PK/PD Models for Time-to-Event Outcomes

1. What Is a Time-to-Event PK/PD Model?

Many pharmacodynamic endpoints are not measured as a continuous response at a fixed time. Instead, the scientific question is when an event occurs: disease progression, relapse, seizure, rescue medication, treatment failure, death, or another clinically defined event.

These outcomes are commonly analyzed as time-to-event outcomes. The event time is informative, but some subjects may not experience the event during observation. Those observations are typically treated as right-censored.

A PK/PD model for a time-to-event endpoint connects drug exposure to the instantaneous risk of experiencing the event.

2. The Survival Analysis Framework

Let \(T\) denote the time until an event occurs. The two central quantities are the survival function and the hazard function.

Survival function

The survival function is the probability that the event has not occurred by time \(t\):

Thus, \(S(t)=0.80\) means that 80% of the population is predicted to remain event-free beyond time \(t\).

Hazard function

The hazard describes the instantaneous event rate among subjects who have remained event-free immediately before time \(t\):

3. Linking PK Exposure to Event Hazard

A common starting point is a proportional-hazards model in which the baseline hazard is modified by drug exposure.

Here:

• \(h(t)\) is the hazard at time \(t\).
• \(h_0(t)\) is the baseline hazard when the exposure effect is zero.
• \(C(t)\) is drug concentration predicted by the PK model.
• \(g(C(t))\) is the exposure-response function.
• \(\beta\) determines the magnitude and direction of the exposure effect.

The simplest model uses concentration directly:

But a linear concentration effect is not always scientifically appropriate. Saturable exposure-response relationships can be incorporated into the hazard model as well.

4. Emax Relationships for Time-to-Event Outcomes

If increasing concentration eventually produces little additional effect, an \(E_{\max}\) model can describe the exposure-response relationship.

That exposure effect can then enter a hazard model. For example:

The negative sign represents a model in which increasing exposure reduces the hazard. A positive sign could instead represent increasing hazard with increasing exposure.

The parameter \(EC_{50}\) represents the concentration producing half of the modeled maximum effect, while \(E_{\max}\) controls the maximum magnitude of the exposure effect on the log-hazard scale.

5. What Is the Baseline Hazard?

The baseline hazard \(h_0(t)\) describes how event risk changes over time before accounting for the modeled exposure effect.

Several approaches are possible.

Choosing a baseline hazard is important because exposure effects and baseline time trends can otherwise be difficult to distinguish.

6. The Weibull Time-to-Event PK/PD Model

The Weibull model is particularly useful because it allows the baseline hazard to change over time.

One common parameterization is:

where \(\lambda>0\) controls the scale and \(\gamma>0\) controls how the hazard changes with time.

• If \(\gamma=1\), the baseline hazard is constant.
• If \(\gamma>1\), the baseline hazard increases with time.
• If \(\gamma<1\), the baseline hazard decreases with time.

Combining this with a PK-driven exposure effect gives:

The exact parameterization and sign convention should always be stated because different software and publications use different Weibull parameterizations.

7. Turning the Hazard Into Survival Probability

The hazard function determines the cumulative hazard:

The survival function is then:

This relationship is particularly important for PK/PD modeling because the predicted concentration \(C(t)\) can vary continuously over time, causing the hazard to vary continuously as well.

For example, if drug exposure reduces hazard after dosing, the cumulative hazard may grow more slowly during periods of high exposure. This produces a higher predicted probability of remaining event-free.

8. How the PK Model Supplies Exposure

The time-to-event component does not need to estimate concentration independently if a PK model is already available. Instead, the PK model supplies the exposure history used by the hazard model.

For example, a one-compartment IV bolus model gives:

A time-to-event model can then use this concentration directly:

This creates a mechanistic sequence:

More complicated PK models can be substituted without changing the basic conceptual structure.

9. Why Time-Varying Exposure Matters

Drug concentration usually changes substantially over time. Therefore, using only a single exposure summary may discard information that is relevant to event risk.

A time-varying model allows the hazard to respond to the predicted concentration at each time:

This can be particularly useful when the event hazard is expected to track current or recent drug exposure.

However, the appropriate exposure metric depends on the pharmacology. The event may respond to:

• Current concentration.
• Effect-compartment concentration.
• Recent average concentration.
• Cumulative exposure such as AUC.
• Peak concentration.
• A delayed or filtered exposure measure.

10. Effect-Compartment Models for Delayed Event Effects

The event process may respond more slowly than plasma concentration changes. In that situation, an effect compartment can provide a useful intermediate exposure variable.

A simple effect-compartment model is:

where \(C_e(t)\) is the effect-compartment concentration and \(k_{e0}\) controls equilibration between plasma and the effect compartment.

The hazard can then depend on \(C_e(t)\) rather than plasma concentration:

This structure can capture situations in which event risk does not immediately follow changes in plasma concentration.

11. Adding Covariates to the Time-to-Event Model

Patient characteristics can affect both PK exposure and the underlying event process.

A covariate-extended hazard model might be written as:

where \(\mathbf{X}_i\) represents patient-level covariates such as baseline characteristics, disease severity, treatment group, or other clinically relevant predictors.

Covariates can also enter the PK model. For example:

This means a patient characteristic can influence the predicted concentration trajectory, which subsequently influences the event hazard.

12. How Censoring Enters the Analysis

Time-to-event datasets commonly contain subjects for whom the event has not been observed by the end of follow-up. Their exact event time is unknown, but their event-free follow-up time is known.

For a subject with observed follow-up time \(t_i\), define:

• \(\delta_i=1\) if the event was observed.
• \(\delta_i=0\) if the observation was right-censored.

The likelihood contribution can be expressed using the hazard and survival functions:

Equivalently, the log-likelihood contribution is:

This is one reason time-to-event PK/PD models differ fundamentally from ordinary continuous-response models: the likelihood must account for both event times and censored observations.

13. Estimating a PK/PD Time-to-Event Model

The model parameters are estimated by comparing observed event-time data with the event process predicted by the PK/PD model.

A practical workflow is:

1. Develop or obtain the PK model. Estimate concentration-time behavior and relevant PK variability.
2. Define the event. Specify precisely what constitutes an event and how censoring is handled.
3. Explore event-time patterns. Examine Kaplan–Meier curves, event counts, follow-up, and possible treatment differences.
4. Select a baseline hazard structure. Consider constant, Weibull, piecewise, or other suitable approaches.
5. Specify the exposure-response relationship. Consider linear, Emax, sigmoid Emax, delayed, or other biologically motivated functions.
6. Estimate parameters. Fit the joint or sequential model using an appropriate likelihood-based method.
7. Evaluate diagnostics. Assess event predictions, residual behavior, parameter plausibility, and sensitivity to model assumptions.
8. Simulate. Generate event-time distributions under alternative exposure or dosing scenarios when appropriate.

Depending on the analysis, PK and time-to-event components can be estimated sequentially or jointly.

14. Worked Example: Exposure and Time to Disease Progression

Consider a hypothetical drug studied in patients with a chronic disease. The event is disease progression. A one-compartment PK model predicts concentration after an IV dose.

Suppose:

• Dose \(D=400\) mg.
• Volume \(V=40\) L.
• Clearance \(CL=4\) L/h.
• Baseline hazard is constant with \(h_0=0.020\) h\(^{-1}\).
• The exposure effect is \(h(t)=h_0\exp[-\beta C(t)]\).
• \(\beta=0.05\) L/mg.

Step 1: Calculate the initial concentration

Step 2: Calculate the elimination rate constant

Step 3: Write the concentration-time model

Step 4: Calculate concentration at 5 hours

Step 5: Calculate the hazard at 5 hours

Thus, under this hypothetical model, the predicted instantaneous hazard at 5 hours is approximately \(0.0148\) h\(^{-1}\), compared with the baseline hazard of \(0.020\) h\(^{-1}\).

Step 6: Interpret the result carefully

The model does not say that 5 hours after dosing a patient's probability of progression is 1.48%. Hazard is an instantaneous rate, not a probability. The corresponding survival probability requires integrating the hazard over the relevant follow-up period.

15. Interpreting Exposure Effects

In a proportional-hazards model with a linear concentration effect:

A one-unit increase in concentration multiplies the hazard by:

For example, if \(\beta=-0.10\), then:

So a one-unit increase in concentration is associated with a multiplicative hazard factor of approximately 0.905 under the model.

However, when concentration changes continuously over time, the hazard ratio is also time-dependent:

This makes interpretation different from a simple treatment-group hazard ratio estimated from a conventional Cox model with a fixed treatment indicator.

16. Concentration, AUC, and Other Exposure Metrics

There is no universally correct exposure metric for every time-to-event endpoint. The appropriate choice depends on the pharmacology and biological mechanism.

Comparing several scientifically plausible exposure models can help determine whether the event process is better described by current, cumulative, delayed, or other measures of exposure.

17. Repeated Dosing and Time-to-Event Risk

Under repeated dosing, concentration can fluctuate around a steady-state pattern. The PK model can generate the full exposure trajectory rather than requiring a single summary exposure.

The time-to-event component can then be written as:

where \(E[C(t)]\) represents the chosen exposure-response function.

This structure allows the model to represent changing hazard as concentrations rise and fall during each dosing interval.

For chronic treatment, the model can therefore be used to investigate how changes in dose, dosing interval, clearance, or adherence might alter predicted event-time distributions.

18. Using the Model for Simulation

One of the major advantages of a PK/PD time-to-event model is that it can be used to simulate event-time outcomes under exposure conditions that were not directly observed.

A simulation workflow might be:

1. Specify the dosing regimen.
2. Simulate individual PK parameters and concentration-time profiles.
3. Calculate the exposure-response function at each time point.
4. Combine exposure with the baseline hazard.
5. Integrate the hazard to obtain the survival function.
6. Generate event times from the predicted survival distribution.
7. Repeat across many simulated subjects and scenarios.

This can support questions such as:

• How might a different dose change the predicted progression-time distribution?
• How does variability in clearance affect event risk?
• What happens if exposure is substantially lower than expected?
• How does a delayed PD effect alter the predicted timing of events?

19. Important Modeling Considerations

Time-to-event PK/PD models can be powerful, but several issues require particular attention.

• Exposure may be endogenous to the event process. Disease progression can sometimes alter dosing, adherence, clearance, or subsequent measurements.
• Time-varying exposure must be aligned correctly. Using exposure information that would not have been available at the prediction time can introduce bias.
• The exposure-response relationship may be misspecified. A linear model may be inadequate when effects saturate or are delayed.
• Baseline hazard matters. An overly restrictive baseline hazard can force the exposure component to explain patterns that actually arise from time-dependent baseline risk.
• Censoring assumptions matter. Informative dropout can require additional modeling.
• PK uncertainty propagates into PD predictions. Treating predicted exposure as known without accounting for relevant uncertainty can understate uncertainty in event-risk predictions.
• Identifiability can be difficult. Sparse event data may not support simultaneous estimation of a complex PK/PD exposure effect and a flexible baseline hazard.
• Competing events may matter. If different events prevent the event of interest from occurring, competing-risk methods may be needed.

20. A Practical PK/PD Time-to-Event Workflow

1. Define the clinical event. Specify the event, time origin, follow-up period, and censoring rules.
2. Develop the PK model. Establish the concentration-time relationship and relevant variability.
3. Explore the event data. Examine event timing, censoring, follow-up, and empirical survival patterns.
4. Select a baseline hazard. Use a structure appropriate for the observed time-to-event pattern.
5. Choose the exposure metric. Consider concentration, effect-compartment concentration, AUC, or another mechanistically justified measure.
6. Specify the exposure-response relationship. Consider linear, Emax, sigmoid, delayed, or other appropriate functions.
7. Add clinically justified covariates. Distinguish direct effects on hazard from effects mediated through PK exposure.
8. Estimate the model. Use the appropriate likelihood for event and censored observations.
9. Evaluate model adequacy. Check parameter estimates, diagnostics, predicted survival, event-time distributions, and sensitivity to structural assumptions.
10. Simulate and predict. Use the model for dose-exposure-event predictions only within a scientifically justified range.

21. How Time-to-Event PK/PD Differs From Other PD Models

The defining feature of the time-to-event approach is that it models when the event occurs, rather than simply whether it occurs or how large a response becomes.

Where to Go Next

A natural progression is to study PK/PD Models for Survival Analysis in greater detail, followed by Weibull exposure-response models, time-varying hazard models, recurrent-event PK/PD models, competing-risk models, and joint PK/PD models for longitudinal and time-to-event endpoints.

For more advanced applications, the next step is to connect population PK variability with individual event-risk predictions and use simulation to evaluate dose-response relationships, exposure targets, and clinical trial scenarios.