\), where \(\ell\) is the maximized log-likelihood and \(p\) is the number of estimated parameters. Under independent, normally distributed residuals with a common unknown variance, constants shared among models can be omitted for a relative comparison, giving \(AIC^*=n\ln(SSE/n)+2p\). Smaller values indicate the better relative trade-off among the candidate models considered; they do not certify that any candidate is adequate.

The three models offer different balances of simplicity and shape flexibility. A zero-order line is readily interpreted over a range with nearly constant release rate, but it can misrepresent a profile that curves or approaches a plateau. A first-order curve is parsimonious and has an interpretable rate constant, but its exponential shape may not capture a lag, a changing curvature, or a complex release pattern. The Weibull curve can represent a wider range of shapes, but that flexibility can make its parameters less precise, especially when there are few time points or no observations near the plateau.

Diagnostics, failure modes, and limits

Plot observed values and fitted values against time, and inspect residuals against time and fitted values. A systematic sequence of positive residuals followed by negative residuals, or the reverse, indicates that the curve is missing a feature of the profile rather than simply making random errors. Increasing residual spread as dissolution increases suggests nonconstant variance. A few influential observations can dominate an unweighted least-squares fit, so examine the data and measurement process rather than deleting points solely to improve the curve.

Check that predictions remain plausible over the times being interpreted: cumulative release should not decrease, and a percentage released should not exceed its defensible upper bound. Inspect parameter estimates and uncertainty. A fitted plateau far beyond the observed range, very large or small \(\beta\), or strong dependence between estimated parameters can signal weak identification. A numerical fit that changes substantially with starting values may indicate a difficult optimization, multiple local minima, or insufficient information in the data.

Do not infer the release mechanism from the best-fitting equation alone. Different physical processes can generate similar curves over a limited observation window, and a flexible empirical model can fit well without representing a unique mechanism. Mechanistic interpretation requires evidence beyond curve fit, such as relevant formulation knowledge and experiments designed to discriminate between explanations. These equations summarize in vitro profiles; they do not, by themselves, establish a relationship between dissolution and in vivo performance.

The similarity factor \(f_2\) is another established way to compare dissolution profiles, with a different purpose and set of conditions. Model-based curve summaries and \(f_2\) answer different questions: a fitted parameter may compactly describe a profile, while a similarity analysis evaluates profile agreement under its own framework. Neither should be treated as proof of a release mechanism. The choice of method should match the regulatory or scientific question, and conclusions should account for replicate-unit variability rather than relying only on rounded mean profiles.

Common mistake