Why Diagnose Pathways From the Profile?
A concentration-time profile can suggest that a dose enters the systemic circulation through more than one route or process. For example, one portion may be absorbed relatively rapidly while another enters more slowly. The resulting curve may rise quickly at first and then continue rising, flatten, or show a shoulder. But a bend in a plotted curve is not, by itself, evidence for two pathways: measurement noise, sparse sampling, and other absorption models can produce similar appearances.
The purpose of diagnosis is therefore not to label every irregular curve as a mixture. It is to ask whether a parallel-pathway model gives a coherent explanation of the observed profile, whether the available samples can reveal the relevant features, and whether simpler explanations remain plausible. The earlier material on parallel absorption pathways provides the model structure; here the emphasis is on interpreting profile evidence and its limits.
Reference Profile and Its Assumptions
Consider a single oral dose divided between two immediate first-order absorption pathways. Let \(D\) be the dose, \(F\) the overall bioavailability, \(V\) the apparent volume of distribution, \(f\) the fraction assigned to the rapid pathway, \(k_{a,r}\) and \(k_{a,s}\) the rapid and slow absorption rate constants, and \(k_e\) the elimination rate constant. The fractions are \(f\) and \(1-f\), with \(0\leq f\leq1\). All rate constants have units of inverse time.
For this reference profile, assume a single well-mixed disposition compartment, first-order elimination, linear disposition, first-order absorption in each pathway, and a common disposition process after absorption. Both pathways begin at the dose time, there is no lag, and the input fractions together account for the dose. These assumptions are a diagnostic reference, not facts guaranteed by the observed curve. For distinct absorption and elimination rates, the predicted concentration is
Each bracketed term is the concentration contribution from one pathway. Their sum is the predicted total concentration. When an absorption rate equals \(k_e\), the corresponding fraction in this expression has a zero denominator; its continuous limit is used instead, giving a term proportional to \(t e^{-k_e t}\). In practice, a fitted model must handle that limiting case correctly rather than evaluate the singular-looking expression directly.
The equation follows from adding the contributions of the two inputs. For a pathway receiving fraction \(g\) of the dose with absorption rate \(k_a\), the input rate at time \(s\) is \(F D g k_a e^{-k_a s}\). Under one-compartment first-order elimination, a unit amount entering at time \(s\) contributes a fraction \(e^{-k_e(t-s)}/V\) to concentration at time \(t\). Integrating input over all entry times from zero to \(t\) gives
The total profile is obtained by setting \(g=f\) for the rapid pathway and \(g=1-f\) for the slow pathway, then adding the two contributions. This derivation also explains an important diagnostic limitation: the observed concentration is a combined result of input and elimination, not a direct measurement of either pathway’s input rate.
Profile Features That Can Support a Mixture
A rapid-and-slow mixture may produce a steep early increase followed by continued accumulation from the slow component. Depending on the rate constants, dose fractions, and sampling times, this can appear as a shoulder, a flatter segment near the peak, or a change in the rate of decline. No one feature is required. The relative contributions overlap in time, and the total concentration can remain smooth even when the inputs have distinct rates.
A useful diagnostic is whether the pattern is coherent across several observations. For example, a model with one absorption rate may repeatedly underpredict the early rise and overpredict concentrations around a later shoulder, while a two-pathway model reduces both patterns. A single high or low residual is weaker evidence: it may result from ordinary measurement variation or an isolated data problem.
Residuals are the observed concentrations minus the model-predicted concentrations at the corresponding times. Their size should be judged relative to the observation-error scale, and their pattern across time matters. Residuals that alternate around zero without a time-related trend are generally less suggestive of a missing pathway than a sequence of deviations that consistently follows a profile feature. A more flexible model will often fit the data better simply because it has more parameters; improved fit alone does not establish that the extra pathway is needed.
Worked Example: A Coherent Profile Pattern
Suppose hypothetical data are collected after a single dose. For illustration, take \(FD/V=10.0\) mg/L, \(f=0.600\), \(k_{a,r}=2.00\) h\(^{-1}\), \(k_{a,s}=0.250\) h\(^{-1}\), and \(k_e=0.200\) h\(^{-1}\). Assume the parallel-pathway model and an additive observation-error standard deviation of \(0.150\) mg/L for this diagnostic illustration. The observed values below are hypothetical.
| Time (h) | Observed concentration (mg/L) | Two-pathway prediction (mg/L) | Residual (mg/L) |
|---|---|---|---|
| 0.25 | 2.590 | 2.534 | 0.056 |
| 0.50 | 3.950 | 4.027 | -0.077 |
| 1.00 | 5.430 | 5.355 | 0.075 |
| 2.00 | 5.500 | 5.622 | -0.122 |
| 4.00 | 4.550 | 4.622 | -0.072 |
At \(t=1.00\) h, the rapid-pathway contribution is \(10.0(0.600)(2.00/1.80)(e^{-0.200}-e^{-2.00})=4.556\) mg/L, and the slow-pathway contribution is \(10.0(0.400)(0.250/0.0500)(e^{-0.200}-e^{-0.250})=0.7986\) mg/L. Their unrounded sum is approximately \(5.3546\) mg/L, which rounds to \(5.355\) mg/L. The rapid component contributes about \(4.556/5.3546=85.1\%\) of that predicted concentration at this time.
For example, at \(t=0.25\) h, the model prediction is approximately \(2.534\) mg/L. The rapid contribution is \(6.6667(e^{-0.0500}-e^{-0.500})=2.298\) mg/L, and the slow contribution is \(20.00(e^{-0.0500}-e^{-0.0625})=0.2363\) mg/L. The sum is \(2.534\) mg/L. At the other listed times, the model predictions are approximately \(4.027\), \(5.355\), \(5.622\), and \(4.622\) mg/L, respectively.
The residuals range from \(-0.122\) to \(0.075\) mg/L. Dividing each by the illustrative \(0.150\) mg/L error standard deviation gives standardized residuals of approximately \(0.37\), \(-0.51\), \(0.50\), \(-0.81\), and \(-0.48\). These values do not show an obvious large, time-ordered discrepancy from this candidate profile. They are compatible with the two-pathway prediction under the stated error scale, but do not demonstrate that two pathways are necessary. A fair model comparison would also fit a one-pathway alternative and examine its residual pattern and uncertainty.
Why Sparse Sampling Can Mislead
The most informative samples for distinguishing a rapid-and-slow mixture are often early samples on the rising limb, together with enough later samples to characterize continued input and the decline. If sampling begins after the rapid input has largely occurred, the evidence for that component may be missed. Conversely, a few isolated early samples may not establish a shoulder or separate its shape from noise.
The example below uses the same hypothetical model parameters as above. Its pathway contributions show why later concentrations do not directly reveal how many inputs generated them.
Worked Example: What Later Samples Can and Cannot Show
At \(t=4\) h, the rapid-pathway contribution is \(6.6667(e^{-0.800}-e^{-8.00})=2.993\) mg/L, and the slow-pathway contribution is \(20.00(e^{-0.800}-e^{-1.00})=1.629\) mg/L. The total is approximately \(4.622\) mg/L. At \(t=8\) h, the corresponding contributions are \(6.6667(e^{-1.600}-e^{-16.00})=1.346\) mg/L and \(20.00(e^{-1.600}-e^{-2.00})=1.331\) mg/L, giving approximately \(2.677\) mg/L. At \(t=12\) h, they are approximately \(0.6048\) and \(0.8186\) mg/L, giving \(1.423\) mg/L.
Imagine that only the 4, 8, and 12 h samples were collected. These observations describe a declining profile, but they provide no direct observation of the early rise or peak. The contributions at 4 h still include concentration resulting from the rapid input that occurred earlier; they do not mean that rapid absorption is continuing at 4 h. With no early samples, the data may be unable to distinguish a rapid-and-slow mixture from a suitably adjusted alternative model. This is a limitation of the sampling design, not evidence that the rapid pathway is absent.
The calculation checks the predictions at those times against the same two-pathway equation. It does not establish that those rates or fractions could be estimated from only three late observations. Estimability depends on the full model, the unknown parameters, observation variability, and the actual sample schedule.
Noise, Alternative Explanations, and Diagnostics
An apparent kink in a profile may be caused by a measurement error, a mistimed sample, or ordinary observation variability. In a simple diagnostic exercise, compare the discrepancy with the expected error scale, but do not treat a single standardized residual as a formal test of a pathway. In a fitted analysis, inspect residuals over time, model-predicted profiles, and parameter uncertainty; also ask whether the fitted rates and fractions are stable when reasonable data or model choices change.
The next example illustrates why an isolated outlier is not enough to establish a second input component. Suppose the two-pathway prediction at 0.50 h is \(4.027\) mg/L, as in the earlier hypothetical calculation.
Worked Example: An Isolated High Observation
Assume the additive observation-error standard deviation is \(0.150\) mg/L and a hypothetical observation at 0.50 h is \(4.600\) mg/L. Its residual relative to the stated prediction is \(4.600-4.027=0.573\) mg/L. In units of the assumed error standard deviation, this is \(0.573/0.150=3.82\).
This observation deserves investigation because it is large relative to the illustrative error scale. But one high value does not identify a rapid pathway: the candidate model already contains one, and a single discrepancy cannot distinguish an unusual measurement from model misspecification or other causes. Check the sample time and record, examine nearby observations and replicate measurements if available, and determine whether a systematic profile pattern remains. If the rest of the profile agrees with a simpler model and the discrepancy is isolated, the isolated point is not persuasive evidence for an additional pathway.
Several sources of non-uniqueness deserve particular attention:
- Similar absorption rates: if \(k_{a,r}\) and \(k_{a,s}\) are close, their component curves may be difficult to separate. The sum can look much like a single pathway with an intermediate rate.
- A small pathway fraction: if \(f\) is near zero or one, the component assigned a very small fraction may have little visible effect on the total profile.
- Unobserved early input: if the first sample is late, the size and rate of a rapid component can be hard to distinguish from overall exposure or other model features.
- Slow absorption and elimination: a slowly declining concentration tail can reflect continued absorption as well as elimination. A terminal slope should not automatically be interpreted as \(k_e\) when the input may still be contributing.
- Competing model structures: transit, lag-time, or extended-input models can sometimes describe features that a two-rate mixture also describes. The curve shape alone may not select a unique mechanism.
Comparing Candidate Explanations
Begin with a plausible simpler model, such as one first-order absorption pathway, and compare it with the parallel-pathway candidate under the same assumptions about disposition and observation error. A useful comparison asks whether the more complex model removes a structured residual pattern, whether its extra parameters are estimated with meaningful precision, and whether the resulting profile is plausible over the sampled time range. A lower residual sum alone is insufficient: additional parameters can improve fit even when the apparent improvement is mostly noise.
Other absorption structures should be considered when they have a plausible rationale. A transit model represents distributed delay through a sequence of transit stages; a lag-time model shifts the onset of input; and zero-order input spreads a portion of the dose over time. These structures can create profile features resembling a mixture. They should be compared as alternative explanations, not combined automatically just because each can improve fit.
Evidence is strongest when the sampling schedule covers the profile features that separate the candidates, the parallel model improves the fit in a time-coherent way, and the pathway parameters are not excessively uncertain or unstable. If both models describe the observed concentrations comparably, report that the data do not clearly distinguish them rather than choosing the more elaborate explanation by default.
Common Mistake and Key Takeaway
Common mistake: “The concentration curve has a shoulder, so the dose must have entered through two absorption pathways.” The shoulder may instead be sampling variation, a single pathway observed sparsely, or a feature compatible with another input model. The correction is to treat the curve shape as a diagnostic clue, check whether it recurs across observations, assess it relative to measurement error, and compare plausible candidate models.
Check Your Understanding
Answer each question using the diagnostic principles above. Concise calculations are shown in the solutions.
- In the reference model, what does \(f\) represent, and what restriction must it satisfy?
- A model predicts \(3.20\) mg/L at a sample time, and the hypothetical observation is \(3.50\) mg/L. If the additive error standard deviation is \(0.10\) mg/L, calculate the residual and its standardized value. Does this single observation prove a second pathway?
- Why can samples collected only during the declining portion of a profile provide weak evidence about a rapid absorption component?
- If the two estimated absorption rates are nearly equal, what happens to the diagnostic value of a two-pathway fit?
Solutions. (1) \(f\) is the fraction assigned to the rapid pathway; the other fraction is \(1-f\), and \(0\leq f\leq1\). (2) The residual is \(3.50-3.20=0.30\) mg/L, and the standardized residual is \(0.30/0.10=3.00\). It is large relative to the stated error scale, but by itself it does not prove a second pathway; check the surrounding profile and possible data issues. (3) The rapid input and early rise may have occurred before the first sample. Later concentrations combine prior input with elimination and may be compatible with several candidate models. (4) The component profiles become similar, so their sum can resemble a single pathway; the extra rates and fractions may be poorly distinguished even if the two-pathway model fits well.