Why Optimize More Than One Outcome?
A regimen can improve predicted efficacy while also increasing predicted toxicity. When neither outcome can be improved without worsening the other, there may be no single regimen that is best on both measures. Multiobjective optimization makes this trade-off explicit: first identify regimens that are not unambiguously inferior, then use a stated clinical preference to choose among them.
The result is not a substitute for clinical judgment or for evaluating model adequacy. It is a structured way to connect model predictions to a decision. The comparison is meaningful only when efficacy and toxicity predictions refer to a clearly defined population, time horizon, and set of feasible regimens. As in the earlier material on regimen utility and population variability, the quantities being compared should represent the intended treatment population and decision context.
Dominance requires no trade-off judgment: a dominated regimen has no greater efficacy and no lower toxicity than an alternative. By contrast, Pareto efficiency does not mean that a regimen is clinically optimal. It means only that improving one objective from that regimen requires accepting a worse value on the other, among the regimens considered.
Constructing the Pareto Set
For a finite list of candidate regimens, compare each candidate with every other feasible candidate. Remove a regimen if another has at least as much predicted efficacy and no more predicted toxicity, with a strict improvement in at least one outcome. Retain every candidate that is not dominated. If the choices are continuous rather than a finite list, the same definition applies over the full feasible set, although locating its boundary may require numerical optimization.
The comparison assumes that each outcome is oriented consistently: higher efficacy is better and lower toxicity is better. If an endpoint is instead a harm measure or a symptom burden, its direction must be defined accordingly before comparing regimens. Predictions should also be on compatible scales and refer to the same target population and time horizon. Comparing, for example, efficacy at one time point with toxicity accumulated over a different period can produce a mathematically valid but clinically misleading frontier.
Worked Example: Removing Dominated Regimens
Suppose a hypothetical model predicts response percentages and the percentage of patients with a specified toxicity by a common follow-up time for five feasible regimens. The figures are illustrative, not results from a real study.
| Regimen | Predicted efficacy (response, %) | Predicted toxicity (%) |
|---|---|---|
| R1 | 50 | 2 |
| R2 | 65 | 4 |
| R3 | 72 | 8 |
| R4 | 60 | 5 |
| R5 | 68 | 3 |
R4 is dominated by R2: \(65\%\ge60\%\) efficacy and \(4\%\le5\%\) toxicity, with both comparisons strict. R4 is also dominated by R5. R2 is dominated by R5 because R5 has \(68\%\) efficacy rather than \(65\%\), and \(3\%\) toxicity rather than \(4\%\).
R1, R3, and R5 are not dominated. R1 has the lowest toxicity among these three, R3 has the highest efficacy, and R5 lies between them. Each offers an outcome advantage that cannot be obtained from either of the other two without worsening the other outcome. The Pareto set among the five listed candidates is therefore R1, R3, and R5.
Why Dominated Regimens Can Be Removed
A useful justification follows directly for a common utility function. Suppose clinical utility is \(U(r)=E(r)-\lambda T(r)\), where \(\lambda>0\) is the penalty, in utility points, for each unit of toxicity. If \(r_1\) dominates \(r_2\), then
Each bracketed term is nonnegative by the definition of dominance, and at least one is positive. Thus \(U(r_1)>U(r_2)\): under this utility, a dominated regimen cannot be the preferred choice. More generally, the same conclusion holds for any utility that increases with efficacy and decreases with toxicity. This is why finding the Pareto set is a useful first step, but it does not determine which efficient regimen should be selected.
Using a Specified Clinical Utility
A utility function expresses how the decision maker values efficacy relative to toxicity. The function must be specified in a way that is interpretable for the decision: its outcome definitions, units, and preference parameters should be stated. For \(U=E-\lambda T\), the value of \(\lambda\) determines the trade-off. If both outcomes are percentages and utility is reported in efficacy-percentage-point equivalents, then \(\lambda\) has units of utility points per toxicity percentage point. A weight without units or an explanation of its clinical meaning is easy to misinterpret.
Worked Example: Choosing Among Efficient Candidates
Use R1, R3, and R5 from the hypothetical table, and suppose the prespecified utility is \(U=E-2T\), where \(E\) and \(T\) are percentage points and the toxicity penalty is 2 utility points per toxicity percentage point.
For R1, \(U=50-2(2)=46\) utility points. For R3, \(U=72-2(8)=56\) utility points. For R5, \(U=68-2(3)=62\) utility points. The largest value is 62, so R5 is preferred under this specified utility. This selection is conditional on the chosen outcomes and penalty; it is not a claim that R5 is best under every clinical preference.
A specified utility can also be optimized when dose is continuous. The following example shows the assumptions and calculus for a simple model. The dose variable \(x\) is normalized and ranges from 0 to 2; it is not a dose in milligrams.
Worked Example: Optimizing a Continuous Dose
Suppose hypothetical model predictions over \(0\le x\le2\) are \(E(x)=70x/(1+x)\) response-percentage points and \(T(x)=4x^2\) toxicity-percentage points. Assume the functions describe the same target population and follow-up period throughout this dose range. Let \(U(x)=E(x)-1.5T(x)\), with the toxicity penalty expressed as 1.5 utility points per toxicity percentage point.
Substitution gives \(U(x)=70x/(1+x)-6x^2\). A differentiable interior maximum must satisfy
The second derivative is negative throughout the feasible range, so the utility is strictly concave; any interior stationary point is the unique maximum. Solving \(70=12x(1+x)^2\) numerically gives \(x\approx1.2025\), which is inside the allowed range. At that dose, \(E(x)=70(1.2025)/(2.2025)\approx38.2179\%\), \(T(x)=4(1.2025)^2=5.784025\%\), and \(U(x)=38.2179-1.5(5.784025)\approx29.5419\) utility points, or 29.54 to two decimal places.
The endpoint utilities are \(U(0)=0\) and \(U(2)=70(2)/(3)-6(2)^2\approx22.67\) points. The interior value is larger than both, consistent with the unique maximum. In this model, both efficacy and toxicity increase with \(x\), so no dose strictly dominates another dose: lower doses have less toxicity but also less efficacy. The utility selects one point from this continuous trade-off.
Preferences, Alternatives, and Diagnostics
A weighted utility is not the only way to make a choice. A constraint-based approach might require toxicity to remain below a specified limit and then maximize efficacy among regimens meeting that limit. A lexicographic approach might first require an efficacy target, then choose the least toxic regimen among those that meet it. These approaches can be easier to explain when a clinical requirement is genuinely a threshold or priority. Their conclusions depend on the threshold or ordering, just as a weighted utility depends on its penalty. The Pareto set is useful under all three approaches because a dominated option cannot improve either objective.
For the finite example, the preference weight can change which efficient regimen wins. With \(U=E-\lambda T\), the utility difference between R3 and R5 is \(U(R3)-U(R5)=4-5\lambda\). R3 is preferred to R5 when \(\lambda<0.8\), and R5 is preferred when \(\lambda>0.8\). R5 and R1 tie when \(68-3\lambda=50-2\lambda\), or \(\lambda=18\). Accordingly, R3 is preferred for \(\lambda<0.8\), R5 for \(0.8<\lambda<18\), and R1 for \(\lambda>18\); at either boundary the adjacent regimens tie. This makes the clinical role of the toxicity penalty visible rather than hiding it in a single selected dose.
Before acting on a result, check the analysis for signs that its assumptions or implementation are failing:
- Inconsistent outcomes: confirm that efficacy and toxicity use the same population, time horizon, and regimen definition. Otherwise, the apparent trade-off may compare unlike quantities.
- Incorrect dominance calculations: verify both directions and units. A lower toxicity percentage is favorable, while a higher efficacy percentage is favorable.
- Unexplained utility weights: show how the penalty reflects the intended decision. Recalculate choices over clinically plausible weights when the preference is not uniquely fixed.
- Overinterpretation of predictions: model-based point predictions depend on model assumptions and data support. Pareto efficiency does not establish that predicted differences are precise or clinically important.
- Unexamined feasible choices: the frontier is only as complete as the candidate set or dose range searched. Omitting feasible regimens can omit efficient options.
A frequent mistake is to call the regimen with the highest efficacy “optimal” even when it has substantially greater toxicity. The correction is to distinguish objectives from preferences: first identify the efficient trade-off set, then select a regimen using a stated utility or an explicit constraint. Another mistake is to treat every Pareto-efficient regimen as equally desirable. Efficiency rules out clear inferiority, not all preference differences.
Check Your Understanding
Answer each question using the definitions and decision rules above. Solutions are included below the questions.
- Regimen A predicts 40% efficacy and 2% toxicity; regimen B predicts 45% efficacy and 3% toxicity. Does either dominate the other?
- For \(U=E-1.5T\), calculate the utilities of A and B in Question 1. Which is preferred under this utility?
- For the five-regimen example, at \(\lambda=0.5\), calculate the utilities of R1, R3, and R5 and identify the preferred regimen.
- State one reason why a Pareto-efficient regimen is not necessarily the clinically preferred regimen.
Solutions. 1. Neither dominates: B has greater efficacy but also greater toxicity, so each has an advantage on one objective. 2. \(U(A)=40-1.5(2)=37\) points and \(U(B)=45-1.5(3)=40.5\) points, so B is preferred under this utility. 3. \(U(R1)=50-0.5(2)=49\), \(U(R3)=72-0.5(8)=68\), and \(U(R5)=68-0.5(3)=66.5\) utility points; R3 is preferred. 4. Pareto efficiency only means that no feasible alternative improves one outcome without worsening the other. The preferred regimen depends on the clinical utility, constraints, or priorities used to value that trade-off.