From Exposure Predictions to a Regimen Decision
An exposure-response model can predict an efficacy outcome or a toxicity outcome at a specified exposure. A regimen decision requires one further step: expressing those predictions on a common decision scale. A regimen with greater predicted efficacy may also carry greater toxicity burden, and the model predictions alone do not say how much of one outcome is worth accepting for the other. Regimen utility makes that tradeoff explicit by combining predicted outcomes with stated values or weights.
The focus here is translating candidate-schedule exposure summaries into predicted outcomes and then into utility. Exposure values such as AUC, maximum concentration (\(C_{\max}\)), or trough concentration are inputs to the relevant exposure-response models. The schedule’s exposure predictions must come from an appropriate pharmacokinetic model or data source; utility calculations do not themselves establish those exposures.
For example, if efficacy is a binary response, its expected benefit may be the response probability \(p_E\). If toxicity is also binary, its expected burden may be the toxicity probability \(p_T\). A simple score is then \(U=w_Ep_E-w_Tp_T\). This is not a universal clinical utility function: it is an explicit representation of a particular decision’s values. The score’s units are utility points if the weights are defined in points per unit of outcome.
Assumptions Behind the Translation
A prediction is meaningful only when the exposure and outcome models are used within their intended scope. For each candidate schedule, the exposure metric must use the same definition and units as the model’s input. For example, an AUC over one dosing interval cannot be substituted for an AUC over a different interval without a justified conversion. The endpoint definition, follow-up period, and population represented by the exposure-response model must also match the decision being evaluated.
The examples below assume that the stated exposure is the relevant value for the patient or representative scenario being scored, that the exposure-response models are calibrated for that setting, and that the utility is additive in the stated efficacy and toxicity outcomes. They treat model coefficients and predicted exposures as fixed. They do not include uncertainty in estimated coefficients or exposure predictions.
A further assumption is that each component of the utility is defined without accidental double counting. For instance, if a composite efficacy endpoint already includes an event that is also separately penalized as toxicity, subtracting that event again may count the same clinical consequence twice. If utility depends on whether efficacy and toxicity occur together, separate marginal probabilities may not be enough: a joint outcome model, or a justified dependence assumption, is needed.
Why the Utility Formula Works
Suppose the score for one patient is \(u=w_E B-w_T H\), with fixed weights. Taking expectations and using linearity of expectation gives the expected score:
This justification does not require \(B\) and \(H\) to be independent. It does require the score to be additive in \(B\) and \(H\), and the weights to be fixed. If the score includes an interaction term, such as an additional penalty when both outcomes occur, its expectation depends on the joint distribution. In that case, marginal efficacy and toxicity predictions alone cannot determine expected utility.
For a binary outcome \(I\), which equals 1 when the event occurs and 0 otherwise, \(E(I)=P(I=1)\). Thus its expected value is its event probability, which explains why the additive utility can use \(p_E\) and \(p_T\). For continuous outcomes, use the model-predicted expected outcome on its original scale, then apply the utility’s stated conversion weight.
Worked Example: Combining Trough-Based Efficacy and Peak-Based Toxicity
Consider three hypothetical schedules. Their representative trough concentrations and maximum concentrations are supplied as inputs. The efficacy model predicts response probability from trough concentration \(C_{\mathrm{trough}}\), and the toxicity model predicts toxicity probability from \(C_{\max}\):
Here \(C_{\mathrm{trough}}\) is in the model’s specified concentration units, \(C_{\max}\) is in ng/mL, and \(\operatorname{logit}(p)=\ln[p/(1-p)]\). Let the score be \(U=100p_E-150p_T\) utility points. The weights therefore assign 100 points to a unit increase in response probability and subtract 150 points for a unit increase in toxicity probability.
| Schedule | \(C_{\mathrm{trough}}\) | \(C_{\max}\) (ng/mL) |
|---|---|---|
| A | 1 | 70 |
| B | 2 | 55 |
| C | 3 | 48 |
For Schedule A, the efficacy logit is \(-1.5+0.8(1)=-0.7\), so \(p_E=1/(1+e^{0.7})\approx0.33181223\). Its toxicity logit is \(-3+0.025(70)=-1.25\), so \(p_T=1/(1+e^{1.25})\approx0.22270014\). Using the unrounded predictions,
Schedule A has an expected score of approximately \(-0.22\) utility points.
For Schedule B, the efficacy logit is \(-1.5+0.8(2)=0.1\), giving \(p_E=1/(1+e^{-0.1})\approx0.52497919\). The toxicity logit is \(-3+0.025(55)=-1.625\), giving \(p_T=1/(1+e^{1.625})\approx0.164516\). Therefore,
For Schedule C, the efficacy logit is \(-1.5+0.8(3)=0.9\), giving \(p_E=1/(1+e^{-0.9})\approx0.71094950\). Its toxicity logit is \(-3+0.025(48)=-1.8\), giving \(p_T=1/(1+e^{1.8})\approx0.14185106\). Thus,
Under this model, these assumed weights, and the stated schedules, Schedule C has the highest expected score. This is a model-based comparison, not proof that the schedule will produce those outcomes. In particular, it does not establish that the weights are clinically appropriate or that any separate safety constraint is satisfied.
Continuous Outcomes and Different Scales
Not every efficacy endpoint is a response probability. A model may predict an expected change in a continuous clinical measure, while toxicity is represented by an event probability. These can still be combined if the conversion weights explicitly map both quantities to a common score. The units matter: a weight applied to symptom-score units cannot be used as if it were a probability weight.
Worked Example: Combining an AUC-Based Improvement With Toxicity Risk
Suppose a hypothetical efficacy model predicts mean symptom improvement \(M\), measured in symptom-score units, from AUC:
A separate model predicts toxicity probability from \(C_{\max}\): \(\operatorname{logit}(p_T)=-4+0.08C_{\max}\), with \(C_{\max}\) in ng/mL. Define utility as \(U= M-10p_T\) points, where one symptom-score unit of improvement is worth one point and each unit of toxicity probability is penalized by 10 points. Compare two hypothetical schedules.
For Schedule A, \(AUC=50\) mg·h/L and \(C_{\max}=25\) ng/mL. The predicted improvement is \(M=12(50)/(30+50)=600/80=7.5\) symptom-score units. The toxicity logit is \(-4+0.08(25)=-2\), so \(p_T=1/(1+e^2)\approx0.11920292\). Hence,
For Schedule B, \(AUC=80\) mg·h/L and \(C_{\max}=35\) ng/mL. The improvement prediction is \(M=12(80)/(30+80)=960/110\approx8.7272727\). The toxicity logit is \(-4+0.08(35)=-1.2\), so \(p_T=1/(1+e^{1.2})\approx0.23147522\). Retaining the unrounded model predictions through the score calculation gives
Schedule B’s expected score is higher by about \(6.4125205-6.3079708=0.10455\) points. The efficacy improvement is larger, but so is the predicted toxicity probability; the utility calculation expresses the net result using the specified tradeoff. The small difference also signals that the ranking may be sensitive to assumptions about model estimates and utility weights.
Worked Example: Comparing Trough-Based Toxicity Risk With Response Benefit
Suppose a hypothetical model predicts a continuous mean benefit \(R\), in symptom-score units, as \(R=0.4C_{\mathrm{trough}}/(20+C_{\mathrm{trough}})\). A toxicity model gives \(\operatorname{logit}(p_T)=-3+0.3C_{\mathrm{trough}}\). Let \(U=20R-30p_T\) points, with trough concentration in the model’s specified units. Compare schedules with trough values 4 and 2.
At \(C_{\mathrm{trough}}=4\), \(R=0.4(4)/(20+4)=1.6/24=0.0666667\) symptom-score units, so the benefit component is \(20R=1.33333\) points. The toxicity logit is \(-3+0.3(4)=-1.8\), giving \(p_T=1/(1+e^{1.8})\approx0.14185106\). Therefore \(U=1.33333-30(0.14185106)\approx-2.92220\) points.
At \(C_{\mathrm{trough}}=2\), \(R=0.4(2)/(20+2)=0.8/22=0.0363636\), giving a benefit component of \(0.727273\) points. The toxicity logit is \(-3+0.3(2)=-2.4\), so \(p_T=1/(1+e^{2.4})\approx0.08317270\). Thus \(U=0.727273-30(0.08317270)\approx-1.76791\) points. Although the second schedule has less predicted benefit, it has the higher score under these weights because its predicted toxicity burden is lower. The negative scores indicate that each schedule’s expected penalized burden exceeds its benefit on this defined scale; they are not probabilities.
Diagnostics, Alternatives, and Limitations
Before interpreting a ranking, check the translation from exposure to score. Confirm that the exposure units and summary definitions match the fitted models, that model predictions remain in a region supported by data, and that the modeled endpoints correspond to the clinical decision and time horizon. Examine each component prediction separately: a plausible total score can conceal an implausible efficacy or toxicity prediction. Check arithmetic using unrounded probabilities and expected outcomes, rounding only for reporting.
A common mistake is to treat the largest efficacy prediction as the best regimen and regard toxicity as a separate descriptive detail. That reasoning ignores the stated decision tradeoff. The correction is to calculate the expected score for each candidate using the same models, endpoints, weights, and units, then interpret the components as well as the total. A different mistake is to round each component early and calculate the score from rounded values; small errors can change a close ranking. Carry adequate precision through the calculation.
Weighted utility is not the only decision approach. Efficacy-only selection ignores toxicity unless safety is addressed elsewhere. A safety-constrained approach instead excludes schedules that fail a prespecified safety requirement and compares efficacy among those remaining. That approach avoids implicitly allowing efficacy to compensate for a safety violation. Weighted utility makes compensation explicit, but its ranking depends on the chosen weights. If a hard clinical constraint exists, it should not be silently replaced by a penalty weight.
Utility also does not remove uncertainty. A calculated score is conditional on the model predictions, exposure inputs, and weights used. If a candidate lies outside the observed exposure range, the score may rely heavily on extrapolation. If exposure varies between patients or over time, applying a nonlinear exposure-response model to a single average exposure generally need not equal the average predicted outcome. The score should therefore be described at the level it represents, rather than as a guaranteed population result. Where the decision depends on joint events, averaging over patient variability, or uncertainty in model parameters, additional modeling is needed.
The practical value of this translation is traceability: a reader can see which exposure predictions generated each outcome, how outcomes were converted to utility, and which assumptions drove the ranking. Reporting the efficacy and toxicity predictions alongside the score makes it easier to identify whether a recommendation reflects strong benefit, low toxicity, or mainly the chosen tradeoff.
Check Your Understanding
Answer each question, then compare with the solutions below.
- If \(p_E=0.70\), \(p_T=0.10\), and \(U=40p_E-60p_T\), what is the expected utility? State what the number represents.
- A model has \(\operatorname{logit}(p_E)=-2+0.5C\). Calculate the response probability at \(C=4\).
- Why can substituting mean exposure into a nonlinear exposure-response model fail to give the mean predicted outcome across patients?
- A utility includes a penalty only when efficacy and toxicity occur together. Are the two marginal event probabilities sufficient to calculate it? Explain.
Solutions. 1. \(U=40(0.70)-60(0.10)=28-6=22\) utility points. This is the expected score under the stated weights, not a 22% response rate. 2. The logit is \(-2+0.5(4)=0\), so \(p_E=1/(1+e^0)=0.5000\). 3. In general, for nonlinear \(f\), \(E[f(X)]\ne f(E[X])\); the distribution of exposure affects the average prediction. 4. No. A penalty for joint occurrence depends on \(P(E\text{ and }T)\), which is not determined by the two marginal probabilities without additional information or a dependence assumption.