Why Compare Regimens by Predicted Exposure?
Two dose regimens can have different clinical value even when neither is clearly distinguished by observed toxicity at an interim analysis. One may leave many participants below an exposure associated with benefit; another may place more participants in a useful exposure range but also move some toward exposures associated with harm. Pharmacometric prediction makes these trade-offs explicit by translating each candidate regimen into a distribution of exposure and, where justified, expected response and toxicity.
The decision is not simply “choose the dose with the largest exposure.” Exposure is an intermediate quantity, not the clinical outcome. A useful comparison needs a prespecified target or decision criterion, a model linking regimen to exposure, and an adequately justified interpretation of the exposure-response evidence. The approach differs from toxicity-based dose-escalation rules, which make decisions primarily from observed toxicity counts or estimated toxicity probabilities.
Set the Comparison Before Looking at the Predictions
A regimen comparison begins with a defined exposure measure and a decision target. For example, \(A_d\) could be steady-state AUC over 24 hours, measured in mg·h/L, and the target might be an interval \([L,U]\) linked to a desired clinical response and acceptable safety. The interval should come from relevant pharmacology and clinical evidence, not from whichever interval makes one candidate look best after its predictions are seen.
A common working assumption is that log exposure is approximately normal for a given regimen:
Here \(m_d\) and \(\sigma_d\) describe the predicted distribution of log exposure for regimen \(d\). This implies that \(A_d\) is lognormal. It is an assumption about the distribution of exposure across the relevant population or predictive individuals, not a claim that every participant has the same exposure. If a model provides a different distribution, the target probability should be calculated from that distribution instead.
Under the lognormal assumption, taking logarithms converts the target event to an interval for a normal random variable. The standardized limits give:
where \(\Phi\) is the standard normal cumulative distribution function. This is a probability of exposure falling in the target interval under the specified predictive distribution. It is not the probability that the target interval is scientifically correct.
Worked Example: Calculating Target-Range Attainment
Suppose a hypothetical regimen produces exposure \(A_d\), measured in mg·h/L, with \(\ln(A_d)\sim N(4.2324,0.3607^2)\). The prespecified target interval is 40 to 100 mg·h/L. Calculate the probability of target-range attainment.
The standardized upper and lower limits are approximately:
Using standard normal probabilities, \(\Phi(1.0335)\approx0.8493\) and \(\Phi(-1.5068)\approx0.0659\). Therefore:
The predicted probability that exposure lies between 40 and 100 mg·h/L is about 78.34% under this model. This calculation alone does not show that the regimen is clinically effective or safe; those conclusions require the rationale for the target range and relevant response and toxicity evidence.
Compare Predicted Response Alongside Exposure
Target attainment is especially useful when an exposure interval has a defensible clinical interpretation. In other settings, the decision may require predicted response and toxicity across the full exposure distribution. If \(r(a)\) is the modeled probability of a beneficial response at exposure \(a\), then the regimen-level predicted response probability is the average of \(r(A_d)\) over the regimen’s exposure distribution. Likewise, if \(q(a)\) is the modeled probability of a specified toxicity at exposure \(a\), the regimen-level toxicity probability averages \(q(A_d)\) over that distribution.
A comparison should state how these quantities affect the decision. For example, a protocol might require a minimum response probability and a maximum toxicity probability, or might compare a prespecified benefit-risk summary. The criteria must be specified before applying them. A regimen is not automatically preferred merely because it has the highest mean exposure, highest target-attainment probability, or highest response prediction considered in isolation.
Worked Example: Comparing Exposure, Response, and Toxicity
Consider two hypothetical regimens and a target exposure interval of 40 to 100 mg·h/L. For transparent arithmetic, suppose each regimen’s predictive exposure distribution has two equally likely values. Regimen A has exposures of 40 and 80 mg·h/L; regimen B has exposures of 80 and 160 mg·h/L. At these exposures, the model predicts response probabilities of 0.30 and 0.60 for A’s two values, and 0.60 and 0.75 for B’s. Corresponding toxicity probabilities are 0.04 and 0.10 for A, and 0.10 and 0.24 for B. Suppose the decision criteria require target attainment of at least 0.80 and regimen-level predicted toxicity no greater than 0.15.
For regimen A, both exposure values are within the inclusive target interval, so target attainment is \(2/2=1.00\). Its average response probability is \((0.30+0.60)/2=0.450\), and its average toxicity probability is \((0.04+0.10)/2=0.070\).
For regimen B, only 80 mg·h/L is within the target interval, so target attainment is \(1/2=0.500\). Its average response probability is \((0.60+0.75)/2=0.675\), and its average toxicity probability is \((0.10+0.24)/2=0.170\).
Regimen B has higher predicted response in this example, but it fails both prespecified criteria: target attainment is below 0.80 and predicted toxicity exceeds 0.15. Regimen A meets both criteria. Under the stated decision rule, A is the eligible choice; the calculation does not establish that A is universally superior. The discrete exposure distributions and response probabilities are hypothetical and serve only to show how the quantities are combined.
How This Differs From Toxicity-Based Escalation
A toxicity-based escalation rule uses observed toxicity outcomes, or a model of toxicity risk, to determine whether a dose or regimen is sufficiently tolerable to continue or increase. Its principal signal is harm. An exposure-based comparison instead predicts where the regimen places participants on the exposure scale and relates that distribution to a prespecified target, response model, or benefit-risk criterion. Both may use toxicity constraints, but they answer different questions.
| Feature | Predicted-exposure comparison | Toxicity-based escalation |
|---|---|---|
| Main input | Predicted exposure distribution, often with response and toxicity models | Observed toxicity outcomes or estimated toxicity risk |
| Primary question | Does the regimen place enough participants in a useful exposure range or yield acceptable predicted outcomes? | Is the regimen tolerable enough to continue, escalate, or avoid? |
| Key assumptions | PK model predicts exposure adequately; target and exposure-response interpretation are relevant | Toxicity outcomes are defined and observed over an adequate window; the escalation rule is appropriate |
| Important limitation | A plausible exposure prediction may still mislead if the target or response model is wrong | Observed toxicity alone may not reveal whether exposure is sufficient for benefit |
A toxicity observation cannot generally substitute for exposure prediction: low observed toxicity may mean a regimen is tolerable, but it does not demonstrate that exposure is adequate for benefit. Conversely, a predicted exposure in a target range does not establish safety. The decision may use both sources of evidence, while keeping their roles distinct.
Worked Example: A Tolerable Regimen May Still Underexpose
At an interim review in a hypothetical trial, six participants have completed the relevant toxicity observation window at regimen C, and none has had the specified toxicity. A pharmacometric model predicts that only 35% of participants receiving C will reach the prespecified exposure target. Regimen D is predicted to have 82% target attainment and a toxicity probability of 0.10, below a protocol maximum of 0.20.
The observed count of zero toxicities in six participants is reassuring as a description of those six outcomes, but it is not proof that the toxicity probability is zero. Under a simple independent Bernoulli model with a common toxicity probability \(p\), the probability of observing zero toxicities is \((1-p)^6\). For example, if \(p=0.20\), that probability is \(0.8^6\approx0.2621\). Thus zero events in a small cohort can occur even when the underlying risk is not negligible.
The exposure predictions address a separate concern: C is predicted to underachieve the target, whereas D meets the stated attainment and toxicity criteria. If the models and criteria are credible, the exposure comparison supports D over C, subject to the trial’s prespecified decision process. A toxicity-only rule might identify C as tolerable without detecting that it is likely to underexpose participants.
Assumptions, Diagnostics, and Failure Modes
A prediction-based decision is only as credible as its inputs. Before comparing regimens, check that the predicted exposure measure matches the measure used to define the target; the regimen, dosing interval, and population are represented adequately by the PK model; and the exposure-response relationship is supported over the exposures being predicted. When response predictions are used, the model must connect exposure to the relevant clinical outcome, and not merely to a convenient biomarker.
Several diagnostics are particularly useful at an interim analysis:
- Compare observed and predicted concentrations. Systematic discrepancies, such as predictions consistently above measured concentrations, suggest model miscalibration or implementation problems.
- Check the range of prediction. Predictions far beyond the exposure range informing the response model rely on extrapolation and should not be treated as equally secure as interpolation.
- Inspect the distribution, not only its mean. A mean exposure can conceal a substantial fraction below the target or a high-exposure tail with potential safety implications.
- Check target and outcome definitions. A target based on one exposure summary or observation period cannot be applied uncritically to a different summary or schedule.
- Review uncertainty. A single plug-in prediction can appear more precise than the available data justify. Where the decision rule uses posterior target-attainment probabilities, follow the uncertainty integration specified in that rule rather than substituting a point estimate.
The lognormal formula assumes positive exposure and an approximately normal distribution of log exposure. Strong multimodality, unusual tails, or systematic subgroups can make that approximation inadequate. A model may also fit concentrations reasonably yet predict clinical response poorly if the exposure-response link is misspecified or the outcome is affected by factors not represented in the model. These are reasons to question a decision, not reasons to select whichever result is most favorable.
Justification of the Target-Attainment Formula
The formula follows by translating the exposure interval to the log scale. For \(0<L\leq U\) and \(\sigma_d>0\), since \(A_d>0\), the logarithm is defined at both limits and is increasing. Therefore \(L\leq A_d\leq U\) is equivalent to \(\ln L\leq\ln A_d\leq\ln U\). Under the assumed normal model, standardizing \(\ln(A_d)\) gives a standard normal variable \(Z=(\ln(A_d)-m_d)/\sigma_d\). Hence:
The subtraction is the normal probability between the two standardized limits. Reversing the order would produce a negative value, which is a useful check on the calculation.
Common Mistakes and Practical Limits
- Wrong reasoning: “There were no toxicities, so this regimen is the best choice.” Correction: Low observed toxicity does not establish adequate exposure or predicted benefit; assess each criterion separately.
- Wrong reasoning: “The regimen with the largest mean exposure must be preferred.” Correction: Compare the exposure distribution with the target and evaluate response and toxicity under the prespecified decision rule.
- Wrong reasoning: “The target-attainment percentage is the probability the regimen will work.” Correction: It is the predicted probability of exposure in a defined interval. Clinical benefit requires evidence linking that interval to the outcome.
- Wrong reasoning: “A precise model output removes uncertainty.” Correction: Prediction depends on model parameters, structure, data quality, and population relevance. A displayed probability does not eliminate those uncertainties.
Exposure-based selection is limited when exposure measurements are sparse, the PK model is poorly identified, the target is weakly supported, or the response relationship is uncertain. It also does not replace safety monitoring: predicted tolerability is not observed tolerability. In a well-specified adaptation, the exposure comparison complements rather than erases clinical outcome and toxicity evidence.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions below.
- If a regimen has predicted target attainment of 0.72, what does that probability mean, and what does it not mean?
- For \(\ln(A_d)\sim N(m_d,\sigma_d^2)\), write the formula for target attainment over \([L,U]\).
- A toxicity-based review finds few events. Name one important question that this finding does not answer about exposure.
- Why can two regimens with the same mean exposure have different target-attainment probabilities?
Solutions. (1) Under the predictive model, 72% of exposures are expected to lie in the specified target interval; it is not a 72% probability that a particular participant benefits or that the target is correct. (2) For \(0<L\leq U\) and \(\sigma_d>0\), the probability is \(\Phi((\ln U-m_d)/\sigma_d)-\Phi((\ln L-m_d)/\sigma_d)\). (3) Few toxicities do not establish whether exposure is sufficient to achieve the intended benefit. (4) Their exposure distributions may differ in spread or shape, changing the fraction inside the interval even when their means agree.