BQL Records During Absorption

Why BQL Timing Matters During Absorption

After an oral dose, concentrations commonly rise as drug is absorbed, reach a peak, and then decline. A below-quantification (BQL) record during the rising part of the profile can be entirely compatible with this pattern: soon after dosing, the predicted concentration may still be below the assay’s lower limit of quantification (LLOQ). But a BQL record at a time when the model predicts a concentration far above the LLOQ is different. It is evidence against that prediction, not a measured concentration of zero.

The timing of BQL records can therefore help evaluate absorption predictions. The evidence depends on the predicted concentration at each sampling time, the assay limit, and the observation model. It does not, on its own, identify an absorption rate or prove that the structural PK model is wrong. Dose, bioavailability, lag time, individual variability, assay performance, and sampling-time accuracy can also matter.

Definition: A BQL record is an observation reported as below the assay’s LLOQ. It indicates that the concentration was not quantified at or above that limit; it is not a numerical measurement equal to zero or to the LLOQ. For a predicted concentration \(C_{\mathrm{pred}}(t)\), its interpretation depends on the probability that the observation model would produce a BQL record at time \(t\).

Connecting BQL Timing to an Oral Absorption Curve

For a single oral dose with first-order absorption and elimination in a one-compartment model, assume no lag time, constant parameters, and an absorption rate constant \(k_a\) greater than the elimination rate constant \(k_e\). At time \(t>0\), the predicted concentration is

C_{\mathrm{pred}}(t)=\frac{F D k_a}{V(k_a-k_e)} \left(e^{-k_e t}-e^{-k_a t}\right),

where \(D\) is dose, \(F\) is bioavailability, and \(V\) is apparent volume of distribution. The rate constants have units of inverse time, and concentration units are determined by \(D/V\). This expression describes a rising concentration followed by a decline. Its time of peak is

t_{\max}=\frac{\log(k_a/k_e)}{k_a-k_e}.

This peak time is a model prediction, not necessarily the time of the largest observed sample. Sparse sampling can miss the peak, and individual profiles can differ from a typical population prediction.

Interpretation: Before the predicted peak, an early BQL record may be expected if the predicted concentration is below or near the LLOQ. A BQL record near the predicted peak is more challenging when the model predicts a concentration well above the LLOQ. The comparison is probabilistic: residual error means that an observation can differ from its prediction.

From Predicted Concentration to BQL Probability

For a concrete probability calculation, suppose the residual observation model is lognormal:

C_{\mathrm{obs}}(t)=C_{\mathrm{pred}}(t)e^{\epsilon}, \qquad \epsilon\sim N(0,\sigma_{\epsilon}^{2}).

Here \(C_{\mathrm{pred}}(t)\) is the conditional median observation under this model, and \(\sigma_{\epsilon}\) is the standard deviation of log-scale residual error. The model assumes positive concentrations and a common log-scale residual standard deviation for the observation being evaluated. With an LLOQ \(L\), a BQL record means \(C_{\mathrm{obs}}(t)<L\), so

P(\mathrm{BQL}\mid C_{\mathrm{pred}}(t)) = \Phi\left( \frac{\log\{L/C_{\mathrm{pred}}(t)\}}{\sigma_{\epsilon}} \right),

where \(\Phi\) is the standard normal cumulative distribution function. This formula concerns the observation probability conditional on the stated prediction and residual model. If the prediction is based on typical population parameters, it does not automatically account for between-subject variability; that distinction matters when interpreting population-level predictions.

Worked Example: An Early BQL Record Followed by a Conflicting One

Consider a hypothetical single-dose oral profile with \(k_a=1.2\ \mathrm{h}^{-1}\), \(k_e=0.20\ \mathrm{h}^{-1}\), and \(FD/V=20\ \mathrm{mg/L}\). Assume the lognormal residual model above with \(\sigma_{\epsilon}=0.30\), and an LLOQ of \(0.50\ \mathrm{mg/L}\). Compare predictions at \(t=0.01\) h and \(t=0.50\) h.

The concentration model becomes \(C_{\mathrm{pred}}(t)=24(e^{-0.20t}-e^{-1.2t})\ \mathrm{mg/L}\), because \(20(1.2)/(1.2-0.20)=24\). At \(0.01\) h,

C_{\mathrm{pred}}(0.01) =24(e^{-0.002}-e^{-0.012}) \approx 24(0.998002-0.988072) \approx 0.2383\ \mathrm{mg/L}.

This is below the LLOQ. The BQL probability is

\Phi\left( \frac{\log(0.50/0.2383)}{0.30} \right) = \Phi(2.471) \approx 0.993.

An early BQL record is therefore unsurprising under these assumptions. At \(0.50\) h, the predicted concentration is

C_{\mathrm{pred}}(0.50) =24(e^{-0.10}-e^{-0.60}) =24(0.904837-0.548812) \approx 8.545\ \mathrm{mg/L}.

The corresponding BQL probability is

\Phi\left( \frac{\log(0.50/8.545)}{0.30} \right) = \Phi(-9.45),

which is effectively zero at ordinary reporting precision. Thus, the same BQL label has very different implications at the two times. A record at \(0.01\) h agrees with the predicted low concentration; one at \(0.50\) h would be highly inconsistent with this particular prediction and residual model. The latter calls for checking assumptions and data as well as reconsidering the model.

Why the Peak-Time Formula Holds

The relationship between \(k_a\), \(k_e\), and the predicted peak follows by differentiating the concentration curve. The constant multiplier does not affect the time at which the curve reaches its maximum. Setting the derivative to zero gives

\frac{dC_{\mathrm{pred}}(t)}{dt}=0 \quad\Longleftrightarrow\quad k_a e^{-k_a t}=k_e e^{-k_e t}.

Rearranging yields \(e^{(k_a-k_e)t}=k_a/k_e\). Taking logarithms and solving for \(t\) gives \(t_{\max}=\log(k_a/k_e)/(k_a-k_e)\). Under the stated condition \(k_a>k_e>0\), this time is positive. This derivation explains why a change in the predicted absorption rate can shift the period during which BQL records are expected.

Worked Example: Comparing Fast and Slow Absorption Predictions

Suppose two hypothetical model predictions share \(k_e=0.20\ \mathrm{h}^{-1}\) and \(FD/V=20\ \mathrm{mg/L}\), but differ in absorption rate: one has \(k_a=1.2\ \mathrm{h}^{-1}\), the other \(k_a=0.25\ \mathrm{h}^{-1}\). At \(t=1.0\) h, compare their predictions against an LLOQ of \(5.0\ \mathrm{mg/L}\), using \(\sigma_{\epsilon}=0.30\).

For \(k_a=1.2\ \mathrm{h}^{-1}\), the multiplier is \(20(1.2)/(1.2-0.20)=24\ \mathrm{mg/L}\), giving

C_{\mathrm{pred}}(1.0) =24(e^{-0.20}-e^{-1.20}) =24(0.818731-0.301194) \approx 12.42\ \mathrm{mg/L}.

Its BQL probability is \(\Phi\{\log(5.0/12.42)/0.30\}=\Phi(-3.032)\approx0.0012\). For \(k_a=0.25\ \mathrm{h}^{-1}\), the multiplier is \(20(0.25)/(0.25-0.20)=100\ \mathrm{mg/L}\), so

C_{\mathrm{pred}}(1.0) =100(e^{-0.20}-e^{-0.25}) =100(0.818731-0.778801) \approx 3.993\ \mathrm{mg/L}.

The BQL probability under this slower-absorption prediction is \(\Phi\{\log(5.0/3.993)/0.30\}=\Phi(0.749)\approx0.773\). Therefore, a BQL record at one hour is much more compatible with the slower-absorption prediction in this calculation. It does not establish that \(k_a=0.25\ \mathrm{h}^{-1}\) is correct: the comparison holds other quantities and the residual model fixed, and those assumptions may not describe the data.

Worked Example: A BQL Record Near the Predicted Peak

Return to the hypothetical parameters \(k_a=1.2\ \mathrm{h}^{-1}\), \(k_e=0.20\ \mathrm{h}^{-1}\), and \(FD/V=20\ \mathrm{mg/L}\). First calculate the predicted peak time:

t_{\max} =\frac{\log(1.2/0.20)}{1.2-0.20} =\log(6) \approx 1.792\ \mathrm{h}.

At \(1.8\) h, close to that peak, the prediction is

C_{\mathrm{pred}}(1.8) =24(e^{-0.36}-e^{-2.16}) =24(0.697676-0.115325) \approx13.98\ \mathrm{mg/L}.

For an LLOQ of \(5.0\ \mathrm{mg/L}\) and \(\sigma_{\epsilon}=0.30\), the probability of BQL is \(\Phi\{\log(5.0/13.98)/0.30\}=\Phi(-3.428)\approx0.0003\). A BQL record at this time is consequently difficult to reconcile with the specified prediction. If a sample at \(2.5\) h were quantified at, for example, \(11\ \mathrm{mg/L}\), that later result would not erase the conflict at \(1.8\) h. The full sequence, its times, and the observation model all matter; the later quantified value alone cannot turn the earlier BQL record into a zero or a quantified concentration.

Assessing the Timing Pattern

Interpret BQL records in relation to the predicted curve, not simply as a count. A useful review aligns each sampling time with the model-predicted concentration and the assay LLOQ. It also checks that the time is measured from the correct dose, that dose and sample timestamps are reliable, and that the applicable assay limit is known.

Locate each record in time.
Determine whether it is early after dosing, near the predicted rising-phase peak, or after the peak. Verify dose timing and any relevant lag or dosing history.

Compare prediction with the LLOQ.
Use the model prediction at that specific time. A prediction below the LLOQ makes BQL plausible; a prediction far above it makes BQL less plausible under the stated observation model.

Review the sequence and assumptions.
Consider nearby quantified samples, individual variability, residual error, assay differences, and the possibility of delayed or otherwise misspecified absorption. Do not infer an absorption parameter from one BQL record alone.

A sequence of early BQL records followed by quantified concentrations can be consistent with absorption beginning after dosing, but the transition time is only informative relative to the assay limit and the model’s uncertainty. BQL records near a predicted peak can challenge the model more strongly when the prediction is much greater than the LLOQ. Conversely, a predicted concentration only slightly above the limit may still yield a substantial BQL probability.

Comparisons, Diagnostics, and Limits

A censored observation model uses the information that a record fell below the LLOQ without claiming to know its exact concentration. Deleting the record discards that information. Substituting zero or a fixed fraction of the LLOQ turns a threshold statement into an invented numerical measurement, potentially distorting the apparent rise and the estimated absorption profile. These approaches do not make the same statistical use of the data.

The probability calculation also depends on its assumptions. A lognormal residual model may be inappropriate for a particular assay or concentration range; the stated residual standard deviation may be wrong; and a typical prediction may not represent the individual whose sample was collected. If random effects are important, the individual’s concentration distribution can differ materially from a curve evaluated only at typical parameters. A BQL observation should therefore prompt comparison with the appropriate predictive distribution, not a mechanical conclusion from the fixed-effects curve alone.

Unexpected early BQL records: check whether the model predicts concentrations too high early, but also verify dose time, sample time, formulation, lag assumptions, and assay limit.
BQL near a predicted peak: compare the predicted concentration-to-LLOQ ratio and residual uncertainty. A large ratio with small residual error makes the record harder to explain.
Quantified samples surrounding a BQL record: inspect the order and spacing of all observations. A single record may reflect measurement variability, while a repeated pattern may indicate a systematic mismatch.
Apparent absorption conclusions: remember that \(k_a\), \(F\), dose, and other model features can affect predicted concentrations. Timing alone does not isolate one parameter.

The main pitfall is to read “BQL at 0.5 h” as a concentration of zero, or to claim that any early BQL record proves absorption is slow. The correction is to treat the record as a threshold-censored observation and ask how probable it is at that time under a specified model, assay limit, and error distribution. The timing provides evidence, but its strength depends on those conditions.

Key takeaway: BQL observations before and around the expected oral peak are informative through their timing relative to predicted concentrations and the LLOQ. They are not zero values, and their interpretation is conditional on the PK and observation models.

Check Your Understanding

Answer each question, then compare with the solutions below.

A model predicts \(0.20\ \mathrm{mg/L}\) at an early sample, with an LLOQ of \(0.50\ \mathrm{mg/L}\). Is a BQL record necessarily evidence of slow absorption?
For \(k_a=1.0\ \mathrm{h}^{-1}\) and \(k_e=0.25\ \mathrm{h}^{-1}\), calculate \(t_{\max}\).
Under the lognormal residual model, what is the BQL probability when \(C_{\mathrm{pred}}=L\)?
Why does a BQL record near a predicted peak carry different information from one just after dosing?
Name two checks to make before interpreting a BQL record that conflicts with the model prediction.

Solutions. (1) No. A prediction below the LLOQ makes BQL compatible with the model, but does not by itself establish slow absorption; dose timing, bioavailability, assay limit, and variability also matter. (2) \(t_{\max}=\log(1.0/0.25)/(1.0-0.25)=\log(4)/0.75\approx1.848\ \mathrm{h}\). (3) The standardized value is \(\log(L/L)/\sigma_{\epsilon}=0\), so the probability is \(\Phi(0)=0.5\). (4) Soon after dosing, the model may predict a concentration below the LLOQ, whereas near the peak it may predict a concentration well above it; the latter makes BQL less probable under the same residual model. (5) For example, verify dose and sample timestamps, confirm the assay-specific LLOQ, and review the residual-error and absorption assumptions.

What You'll Learn