Why Severity Interpolation Requires Care
Renal and hepatic impairment are not always studied at every clinically relevant level. A study might report exposure for a healthy group and two impairment groups, while a model is needed for a severity that falls between them—or beyond the most impaired group observed. The calculation may look like ordinary interpolation, but the important question is what is being interpolated: an exposure summary, a physiological input, or a mechanistic model parameter.
These choices are not interchangeable. A smooth curve through observed exposure ratios can give a useful local prediction, but it does not establish that the same curve continues outside the observed range. A PBPK model can make predictions through its physiological structure, but only to the extent that its relevant assumptions and impairment-dependent inputs remain credible.
Here, let \(z\) denote a defined impairment-severity coordinate, with larger values representing greater impairment when that orientation is appropriate. The coordinate might be a normalized renal-function measure, or it might be a set of hepatic physiology covariates rather than one score. A numerical score is not automatically a linear measurement of biological severity: in particular, an ordinal category should not be treated as an interval scale without justification.
Direct Interpolation of Exposure Ratios
A simple method is to interpolate the observed exposure ratios directly. For positive exposures, interpolation on the logarithmic scale is often more defensible than assuming that equal changes in severity produce equal absolute changes in exposure. It treats changes as multiplicative and guarantees a positive predicted ratio.
This calculation presumes that the chosen severity coordinate is meaningful between the two observed groups and that log exposure changes approximately linearly over that interval. It also presumes that the exposure ratios are comparable: for example, they refer to the same exposure measure and use a suitable reference group, and any relevant differences in dose, route, or population are addressed. The equation does not prove that the underlying biological response is linear.
Worked Example: Interpolating Between Two Exposure Ratios
In a hypothetical study, dose-normalized AUC ratios relative to a reference group are 1.00 at \(z=0\) and 1.60 at \(z=0.40\). Predict the ratio at \(z=0.25\), which is inside the observed range.
The interpolation fraction is \(t=(0.25-0)/(0.40-0)=0.625\). Applying the log-scale formula gives \(R(0.25)=\exp[(1-0.625)\ln(1.00)+0.625\ln(1.60)]\).
Since \(\ln(1.00)=0\) and \(\ln(1.60)\approx0.4700\), the predicted ratio is \(\exp[0.625(0.4700)]=\exp(0.2938)\approx1.342\). Under the stated interpolation assumption, predicted dose-normalized AUC at \(z=0.25\) is about 1.34 times the reference AUC.
This is a prediction between the two observed severity values, not evidence that the same trend holds at greater impairment. The calculation also inherits uncertainty in the observed ratios and in the choice of the severity coordinate.
Interpolate the Mechanism When It Is Supported
When a plausible PBPK mechanism is available, it can be more informative to interpolate impairment-dependent physiological inputs or clearance components and then simulate exposure. This preserves relationships between inputs and outputs that direct interpolation ignores. It also makes assumptions visible: the model must specify which processes impairment changes and how.
A simple renal example illustrates the distinction. Suppose AUC is proportional to \(1/CL\) under linear pharmacokinetics, with unchanged bioavailability and dose. If total clearance is the sum of renal and nonrenal clearance, and only renal clearance changes in proportion to a normalized filtration measure \(g\), then \(CL(g)=CL_{nr}+CL_{r,0}g\). Here \(CL_{nr}\) is the unchanged nonrenal clearance and \(CL_{r,0}\) is renal clearance at the reference filtration level \(g=1\).
The expression follows by writing AUC as \(F\cdot\mathrm{Dose}/CL\) in each group and taking the ratio. The common \(F\) and dose cancel only if they are unchanged. This is a deliberately simple model: it assumes linear pharmacokinetics, unchanged nonrenal clearance and bioavailability, and proportional change in the renal component. If impairment affects other processes, the formula is not sufficient.
Worked Example: Predicting AUC From a Renal Clearance Model
For a hypothetical drug, reference renal clearance is \(6.00\ \mathrm{L/h}\), and nonrenal clearance is \(4.00\ \mathrm{L/h}\). Assume renal clearance scales in proportion to filtration, while nonrenal clearance, bioavailability, and dose remain unchanged. Predict the AUC ratio when filtration is \(45\%\) of the reference level, so \(g=0.45\).
At the reference level, total clearance is \(CL(1)=4.00+6.00(1)=10.00\ \mathrm{L/h}\). At \(g=0.45\), renal clearance is \(6.00(0.45)=2.70\ \mathrm{L/h}\), so total clearance is \(4.00+2.70=6.70\ \mathrm{L/h}\).
Thus \(R_{\mathrm{AUC}}(0.45)=10.00/6.70=1.493\). The model predicts dose-normalized AUC about 1.49 times the reference value. This result is conditional on the stated clearance structure; it is not a general rule that AUC changes in inverse proportion to filtration.
Derivation and a Physiological Bound
Under the assumptions above, the reference AUC is \(F\cdot\mathrm{Dose}/(CL_{nr}+CL_{r,0})\), and the impaired AUC is \(F\cdot\mathrm{Dose}/(CL_{nr}+CL_{r,0}g)\). Dividing the latter by the former cancels the shared \(F\) and dose, yielding the ratio shown above. This derivation explains both the usefulness and the limitation of the relationship: its shape depends on a specific clearance decomposition.
If \(CL_{nr}>0\) and \(g\ge0\), then \(CL_{nr}+CL_{r,0}g\ge CL_{nr}\). Therefore, under this model, the AUC ratio cannot exceed \((CL_{nr}+CL_{r,0})/CL_{nr}\). As renal function approaches zero, the model approaches this finite maximum rather than increasing without limit. This is a model-based bound, not a universal biological ceiling; a change in nonrenal clearance or bioavailability would alter it.
Worked Example: Detecting an Implausible Extrapolation
Use the hypothetical clearance values from the preceding example. The corresponding severity coordinate is \(z=1-g\), so \(z=0.50\) gives \(g=0.50\), and \(z=1.00\) gives \(g=0\). The model predicts ratios \(R(0.50)=10/(4+6(0.50))=10/7\approx1.429\) and \(R(1.00)=10/(4+6(0))=10/4=2.50\).
Suppose someone extends a straight line fitted to log ratio against \(z\) from these two points to \(z=1.40\). The log-linear extrapolation is \(R(1.40)=2.50(2.50/1.429)^{(1.40-1.00)/(1.00-0.50)}\). Since \(2.50/1.429\approx1.75\), this is \(2.50(1.75)^{0.8}\approx2.50(1.565)=3.913\).
That extrapolated value exceeds the model’s maximum ratio of \(2.50\), and \(z=1.40\) implies \(g=-0.40\), outside the coordinate’s physiological range. The calculation is mathematically possible but not defensible under the stated model. The appropriate conclusion is not that exposure must equal 2.50 in every real patient; it is that this linear extrapolation conflicts with the particular assumptions used to derive the mechanistic model.
Choosing a Method and Checking Its Limits
Direct interpolation is transparent and can be appropriate when the target lies between well-characterized groups, the exposure relationship is smooth over that interval, and a mechanistic representation is unavailable or not better supported. Interpolating PBPK inputs and rerunning the model is preferable when the relevant physiological changes are understood and the model has been assessed for those processes. Neither approach makes an unsupported extrapolation reliable.
| Approach | Useful when | Main limitation |
|---|---|---|
| Interpolate log exposure ratio | The target is within the observed range and a local smooth trend is plausible. | It can conceal changing mechanisms and is not justified beyond the observed range by the interpolation itself. |
| Interpolate physiological inputs, then simulate PBPK | Impairment effects on relevant inputs and their model relationships are supported. | Input uncertainty and structural model error propagate into the predicted exposure. |
| Use the nearest observed impairment group | A rough reference is needed and differences from the target are small enough for the purpose. | It is not a severity-specific prediction and can misrepresent an intermediate or more extreme group. |
For hepatic impairment, avoid assuming that a single category score changes linearly with every physiological process. Changes in blood flow, functional liver mass, binding, enzyme capacity, and other processes need not follow the same severity trend. If a PBPK model uses several impairment covariates, evaluate their values and dependencies at the target severity and simulate the resulting combination; do not assume that one interpolated exposure ratio captures all of them. Similarly, renal function measures do not, by themselves, guarantee proportional changes in every renal clearance process.
Several checks help reveal when a prediction is going wrong. Confirm that the target is actually within the calibration range before calling the prediction interpolation. Plot observed group ratios against the chosen severity measure and inspect whether a straight line on the log scale is plausible over the relevant interval. Check predictions against known physiological limits, clearance components, and the model’s own parameter ranges. If modest changes in the severity coordinate or model form lead to large changes in predicted exposure, report that sensitivity rather than presenting a single precise value.
For extrapolation, compare more than one scientifically plausible continuation when possible: for example, a trend on log exposure and a mechanism-based prediction with explicitly stated clearance assumptions. Large disagreement is evidence of model-form uncertainty, not a reason to select whichever estimate is more convenient. A prediction can also be unreliable when the severe group is small, its exposure estimate is imprecise, its impairment characteristics differ from the target, or multiple covariates change together. Extrapolation beyond observed severity is most defensible when the mechanism governing the additional range is independently supported and predictions remain plausible under sensitivity checks.
A common mistake is to treat evenly spaced impairment labels as equally spaced biological values, draw a straight line through their exposure ratios, and extend it beyond the most severe group. The correction is to define a defensible severity coordinate, distinguish interpolation from extrapolation, and justify the relationship being extended. When the necessary mechanistic support is absent, state that the prediction is uncertain or not supported rather than implying that the observed trend must continue.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Observed ratios are 1.20 at \(z=0.20\) and 1.80 at \(z=0.80\). Using log-scale interpolation, calculate the ratio at \(z=0.50\).
- Under the clearance model, let \(CL_{nr}=3.00\ \mathrm{L/h}\), \(CL_{r,0}=7.00\ \mathrm{L/h}\), and \(g=0.40\). Calculate the predicted AUC ratio and state one key assumption.
- Why can an ordinal hepatic impairment category be unsuitable as a linear severity coordinate?
- A target severity lies outside the observed range, and a log-linear extrapolation conflicts with a clearance-model bound. What should be concluded?
Solutions. (1) Here \(t=(0.50-0.20)/(0.80-0.20)=0.50\), so \(R=\exp[0.5\ln(1.20)+0.5\ln(1.80)]=\sqrt{1.20(1.80)}=\sqrt{2.16}\approx1.470\). (2) Reference clearance is \(3+7=10.00\ \mathrm{L/h}\); impaired clearance is \(3+7(0.40)=5.80\ \mathrm{L/h}\). Thus \(R_{\mathrm{AUC}}=10/5.80\approx1.724\), assuming, among other conditions, unchanged bioavailability and nonrenal clearance. (3) Ordinal categories rank impairment but do not establish equal biological spacing; different physiological processes may also change differently across categories. (4) The extrapolation is not supported by that mechanistic model as stated. Reassess the severity range and assumptions, and report the conflict and uncertainty rather than treating either prediction as universally valid.