Patient-Level Utility for Multiple Outcomes

Why Score Outcomes at the Patient Level?

A dose can affect more than one outcome for the same patient. It may improve a symptom while also causing toxicity and increasing treatment burden. Looking at each outcome separately preserves useful detail, but it does not by itself define how to compare two patients whose outcome profiles differ. A patient-level utility provides a prespecified rule for mapping that joint profile to one score.

The order of construction matters. First define each outcome and its clinical direction; next transform each component onto a scale with a clear interpretation; then combine the components according to explicit tradeoffs. A numerical score is only as meaningful as these choices. In particular, adding raw efficacy, toxicity, and burden measurements is generally invalid when they use different units or when larger values have different meanings across outcomes.

The subscript \(i\) identifies a patient. The outcomes in the profile must refer to that same patient and to a defined assessment period. For instance, combining a response measured at week 12 with toxicity accumulated over one year requires a deliberate rationale; without one, the score mixes different clinical windows. In a dose-finding setting, the profile can be understood as the outcomes observed for a patient receiving a particular dose. The score summarizes that profile under the chosen rule; it does not, by itself, establish what that patient would have experienced at another dose.

Orient and Scale Each Component

A common convention is to define every component score so that higher always means better. This removes a frequent source of sign errors: efficacy often improves as its raw measure increases, whereas toxicity and burden often improve as their raw measures decrease. Orientation is a property of the transformation, not an assumption that the raw measurements all move in the same direction.

Suppose a continuous raw outcome \(x\) has clinically meaningful lower and upper anchors \(L\) and \(H\), where \(L<H\). If higher raw values are better, a simple bounded component score is:

This maps values at or below \(L\) to 0, values at or above \(H\) to 1, and values between them linearly. If lower raw values are better, reverse the direction:

The anchors should describe clinically meaningful reference points, not merely the smallest and largest values in the observed sample. Sample-based anchors can change when the study population changes, making the same patient profile receive a different score. A linear mapping also makes a substantive assumption: equal changes in the raw outcome produce equal changes in component score between the anchors. When that is not clinically credible, use a prespecified nonlinear or threshold-based component function instead.

For binary outcomes, define the component directly according to the preferred state. For example, if \(A_i=1\) indicates that patient \(i\) experienced an adverse event and \(A_i=0\) indicates no event, then \(u_{T,i}=1-A_i\) scores absence of the event as better. For an ordinal outcome, the score assigned to each category should reflect the intended meaning of the differences between categories; numeric grade labels alone do not imply equal spacing.

Combining the Oriented Components

One transparent combination is an additive weighted score. Let \(u_{E,i}\), \(u_{T,i}\), and \(u_{B,i}\) denote the patient’s oriented component scores, each on the interval from 0 to 1. With nonnegative weights \(w_E,w_T,w_B\) that sum to 1, define:

Because each component lies between 0 and 1, this weighted sum also lies between 0 and 1. If all weights are positive, improving a component while holding the others fixed cannot lower the score. For example, the effect of a change in the efficacy component is \(w_E\) times that change in the component score. That property depends on the components being oriented consistently; a toxicity score that accidentally rises with worse toxicity reverses the intended implication.

The weights specify tradeoffs on the transformed scales. They are not automatically determined by the measurement units, sample variability, or statistical precision of the outcomes. Nor does a weight of 0.4 mean that efficacy is “40% of the clinical benefit” in every patient. It is a coefficient in a specified scoring rule, whose interpretation depends on the component scales and anchors. The scoring rule must be justified for its intended decision context.

Direction and Scale in Practice

A useful way to verify a score is to test each component against simple hypothetical changes. Hold all other outcomes fixed. If efficacy improves, does the score stay the same or increase? If toxicity worsens, does it stay the same or decrease? If treatment burden increases, does it stay the same or decrease? For differentiable component functions, the signs of their derivatives give the same check: a higher-is-better efficacy transformation should have a nonnegative slope, while transformations of raw outcomes for which lower is better should have a nonpositive slope.

For example, on the interior of the anchored ranges above, the derivative of the higher-is-better transformation is \(1/(H-L)>0\). For the lower-is-better transformation it is \(-1/(H-L)<0\). With positive weights, the additive utility therefore changes in the intended direction when a single raw outcome improves and all other components stay fixed. At a clipped boundary the slope can be zero: further change beyond the anchor does not alter the score. That may be appropriate, but it should be a deliberate feature of the scoring rule rather than an unnoticed consequence of the formula.

Assumptions, Diagnostics, and Alternatives

A patient-level score is defensible only when its construction matches the clinical question. Before using it, check that the outcomes refer to a common patient and time horizon; the preferred direction of each raw outcome is known; the transformations preserve that direction; the anchors and weights have a stated rationale; and the combined score is suitable for the intended comparison. If a component is ordinal, verify that the assigned scores do not imply unjustified equal spacing. If a component has a threshold beyond which additional change is not valued, use a mapping that represents that decision rather than assuming linear improvement indefinitely.

A composite can also count the same clinical consequence more than once. For example, a symptom measure and a rescue-medication measure may both partly reflect symptom control. Including both is not automatically wrong, but the scoring rationale should account for their overlap; otherwise, that aspect of the patient’s experience may receive more influence than intended. Missing components require a prespecified rule as well. Simply dropping a missing component and rescaling the remaining weights can make scores for different patients represent different tradeoffs.

The main alternatives answer different questions. Keeping efficacy, toxicity, and burden as separate outcomes retains the full profile and avoids imposing one set of tradeoffs, but it does not produce a single composite ranking. A raw sum is usually unsuitable when outcomes have different units, ranges, or directions. A multiplicative benefit-risk utility is another possible construction, but it makes a different structural choice about how components combine. No single form removes the need to define the clinical meaning of the components and the tradeoffs.

A further limitation is that a single score can hide important differences in profiles. Two patients may have the same utility but very different combinations of response, toxicity, and burden. Therefore, report the component outcomes alongside the composite whenever the profile matters for interpretation. The score is a summary under an explicit rule, not a replacement for the outcomes from which it was built.