Why Compare Models at Intermediate Extraction?
At intermediate extraction, hepatic clearance is neither very close to blood flow nor very small relative to it. In this range, different assumptions about how blood mixes within the liver can yield noticeably different predictions even when the inputs are identical. Comparing the predictions makes the practical consequence of those assumptions visible: the difference is not a change in the drug’s intrinsic properties, but a change in the representation of the organ.
The comparison below uses blood flow, blood unbound fraction, and intrinsic clearance as shared inputs. It focuses on the standard well-stirred, parallel-tube, and dispersion formulations. All numerical values are hypothetical and illustrate the calculations; they are not estimates from a clinical study.
The common ratio \(x\) isolates the combined effect of binding and intrinsic clearance relative to flow. It allows a direct comparison: if \(Q_h\), \(f_{u,b}\), and \(CL_{int}\) are held fixed, any difference in predicted \(CL_h\) arises from the model formulation and its mixing assumption.
The Three Formulations and Their Mixing Assumptions
In the well-stirred model, blood entering the liver is assumed to mix rapidly and completely with blood in the organ. The liver is represented by a uniform concentration, so the concentration relevant to elimination is the outlet concentration. In the parallel-tube model, blood flows through parallel channels with no longitudinal back-mixing; concentration falls along each channel as drug is removed. The dispersion model represents intermediate axial mixing: blood can mix along the flow direction, but the liver is not treated as either perfectly stirred or as a set of entirely unmixed plug-flow channels.
Here \(D_N\) is the dimensionless dispersion number under this formulation. A finite positive value represents intermediate axial mixing; as \(D_N\) approaches zero, the dispersion prediction approaches the parallel-tube limit, and as \(D_N\) becomes large, it approaches the well-stirred limit. A value such as \(D_N=0.17\) can be used as an illustrative conventional value, but it is not a universal constant for every liver or application.
These equations assume steady hepatic flow, linear (nonsaturable) elimination over the concentrations considered, and a homogeneous organ representation within each model. They do not explicitly represent such features as distinct blood-flow regions, shunting, or transporter processes. The models use the same effective inputs here; changing those inputs between formulations would no longer isolate the effect of mixing assumptions.
Why the Well-Stirred and Parallel-Tube Predictions Differ
The contrast between the two limiting mixing assumptions follows directly from mass balance. In the well-stirred model, elimination equals the rate at which drug is delivered to and removed from the organ. With inlet concentration \(C_{in}\) and uniform outlet concentration \(C_{out}\), this balance is \(Q_h(C_{in}-C_{out})=f_{u,b}CL_{int}C_{out}\). In the parallel-tube model, concentration falls continuously along a channel according to the local extraction rate. The derivation shows why the same \(x\) produces different outlet fractions.
Divide the balance by \(Q_hC_{in}\) and use \(x=f_{u,b}CL_{int}/Q_h\). Then \(1-C_{out}/C_{in}=x(C_{out}/C_{in})\), so \(C_{out}/C_{in}=1/(1+x)\). Therefore \(E_h=1-C_{out}/C_{in}=x/(1+x)\).
Let \(u\) be the fraction of total flow that has passed along a channel, from 0 at its inlet to 1 at its outlet. Under the plug-flow assumption, \(dC/du=-xC\). Separating variables and integrating gives \(\ln(C_{out}/C_{in})=-x\), hence \(C_{out}/C_{in}=e^{-x}\) and \(E_h=1-e^{-x}\).
For \(x>0\), \(e^x\geq 1+x\), so \(e^{-x}\leq 1/(1+x)\). Subtracting both sides from 1 gives \(E_{h,\mathrm{PT}}\geq E_{h,\mathrm{WS}}\). Thus parallel-tube clearance is at least as high as well-stirred clearance for these shared positive inputs.
The dispersion model supplies a prediction between these limiting mixing cases for the finite dispersion value used in the comparison. It is not obtained by simply averaging the other two clearances: its prediction follows from its own mixing formulation.
Worked Example: Three Clearances at Intermediate Extraction
Suppose a hypothetical liver has \(Q_h=90.0\) L/h, \(f_{u,b}=0.100\), and \(CL_{int}=900\) L/h. Use \(D_N=0.170\) for the dispersion calculation. The common ratio is
The inputs are dimensionally consistent: \(f_{u,b}\) is dimensionless, and \(x\) is a ratio of clearances. An \(x\) of 1 gives intermediate extraction rather than a flow-limited or very low-extraction extreme.
Well-stirred. \(E_{h,\mathrm{WS}}=1.00/(1+1.00)=0.500\), so \(CL_{h,\mathrm{WS}}=(90.0\ \mathrm{L/h})(0.500)=45.0\ \mathrm{L/h}\).
Parallel-tube. \(E_{h,\mathrm{PT}}=1-e^{-1.00}=0.6321\), so \(CL_{h,\mathrm{PT}}=(90.0\ \mathrm{L/h})(0.6321)=56.89\ \mathrm{L/h}\).
Dispersion. First calculate \(a=\sqrt{1+4(0.170)(1.00)}=\sqrt{1.68}=1.296\). Substitution into the dispersion equation gives
Thus \(CL_{h,\mathrm{Disp}}=(90.0\ \mathrm{L/h})(0.5884)=52.96\ \mathrm{L/h}\). With these fixed inputs and the stated dispersion number, the dispersion prediction lies between the two limiting predictions, as expected.
Quantifying the Spread and Its Consequences
A simple measure of prediction spread is the range, defined here as the largest predicted clearance minus the smallest. In the example, the largest prediction is \(56.89\) L/h and the smallest is \(45.0\) L/h, giving a range of \(56.89-45.0=11.89\) L/h. The dispersion prediction is \(52.96\) L/h, which is \(7.96\) L/h above the well-stirred prediction and \(3.93\) L/h below the parallel-tube prediction. The range summarizes disagreement; it is not a confidence interval or a measure of sampling uncertainty.
Worked Example: Comparing Hepatic Availability
Use the same hypothetical inputs to calculate hepatic availability under each model. Since \(F_h=1-E_h\), the well-stirred model gives \(F_{h,\mathrm{WS}}=1-0.500=0.500\), the parallel-tube model gives \(F_{h,\mathrm{PT}}=1-0.6321=0.3679\), and the dispersion model gives \(F_{h,\mathrm{Disp}}=1-0.5884=0.4116\).
The spread in predicted hepatic availability is \(0.500-0.3679=0.1321\), or 13.21 percentage points. This comparison concerns the fraction escaping hepatic extraction only. For example, interpreting \(F_h=0.4116\) as total oral bioavailability would require additional assumptions and information about processes outside this hepatic model.
The clearance range has a practical interpretation: when inputs are held constant, the two limiting assumptions predict different degrees of extraction, while intermediate dispersion gives a result between them in this example. If a modeling decision depends on distinguishing predictions separated by this amount, model structure may matter to the application. If the application is insensitive to a difference of this size, a simpler representation may be adequate. The comparison alone does not determine which assumption best represents a particular liver.
Worked Example: Reading a Hypothetical Clearance Observation
Suppose, purely for illustration, that an independent hypothetical estimate of hepatic clearance is \(53.0\) L/h. The signed prediction error, defined as model prediction minus this estimate, is \(-8.0\) L/h for the well-stirred model, \(56.89-53.0=+3.89\) L/h for the parallel-tube model, and \(52.96-53.0=-0.04\) L/h for the dispersion model. The dispersion prediction is numerically closest in this comparison.
That arithmetic does not establish that dispersion is the correct liver model. The observation’s uncertainty, the appropriateness of the shared input values, and whether the model assumptions fit the application all matter. A close match for one hypothetical clearance value is not, by itself, evidence that the model will predict accurately in other conditions.
Checks, Failure Modes, and Model Choice
A useful comparison starts by checking the inputs and the output scale. \(x\) must be dimensionless; \(Q_h\), \(CL_{int}\), and \(CL_h\) must use compatible volume-per-time units; and \(f_{u,b}\) must be expressed as a fraction rather than as a percentage. For positive finite \(x\), each model should give \(0\leq E_h<1\) and \(0\leq CL_h<Q_h\). The well-stirred and parallel-tube formulas should also satisfy the ordering derived above. A violated check often points to a unit error, an incorrect formula substitution, or a numerical calculation problem.
For the dispersion calculation, check the definition and value of \(D_N\) used by the implementation: dispersion-number conventions must match the equation. For very small or very large values, direct evaluation of exponentials and subtractions can lose numerical precision. Use a numerically stable implementation and verify that the result approaches the appropriate limiting model. A result outside the plausible range should prompt checks of the formula, parameter convention, and numerical method before any physiological interpretation.
| Formulation | Mixing representation | What the comparison means |
|---|---|---|
| Well-stirred | Rapid, complete mixing; uniform organ concentration | A limiting prediction based on outlet concentration representing the liver |
| Parallel-tube | Parallel channels with no longitudinal back-mixing | A limiting prediction based on concentration decline along the channels |
| Dispersion | Intermediate axial mixing, indexed by \(D_N\) | A prediction that depends on the chosen dispersion number and its mixing formulation |
These alternatives are structural representations, not three independent measurements of the same quantity. The earlier material on hepatic-clearance formulations and flow perturbations provides context for interpreting why models can diverge; here, the essential comparison is at one shared set of intermediate-extraction inputs. The spread should not be mistaken for a probability distribution over the true clearance.
A common mistake is to say, “The three models use the same intrinsic clearance, so they should predict the same hepatic clearance.” The reasoning overlooks that intrinsic clearance is not the only determinant of the prediction: how blood concentration is represented inside the liver also changes the relationship between intrinsic clearance and extraction. The correction is to hold the inputs fixed, calculate each formulation separately, and interpret the resulting spread as a consequence of model structure. Conversely, choosing a model solely because it gives the closest value to one observation ignores uncertainty and the model’s intended assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- For \(Q_h=100\) L/h, \(f_{u,b}=0.20\), and \(CL_{int}=500\) L/h, calculate \(x\) and the well-stirred extraction ratio.
- For any \(x>0\), which predicts greater clearance, the well-stirred or parallel-tube model? State the inequality that justifies your answer.
- In the numerical example, calculate the range of the three predicted hepatic availabilities.
- Why does a close match between one observed clearance and a dispersion prediction not prove that dispersion is the best model?
- Name one unit or boundary check that can reveal an implementation error.
Solutions. (1) \(x=(0.20)(500)/100=1.00\), so \(E_{h,\mathrm{WS}}=1/(1+1)=0.500\). (2) Parallel-tube predicts greater or equal clearance because \(e^x\geq1+x\) implies \(1-e^{-x}\geq x/(1+x)\); equality is approached at \(x=0\). (3) The largest availability is \(0.500\) and the smallest is \(0.3679\), so the range is \(0.1321\), or 13.21 percentage points. (4) One match does not establish predictive performance across conditions, account for observation uncertainty, or validate the model assumptions. (5) For positive finite inputs, check that \(x\) is dimensionless and that \(0\leq CL_h<Q_h\); for example, a predicted clearance above flow signals a problem under these model assumptions.