Discriminating Liver Models With Flow Perturbation Data

Why Perturb Hepatic Flow?

A clearance measurement at one hepatic-flow condition can be consistent with more than one liver model, particularly if model-specific inputs are allowed to vary. Deliberately evaluating predictions across different flow conditions creates a more informative comparison: the models can predict different changes in extraction and clearance even when they use the same effective intrinsic clearance. The design question is therefore not simply how many measurements to collect, but which flow conditions make the predictions separate relative to measurement uncertainty.

The comparison is meaningful only if the perturbation changes hepatic blood flow while other relevant inputs remain stable or are measured and updated appropriately. In particular, an apparent difference between predictions is not evidence for a model if the perturbation also changes unbound fraction, intrinsic metabolic capacity, or the blood-to-plasma relationship and those changes are ignored.

Common Scale and Model Assumptions

Let \(Q\) be hepatic blood flow, \(CL_{\mathrm{int}}\) intrinsic clearance on a consistent blood-concentration basis, and \(f_{u,b}\) the unbound fraction in blood. Define the effective intrinsic clearance \(x=f_{u,b}CL_{\mathrm{int}}\). Both \(Q\) and \(x\) have units of volume per time. The hepatic extraction ratio is \(E=CL_H/Q\), where \(CL_H\) is hepatic clearance; thus \(CL_H=QE\). Comparing extraction ratios as well as clearances helps separate a model's extraction prediction from the direct scaling by \(Q\).

The basic assumptions are that the reported clearance is hepatic clearance, the system is operating in the linear range, and the measured flow represents the flow relevant to the model. The comparisons also assume that \(x\) does not change during the flow perturbation. For the dispersion model, a fixed \(D_N\) is an additional modeling assumption, not an automatic consequence of changing flow. If physiology implies that dispersion changes with flow, that relationship needs to be specified before comparing predictions.

With \(r=x/Q\), the well-stirred model gives \(E=r/(1+r)\); the parallel-tube model gives \(E=1-\exp(-r)\). In the dispersion model, define \(a=\sqrt{1+4D_Nr}\). Its outlet-to-inlet concentration ratio and extraction are:

These equations assume the model structure and parameter definitions just stated; they are not interchangeable formulas for the same physical assumption. At low \(r\), all three models predict similar extraction to first order, so low-extraction conditions often provide little discrimination. Conditions around intermediate extraction can make differences more visible. At very high extraction, clearance approaches the flow limit, so clearance values can again be close in absolute terms even if extraction ratios remain informative.

Why the Predictions Differ

The common scale \(r=x/Q\) makes the comparison transparent. For a well-stirred liver, a single mixed compartment has \(Q(C_{\mathrm{in}}-C_{\mathrm{out}})=xC_{\mathrm{out}}\). Rearranging gives \(C_{\mathrm{out}}/C_{\mathrm{in}}=Q/(Q+x)\), and therefore \(E=x/(Q+x)\). For a parallel tube, if \(z\) is distance along a tube scaled from 0 to 1, the concentration obeys \(dC/dz=-rC\). Integrating from inlet to outlet gives \(C_{\mathrm{out}}/C_{\mathrm{in}}=\exp(-r)\), hence \(E=1-\exp(-r)\). The dispersion expression is the corresponding outlet solution for axial dispersion under its specified boundary conditions.

The different concentration profiles lead to different extraction responses at the same \(r\). The well-stirred model represents complete mixing; the parallel-tube model represents flow through tubes with an axial concentration decline; the dispersion model represents axial mixing between these idealized patterns, with its behavior depending on \(D_N\). Flow perturbation is useful because changing \(Q\) changes \(r\) when \(x\) is fixed, moving the models along their distinct response curves.

Designing a Flow-Perturbation Comparison

A defensible design starts by calculating predictions, not by selecting conditions solely because they are experimentally convenient. Specify plausible values or estimates for \(x\) and, for the dispersion model, \(D_N\). Then calculate each model's \(E\) and \(CL_H\) over candidate flow conditions. Prefer a range that moves \(r\) through regions where predictions separate, rather than collecting several conditions that all lie in a nearly indistinguishable low-extraction range.

Assessing Measurements and Recognizing Failure Modes

Once measurements are available, compare them with predictions made using the same stated inputs and the same measurement scale. A simple descriptive comparison is the sum of squared standardized differences, \(\sum_i[(CL_{H,i}^{\mathrm{obs}}-CL_{H,i}^{\mathrm{pred}})/SE_i]^2\), where \(SE_i\) is the standard error for the clearance measurement at condition \(i\). This quantity makes discrepancies relative to their reported uncertainty; it is not, by itself, a universal model-selection test. Its interpretation also depends on whether the predictions were fixed in advance or their parameters were estimated from the same observations.

For example, the following hypothetical observations are compared with the predictions from the preceding example. Suppose each clearance has a measurement standard error of 1.5 L/h. The standardized sum is calculated for each model using its three predictions.

The main alternatives are to compare models at a single flow or to estimate separate model-specific inputs from each condition. A single-flow comparison may be unable to distinguish models because different assumptions can yield similar clearance at that point. Allowing each condition to have unconstrained values of \(x\) can also absorb the very flow-response differences the design is intended to examine. A useful comparison therefore asks whether a common, plausible input set predicts the response across conditions, while making any parameter estimation explicit.

Failure modes include inaccurate flow measurements, clearance estimates that include substantial extrahepatic elimination, and flow perturbations that change \(x\) or the dispersion behavior. Another warning is model separation smaller than the uncertainty of the observations. In such cases, a numerical ranking can look decisive while the data do not reliably distinguish the predictions. The comparison is also limited by the selected range of flow, the assumed values and constancy of \(D_N\), and the possibility that none of the three model structures adequately represents the organ.

A practical conclusion should therefore be conditional: state which model's predictions agree most closely under the assumptions and inputs used, how large the differences are relative to uncertainty, and whether the perturbation plausibly isolated flow. Do not convert better agreement in one hypothetical or limited dataset into a causal claim about liver architecture.