Why Interruptions Need Their Own Model
A dose hold is not generally a reset to zero. Drug remaining in the body can continue to drive pharmacology after the last dose, and a delayed effect can persist even as plasma concentration falls. Meanwhile, response may recover, remain altered, or move toward a different level during the drug-free interval. When treatment restarts, both residual drug and the state of the response at that time matter.
A useful representation is a piecewise dosing history: doses occur before the hold, none occur during the hold, and dosing resumes at a specified time. The PK and PD states continue to evolve across all three periods. The earlier material on cumulative and recency-weighted dosing history provides ways to summarize past dosing; here the central task is to predict the linked concentration and response trajectories across an interruption, without discarding the states already accumulated.
Piecewise PK and Carryover
For a simple one-compartment model with first-order elimination, let \(C(t)\) be concentration at time \(t\), and let \(k\) be the elimination-rate constant. During a hold that begins at time \(t_h\), with concentration \(C_h\), the model assumes no further drug input and predicts exponential decline. The half-life is \(t_{1/2}=\ln(2)/k\). At restart, a dose changes the concentration according to the assumed route and PK model; elimination continues throughout.
Here \(t_r\) is the restart time. This formula assumes a one-compartment disposition, constant \(k\), and no input during the interval. For an instantaneous IV dose with concentration increment \(\Delta C\), the concentration immediately after the dose is the concentration immediately before it plus \(\Delta C\). For non-instantaneous input, such as an infusion or oral absorption, the input process must instead be represented in the PK model. In either case, residual concentration at restart is part of the initial condition.
Worked Example: Residual Concentration and a Restart Dose
Suppose, hypothetically, concentration is \(8.00\) mg/L when a hold begins. The elimination half-life is \(12.0\) hours. After a \(24.0\)-hour hold, an instantaneous dose adds \(6.00\) mg/L to concentration. Calculate concentration just before restart and \(12.0\) hours afterward.
The elimination rate is \(k=\ln(2)/12.0=0.05776\) per hour. Since \(24.0\) hours is two half-lives, residual concentration just before restart is
The dose adds \(6.00\) mg/L, so concentration immediately afterward is \(2.00+6.00=8.00\) mg/L. Twelve hours later, one half-life after the restart dose,
If the model incorrectly reset concentration to zero at restart, it would predict \(6.00(0.50)=3.00\) mg/L at that time. The \(1.00\) mg/L difference is the decayed contribution from drug already present when dosing resumed. These calculations rely on the stated one-compartment, first-order elimination model and the stipulated instantaneous concentration increment.
Concentration and Pharmacologic Effect Can Have Different Histories
A concentration decline does not necessarily imply an equally rapid decline in pharmacologic effect. One standard representation is an effect compartment, with concentration \(C_e(t)\) approaching plasma concentration \(C(t)\) at rate \(k_{e0}\). During a hold, \(C_e(t)\) can remain appreciable because it reflects both its starting value and the changing plasma concentration. The equations below assume constant \(k\) and \(k_{e0}\), and no direct effect of the hold on either parameter.
If \(C(t)=C_h e^{-k(t-t_h)}\), the solution for \(k_{e0}\ne k\) is
This solution separates two contributions: decay of the effect-compartment state present at the start of the hold, and the effect-compartment response to plasma drug that remains during the hold. It does not claim that the effect compartment is a measured physical space; it is a model for delayed equilibration.
Worked Example: Delayed Effect During a Drug-Free Interval
Use the previous hypothetical concentration history, with \(C_h=C_e(t_h)=8.00\) mg/L, \(k=\ln(2)/12.0=0.05776\) per hour, and \(k_{e0}=0.100\) per hour. Calculate \(C_e\) after \(24.0\) hours without dosing, then calculate inhibition using \(I=C_e/(EC_{50}+C_e)\), where \(EC_{50}=4.00\) mg/L.
The plasma concentration after \(24.0\) hours is \(2.00\) mg/L. The effect-compartment solution gives, using unrounded exponential values in the calculation,
The displayed exponentials and parameters are rounded; the final value is calculated from the unrounded values. The predicted inhibition is
Thus the model predicts about \(48.34\%\) inhibition at that time, despite plasma concentration being \(2.00\) mg/L. Substituting plasma concentration directly into the same inhibition function would give \(2.00/(4.00+2.00)=0.3333\), or \(33.33\%\). That alternative is appropriate only if the direct-effect assumption is supported; it is not interchangeable with the stated delayed-effect model.
Response May Recover or Persist During the Holiday
The response state can have its own dynamics. For an indirect-response example, let \(Y(t)\) be a measured response, \(k_{in}\) a production rate, and \(k_{out}\) a first-order loss rate. Suppose drug inhibits production by fraction \(I(t)\), so that \(dY/dt=k_{in}[1-I(t)]-k_{out}Y(t)\). During a drug-free interval, setting \(I(t)=0\) makes the response move toward the drug-free steady level \(Y_0=k_{in}/k_{out}\), rather than jump immediately to it.
This expression follows by subtracting \(Y_0\) from both sides of the differential equation during the hold: \(d\{Y-Y_0\}/dt=-k_{out}\{Y-Y_0\}\). Solving this first-order equation gives exponential recovery toward \(Y_0\). The result assumes constant turnover and no other time-varying influences on response during the interval. A model for a response that persists after drug removal would use different state dynamics, such as a slower recovery rate or an additional persistent state.
Worked Example: Recovery and Response After Reinitiation
Consider a hypothetical biomarker with drug-free level \(Y_0=100\) units and \(k_{out}=\ln(2)/48=0.01444\) per hour. At the start of a \(48.0\)-hour holiday, \(Y=60.0\) units. First find the response at the end of the holiday. Then suppose treatment restarts and rapidly establishes a constant inhibition of \(0.500\). Predict \(Y\) after a further \(24.0\) hours.
During the holiday, inhibition is assumed to be zero. The response has moved one recovery half-life toward \(Y_0\):
After restart, the assumed constant inhibition makes the production rate \(k_{in}(1-0.500)\). The new steady response is therefore \(Y_{ss}=0.500k_{in}/k_{out}=50.0\) units, given \(Y_0=k_{in}/k_{out}=100\) units. Starting from \(80.0\) units, the response approaches that new level with the same \(k_{out}\):
The calculation assumes that inhibition becomes \(0.500\) immediately and remains constant after restart. A PK-driven model would instead calculate the changing inhibition from concentration or effect-site concentration, and then solve the response equation with that time-varying input. The response at restart, \(80.0\) units, remains the starting condition; it is not replaced by the new steady level.
Model Choice, Diagnostics, and Limits
The main alternatives differ in what states they carry through the interruption. A direct-effect model links response to current concentration; it is simple, but implies no pharmacologic delay beyond the response relationship itself. An effect-compartment model allows delayed equilibration between plasma and effect. An indirect-response or turnover model carries a response state that changes gradually, so recovery and re-suppression need not track concentration. A model with persistent pharmacology may also be needed when the response remains altered after concentrations become negligible. These are scientific alternatives, not merely different ways to calculate the same prediction.
| Model representation | What carries through the hold | Key implication |
|---|---|---|
| Direct effect | Residual concentration | Effect follows concentration without an additional delay state |
| Effect compartment | Residual concentration and \(C_e\) | Effect may lag behind falling concentration |
| Turnover response | Response \(Y\), as well as PK states | Response recovers or re-suppresses gradually |
| Persistent-effect model | An additional slow or persistent state | Response can remain altered after drug exposure wanes |
Predictions across a hold depend on assumptions about event timing, the amount of drug remaining, elimination, effect delay, and response recovery. They are especially sensitive to poorly estimated half-lives or recovery rates when the holiday is short relative to those time scales. A model can fit observations during uninterrupted treatment yet predict interruptions poorly if the hold-period dynamics are not informed by data.
Useful checks include comparing predicted and observed concentrations immediately before and after the hold, and examining whether response residuals show systematic deviations during the holiday or after restart. A persistent pattern of overprediction during the holiday may indicate that the model retains too much drug effect or assumes too slow a recovery. Underprediction after restart may indicate that residual drug, delayed effect, or response-state carryover has been omitted. Such patterns are clues, not unique diagnoses: different parameter combinations or model structures can produce similar discrepancies.
- Check the time origin and event boundaries. State whether a concentration is evaluated just before or just after the restart dose; those values can differ.
- Check state continuity. Unless the model explicitly includes a discontinuous biological change, concentration, effect, and response states should be propagated to the hold and restart times rather than reset.
- Check the assumed holiday mechanism. A zero input does not guarantee zero concentration, zero pharmacologic effect, or immediate biological recovery.
- Check whether the observations can distinguish models. Sparse measurements during a hold may not separate rapid concentration decay from slow pharmacodynamic recovery.
A common mistake is to set both concentration and response to baseline at the restart time because dosing was interrupted. That reasoning confuses a change in input with a reset of the system. The correction is to stop input during the hold, propagate all relevant states under the hold-period equations, and use their predicted values as initial conditions after reinitiation. A true state reset requires a separate biological or treatment mechanism and supporting evidence.
These models also have limits. A simple exponential decline may not describe multicompartment distribution, nonlinear elimination, or a metabolite with a different time course. A fixed recovery rate may not apply across patients or disease states. If interruption timing or doses are uncertain, predictions conditional on a specified history will not capture that uncertainty. Finally, a response model fitted only to on-treatment data may not identify recovery behavior; extrapolation through a holiday should be presented as conditional on the assumed dynamics, not as a directly established fact.
Check Your Understanding
Answer each question, then compare with the short solutions below.
- A drug has a half-life of \(10.0\) hours. If concentration is \(12.0\) mg/L at the start of a \(20.0\)-hour hold, what concentration does a one-compartment first-order model predict at the end?
- Why can effect remain substantial when plasma concentration has fallen during a holiday?
- A response follows turnover dynamics and begins a holiday below its drug-free steady level. In which direction does it move during the hold under the stated zero-inhibition model?
- What initial response value should be used immediately after restart in a continuous turnover model?
- Name one diagnostic pattern that could indicate the model's holiday recovery dynamics are inadequate, and explain why it is not definitive proof of a particular alternative model.
Solutions. (1) Two half-lives elapse, so \(12.0(0.5)^2=3.00\) mg/L. (2) An effect-compartment state can lag behind plasma concentration, and a response state can also evolve more slowly than concentration. (3) It moves upward toward the drug-free steady level, because \(dY/dt=k_{in}-k_{out}Y\) is positive when \(Y<k_{in}/k_{out}\). (4) Use the response predicted or observed at the end of the hold; restarting the input does not itself reset the response state. (5) Systematic response residuals during the holiday may signal inadequate recovery dynamics, but several mechanisms or parameter errors can produce that pattern, so it does not identify a unique replacement model.