Why Treatment Implementation Belongs in the Prediction
A QSP simulation may describe a prescribed regimen precisely while actual treatment is interrupted by missed doses or discontinuation. If the model receives every scheduled dose regardless of treatment implementation, its predictions describe a different input process from the one patients experience. The practical task is to represent uncertain implementation explicitly, then propagate it through exposure, mechanistic states, and the clinical endpoint.
Adherence and discontinuation are not merely adjustments to the final response. They alter the timing and amount of drug entering the system. In a nonlinear model, two patients with the same total delivered dose can have different trajectories if their dose timing differs. Likewise, a population-average adherence rate does not generally determine the average endpoint.
In a simple model, \(A_k\) takes the value 1 when a dose is taken and 0 when it is missed. If discontinuation is permanent, \(R_k\) is 1 until discontinuation and 0 thereafter. More detailed representations can allow partial doses, temporary interruptions, restarts, or distinct reasons for stopping. The representation should match the clinical implementation question and the information available to support it.
From Treatment Events to Clinical Predictions
For a fixed adherence and discontinuation pattern, the QSP model receives the resulting sequence \(U_1,\ldots,U_K\). It then evolves exposure and biological states under that sequence and maps those states to the endpoint. If \(\omega\) denotes a particular treatment pattern and \(Y\) is the endpoint, the prediction across uncertain patterns is a mixture of the conditional predictions for each \(\omega\).
The sum is written for a discrete set of patterns; an integral is used when the uncertain process is continuous. The endpoint mean follows the same conditioning principle: \(E[Y]=\sum_{\omega}E[Y\mid\omega]P(\omega)\). This is an average of model predictions over patterns, not a claim that one average pattern occurred. In particular, when the mechanistic mapping is nonlinear, evaluating the model at the average delivered dose need not reproduce the average prediction.
A simulation can implement this directly: draw an adherence/discontinuation pattern from its specified distribution, construct delivered doses, run the QSP model, and record the endpoint. Repeating the process approximates the predictive distribution. Assumptions about dose-taking must be stated before simulation. For example, independent Bernoulli indicators with \(P(A_k=1)=p\) assume the same probability at each occasion and no serial dependence. Those assumptions are not implied simply by knowing an average adherence proportion.
Worked Example: Uncertain Adherence Propagated to Response
Consider a hypothetical patient scheduled for five equal doses. Let \(A_k\) be independent Bernoulli indicators with \(P(A_k=1)=p\). Assume a simplified mechanistic endpoint \(Y=1-\exp(-0.4N)\), where \(N=\sum_{k=1}^{5}A_k\) is the number of doses taken. The coefficient 0.4 is in inverse dose-count units, so \(Y\) is a dimensionless response between 0 and 1. Suppose there are two equally weighted plausible adherence propensities, \(p=0.5\) and \(p=0.9\).
For a fixed \(p\), \(E[\exp(-0.4N)]\) can be calculated using independence. Each dose contributes a factor of 1 when missed and \(\exp(-0.4)\) when taken, so
Using \(\exp(-0.4)\approx0.670320046\), for \(p=0.5\) the conditional mean is \(1-(0.5+0.5(0.670320046))^5\approx0.593698443\). For \(p=0.9\), it is \(1-(0.1+0.9(0.670320046))^5\approx0.827945449\).
The two adherence propensities have equal weight, so the overall mean prediction is
This is the mean response over both the uncertainty in \(p\) and the dose-taking events conditional on \(p\). It is not the response from a patient who takes a fraction of a dose at each occasion. The assumed independence and constant \(p\) are simplifications; if missed doses cluster, this calculation no longer describes the stated process.
Representing Discontinuation Explicitly
A useful representation of permanent discontinuation is a stopping time \(T\): doses are available before \(T\), and no doses are delivered from \(T\) onward. Alternatively, a discrete-time hazard \(h_k\) can specify the probability of discontinuing at occasion \(k\), conditional on still being on treatment immediately beforehand. If that hazard is constant, then the probability of remaining on treatment through \(m\) occasions is \((1-h)^m\), provided the hazard applies at each occasion under the same conditions.
Adherence conditional on remaining on treatment and discontinuation need not be independent. For instance, a pattern of missed doses may predict stopping. If data or clinical knowledge support such a relationship, the simulation should encode it rather than draw the two processes independently by default. The model should also specify whether discontinuation is permanent and whether missed doses can be followed by catch-up dosing; these are distinct treatment rules.
Worked Example: Two Doses With a Discontinuation Risk
In a hypothetical two-occasion regimen, each prescribed dose is one unit. The first dose is taken with probability 0.8. Before the second occasion, the patient discontinues with probability 0.25; otherwise the patient remains on treatment, and takes the second dose with probability 0.8. Assume all these events are independent except for the stated rule that the second dose is available only if treatment continues. Let \(N\) be the total number of doses taken and let \(Y=1-\exp(-0.5N)\).
The probability of \(N=0\) is \(0.2\) if the first dose is missed and the patient either discontinues or remains on treatment but also misses the second dose: \(0.2(0.25)+0.2(0.75)(0.2)=0.08\). The probability of \(N=1\) includes taking only the first dose, or missing the first and taking the second: \(0.8(0.25)+0.8(0.75)(0.2)+0.2(0.75)(0.8)=0.44\). Taking both doses has probability \(0.8(0.75)(0.8)=0.48\). These probabilities sum to 1.
Using \(1-e^{-0.5}\approx0.393469\) and \(1-e^{-1}\approx0.632121\), the mean is \(0.44(0.393469)+0.48(0.632121)\approx0.4765\). The calculation includes both missed-dose events and permanent discontinuation. Treating everyone as if they received the prescribed two doses would instead predict \(1-e^{-1}\approx0.6321\), which answers a different question.
Dependence Between Missed Doses Matters
A marginal adherence probability describes the chance that a particular scheduled dose is taken, but not how dose-taking varies across occasions. Two processes can have the same marginal adherence and very different trajectories. Clustering of missed doses can create longer exposure gaps; if exposure or response recovers slowly, the clinical consequences may differ from scattered missed doses.
Worked Example: Same Average Adherence, Different Dose Patterns
Compare two hypothetical two-dose processes, each with a 0.5 probability of taking any particular dose. In the first, the patient takes both doses with probability 0.5 and misses both with probability 0.5. In the second, the two doses are taken independently with probability 0.5 each. Both have an expected total of one dose. Suppose response is \(Y=1-\exp(-0.4N)\), with \(N\) the number taken.
For the first, perfectly clustered process, \(N\) is 0 or 2 with equal probability. Therefore, \(E[Y]=0.5(0)+0.5(1-e^{-0.8})\approx0.275336\). For the independent process, the probabilities of \(N=0,1,2\) are 0.25, 0.50, and 0.25. Thus,
The difference arises from the distinct distributions of dose patterns, not from a difference in mean adherence. For a real QSP model, the pattern may also change the shape and timing of the exposure trajectory, so the endpoint should be calculated by simulating the delivered-dose sequence rather than by relying on the total dose alone.
Simulation Workflow and Model Checks
A defensible simulation makes the treatment implementation assumptions visible and carries them through the same mechanistic model used for other clinical predictions. The following sequence is useful when the aim is a predictive distribution under uncertain adherence and discontinuation.
Specify scheduled occasions, prescribed doses, adherence indicators, discontinuation timing, and any allowed interruptions or restarts.
State which quantities are uncertain, their distributions or scenario weights, and whether missed doses, discontinuation, and patient characteristics are dependent.
For each simulated pattern, calculate the actual input at every occasion, respecting the treatment rules after discontinuation.
Simulate exposure and mechanistic states under the realized input sequence, then calculate the prespecified clinical endpoint.
Check simulated adherence rates, discontinuation frequencies and times, exposure trajectories, and endpoint distributions against the assumptions and available evidence.
Diagnostics should begin with the treatment process, not just the final endpoint. Confirm that the simulated proportion of doses taken and the discontinuation frequency match their targets within Monte Carlo error. Plot or tabulate treatment trajectories to detect impossible doses after permanent discontinuation, unintended catch-up doses, or gaps caused by coding errors. Compare results under plausible dependence assumptions when serial patterns are poorly known. A highly skewed or multimodal endpoint distribution can be a genuine consequence of distinct treatment paths, but it can also signal a mixture or implementation error; inspect the contributing paths before interpreting it.
| Approach | What it represents | Main caution |
|---|---|---|
| Deterministic average dose | Uses a fractional or average input in every simulation. | Can miss variability and nonlinear effects of discrete dose events. |
| Independent dose indicators | Draws each dose-taking event separately using specified probabilities. | Rules out serial dependence unless dependence is added explicitly. |
| Explicit pattern simulation | Draws dose-taking and discontinuation paths, then simulates each path. | Requires defensible assumptions about event probabilities and dependence. |
| Scenario analysis | Compares specified adherence or discontinuation patterns. | Scenario weights are needed for a combined predictive distribution. |
The deterministic average-dose approach may be adequate when dose delivery is effectively continuous, the response is approximately linear over the relevant range, and only a mean prediction is needed. Explicit pattern simulation is generally more informative when treatment is administered in discrete doses, response is nonlinear, or endpoint risk depends on exposure timing. Scenario analysis is valuable when the evidence does not support a single probability model, but results should then be reported by scenario or combined only with justified weights.
Common Pitfall and Limitations
- Replacing event patterns with an average dose. The reasoning “average adherence is 70%, so give every simulated patient 70% of each dose” removes the variation and timing of missed doses. This is not generally equivalent to simulating actual treatment events through a nonlinear model.
- Assuming independence without saying so. A per-dose adherence probability does not establish that missed doses are independent over time or independent of discontinuation.
- Applying discontinuation only to the endpoint. Stopping changes later treatment inputs and therefore may change the entire exposure and mechanistic trajectory.
- Over-interpreting poorly supported patterns. If adherence data are sparse, detailed estimates of serial dependence may not be identifiable. Present assumptions and compare plausible alternatives rather than implying the process is known precisely.
Predictions remain conditional on the treatment-process model, the QSP model, and the clinical endpoint definition. Simulating more patterns reduces Monte Carlo error but does not correct misspecified adherence probabilities, omitted dependence, or unrealistic stopping rules. If a treatment policy changes adherence or discontinuation, its pattern distribution may differ from the one used to generate the prediction; the prediction should not be transferred without checking that assumption.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- If each of four scheduled doses is taken independently with probability 0.75, what is the probability of taking all four?
- Why can two adherence processes with the same mean adherence yield different mean clinical responses?
- A permanent discontinuation occurs before dose 3. Which scheduled doses should be delivered after that event under a no-restart rule?
- When is a deterministic average-dose simulation most likely to approximate explicit event simulation adequately?
Solutions. (1) Independence gives \(0.75^4=0.3164\). (2) The processes can have different dose-pattern distributions and timing, and a nonlinear exposure-response mapping need not give the same average response for those distributions. (3) No later doses are delivered; with the stated rule, doses 3 and onward are omitted. (4) It is most defensible when dose delivery is effectively continuous, response is approximately linear over the range considered, and the target is a mean prediction rather than the distribution of outcomes.