Baseline Drift and Turnover-Parameter Confounding

Why Baseline Drift Matters

A turnover model often begins with the assumption that untreated production is constant. That assumption gives a useful reference: when production and loss balance, the biomarker remains at a stable baseline. But untreated biomarkers can drift because of time trends, disease progression, circadian structure, or changes in study conditions. If the model forces input to remain constant, it may explain this drift by changing the estimated turnover rate instead. The resulting fit can describe the observations reasonably while assigning the trend to the wrong mechanism.

The issue is not merely whether a profile rises or falls. A biomarker concentration reflects both how quickly material enters the measured pool and how quickly it leaves. When only the concentration is observed, a time-varying input and a different turnover rate can compensate for one another. The modeler must therefore consider baseline behavior as part of the structural assumptions, rather than treating the untreated baseline as automatically fixed.

Definition: In a one-pool turnover model, let $B(t)$ be biomarker concentration, $R(t)$ the input or production rate, and $k$ the first-order loss rate. The mass-balance equation is $$ \frac{dB(t)}{dt}=R(t)-kB(t). $$ Here $R(t)$ has units of concentration per time, $k$ has units of inverse time, and $B(t)$ has units of concentration. With constant input $R_0$, the steady-state concentration is $B_{\mathrm{ss}}=R_0/k$.

The constant-input model assumes that $R(t)=R_0$ over the modeled interval and that $k$ is constant. A time-varying-input model instead specifies a function for $R(t)$, while retaining the same turnover equation. Both are mechanistic descriptions; neither is automatically correct simply because it can fit observed concentrations. The relevant question is whether its assumptions are plausible and whether the available observations distinguish its parameters.

How a Drifting Input Changes the Profile

Suppose untreated input grows exponentially, $R(t)=R_0e^{gt}$, where $g$ is the fractional rate of change in input. For $k+g\ne0$, the solution starting at concentration $B(0)=B_0$ is

B(t)=\frac{R_0}{k+g}e^{gt} +\left(B_0-\frac{R_0}{k+g}\right)e^{-kt}.

The first term follows the changing input; the second is a transient determined by the initial condition and loss rate. If $g=0$, the solution reduces to the familiar constant-input profile. If $g\ne0$, the eventual profile need not approach a fixed steady state: a continuously growing input can produce a continuing rise in $B(t)$.

A short derivation makes the source of the two terms explicit. Multiply the differential equation by the integrating factor $e^{kt}$, so the left side becomes the derivative of $e^{kt}B(t)$. Integrating from 0 to $t$ gives

e^{kt}B(t)-B_0 =R_0\int_0^t e^{(k+g)s}\,ds =R_0\frac{e^{(k+g)t}-1}{k+g}.

Dividing by $e^{kt}$ and collecting terms gives the stated solution. The derivation assumes the input function and turnover rate are correctly specified over the interval. It does not show that an exponential trend is the right biological model; it shows what that assumption implies.

Worked Example: Comparing a Drifting Input With Constant Input

Consider a hypothetical untreated biomarker with $B_0=20.0$ concentration units (CU), a loss rate $k=0.500\ \mathrm{day}^{-1}$, and initial input $R_0=10.0\ \mathrm{CU/day}$. Suppose input grows as $R(t)=10.0e^{0.100t}\ \mathrm{CU/day}$, with time in days. The initial input equals $kB_0=10.0\ \mathrm{CU/day}$, so the initial rate of change is zero, even though input subsequently increases.

Using the derived profile, $$ B(t)=\frac{10}{0.500+0.100}e^{0.100t} +\left(20-\frac{10}{0.600}\right)e^{-0.500t} =16.6667e^{0.100t}+3.3333e^{-0.500t}. $$ At day 2, $$ B(2)=16.6667e^{0.200}+3.3333e^{-1.000} =20.3567+1.2263=21.5830\ \mathrm{CU}. $$ At day 4, $$ B(4)=16.6667e^{0.400}+3.3333e^{-2.000} =24.8638+0.4511=25.3149\ \mathrm{CU}. $$

Now compare a constant-input model with $R_0=10.0\ \mathrm{CU/day}$, the same initial concentration, and a fitted loss rate $k'=0.388\ \mathrm{day}^{-1}$. Its steady-state value is $10/0.388=25.7732\ \mathrm{CU}$. Its day-2 prediction is $$ B_{\mathrm{fit}}(2)=\frac{10}{0.388} +\left(20-\frac{10}{0.388}\right)e^{-0.388(2)} =23.117\approx23.12\ \mathrm{CU}. $$ At day 4 it predicts $25.7732+(20-25.7732)e^{-1.552}=24.550\ \mathrm{CU}$. These predictions do not reproduce the time-varying-input profile at both times: the constant model is above it at day 2 and below it at day 4. A fitted rate can alter the shape, but it cannot be interpreted as a biological change in loss solely because it helps accommodate a drifting untreated profile.

Why Input and Turnover Can Be Confounded

The mass-balance equation itself gives a direct identifiability argument. If a differentiable concentration profile $B(t)$ were known without error, then for any proposed constant loss rate $k$, the input that makes the profile satisfy the model is

R_k(t)=\frac{dB(t)}{dt}+kB(t).

Thus, when input is allowed to vary freely and is not measured independently, different values of $k$ can produce exactly the same concentration profile by implying different input functions. Concentration data alone do not resolve this ambiguity without additional structure, information, or design. In practice, sparse and noisy measurements make the problem harder because the derivative of $B(t)$ is not directly observed.

Worked Example: Two Turnover Rates, One Biomarker Profile

Suppose a hypothetical observed profile is $B(t)=20e^{0.050t}\ \mathrm{CU}$, with time in days. Its derivative is $dB/dt=0.050B(t)\ \mathrm{CU/day}$. If the turnover rate is $k=0.200\ \mathrm{day}^{-1}$, the input required by the mass-balance equation is $$ R_{0.200}(t)=(0.050+0.200)B(t)=0.250B(t)\ \mathrm{CU/day}. $$ If instead $k=0.500\ \mathrm{day}^{-1}$, the input required is $$ R_{0.500}(t)=(0.050+0.500)B(t)=0.550B(t)\ \mathrm{CU/day}. $$

Both input functions generate exactly the same $B(t)$, despite implying different loss rates and different production rates. At day 0, for example, the respective inputs are $5.00$ and $11.0\ \mathrm{CU/day}$. Neither rate can be selected on the basis of this concentration trajectory alone if the input is otherwise unconstrained. Independent turnover information, direct input measurements, or defensible restrictions on how input changes are needed to distinguish these explanations.

This exact equivalence is the limiting case of a broader practical problem. A constant-input fit can appear adequate over a short interval when the drift is small relative to measurement noise. Conversely, a flexible time-varying input can absorb a trend that the data do not support as a distinct mechanism. A good fit is therefore not sufficient evidence that the turnover estimate is unbiased or that the chosen input function is biologically credible.

Checking the Baseline Assumption

When untreated measurements are available, inspect them over the same time scale and under the same sampling conditions as the treated observations. A systematic trend in untreated values is evidence against a fixed baseline, although it does not by itself identify whether production, loss, or another process caused the trend. Also check whether the apparent change aligns with a known time effect, such as study day or time of day, rather than with drug exposure. A trend in an untreated group should not be attributed to drug action.

A practical model comparison can begin with a constant-input reference and a prespecified time-varying-input alternative. Compare their fitted profiles at the actual observation times, examine parameter uncertainty and correlations, and check whether the added drift parameter is supported by the data. The chosen trend function should be simple enough to estimate and biologically interpretable. A linear trend may be reasonable over a limited interval; an exponential trend is appropriate only when proportional growth or decline is plausible over that interval. Extrapolating either form far beyond the observed period is risky.

Worked Example: Estimating an Untreated Linear Drift

Suppose hypothetical untreated concentrations are $20.0$, $21.0$, and $22.0\ \mathrm{CU}$ on days 0, 2, and 4. For this illustration, represent the observed baseline trend by the exact straight line through these values. Its slope is $$ \frac{22.0-20.0}{4-0}=0.500\ \mathrm{CU/day}, $$ so $B(t)=20.0+0.500t\ \mathrm{CU}$ over the observed interval.

If an independently supported turnover rate is $k=0.250\ \mathrm{day}^{-1}$, the input implied by this profile is $$ R(t)=\frac{dB(t)}{dt}+kB(t) =0.500+0.250(20.0+0.500t) =5.50+0.125t\ \mathrm{CU/day}. $$ The implied input rises from $5.50\ \mathrm{CU/day}$ on day 0 to $6.00\ \mathrm{CU/day}$ on day 4. This is a mass-balance implication conditional on the chosen $k$, not proof that production truly changed linearly.

For comparison, a constant input fixed at the initial implied value, $R=5.50\ \mathrm{CU/day}$, predicts $$ B(t)=\frac{5.50}{0.250} +\left(20.0-\frac{5.50}{0.250}\right)e^{-0.250t} =22.0-2.0e^{-0.250t}. $$ At day 2 this gives $20.787\ \mathrm{CU}$, and at day 4 it gives $21.264\ \mathrm{CU}$, rather than the observed $21.0$ and $22.0\ \mathrm{CU}$. The comparison shows how a constant-input model can miss a sustained drift even when it starts with the correct initial concentration and input.

Modeling Choices, Diagnostics, and Limitations

The main alternatives differ in how they obtain information about the baseline and turnover:

Approach	Useful when	Main limitation
Constant input and constant turnover	Untreated baseline is plausibly stable over the observation window	Unmodeled drift may be absorbed into turnover or other parameters
Prespecified time-varying input	There is a plausible, simple baseline trend and enough data to estimate it	Flexible input can be weakly identified and can mimic changes in turnover
External or independently estimated turnover	Washout or other relevant data provide credible information on loss	Turnover may differ across conditions or populations; external estimates carry uncertainty

Useful diagnostics include plotting untreated observations against time, checking whether a constant-input model leaves sustained deviations in the same direction, and comparing predictions from the competing models at observed times. Examine uncertainty and parameter correlations: a wide interval for $k$, strong dependence between $k$ and drift parameters, or unstable estimates under small changes in the observation window all warn that the data may not separate mechanisms. Residual patterns can flag misspecification, but they cannot uniquely identify its biological cause.

The assumptions behind these checks matter. The simple equation supposes a single well-mixed biomarker pool, first-order loss, and no unmodeled transfers that materially affect the measured concentration. If there are linked precursor pools, nonlinear loss, or feedback, an apparent baseline trend may reflect omitted dynamics rather than a changing input alone. Likewise, a “time-varying input” is a modeling representation; a study-time trend does not establish a causal biological mechanism.

A common mistake is to interpret a lower fitted $k$ as evidence that biomarker degradation slowed whenever the observed baseline rises. That reasoning assumes input is constant, which is precisely the assumption at issue. The correction is to treat the rate estimate as conditional on the input model, assess untreated trends, and use independent turnover or input information where possible. Do not select a more flexible baseline function merely because it reduces deviations: the drift form must also be plausible and estimable.

Key takeaway: Observed biomarker concentration is the net result of input and loss. When untreated input may drift, a fitted turnover parameter is conditional on the assumed baseline model; concentration profiles alone may not distinguish the two mechanisms.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

For a constant input of $8.0\ \mathrm{CU/day}$ and $k=0.400\ \mathrm{day}^{-1}$, what is the steady-state concentration?
If a profile is rising at $0.300\ \mathrm{CU/day}$ when $B=15.0\ \mathrm{CU}$, what input is implied by $k=0.200\ \mathrm{day}^{-1}$?
Why can an observed concentration trajectory alone fail to identify $k$ when the input is unrestricted?
Name one diagnostic sign that a constant-input baseline may be inadequate, and state one limitation of that sign.

Solutions. (1) $B_{\mathrm{ss}}=R_0/k=8.0/0.400=20.0\ \mathrm{CU}$. (2) The mass-balance equation gives $R=dB/dt+kB=0.300+(0.200)(15.0)=3.30\ \mathrm{CU/day}$. (3) For any proposed $k$, choosing $R(t)=dB/dt+kB(t)$ reproduces the same profile, so additional constraints or information are needed. (4) Sustained same-direction deviations from a constant-input fit, or a trend in untreated measurements, can indicate inadequacy. Neither establishes that production changed: omitted pool dynamics, measurement issues, or other time effects can also create a trend.