Utility for Ordinal Toxicity Grades

Why Grade Numbers Are Not Utility Values

Toxicity is often recorded in ordered grades: a higher grade indicates a more severe event. The order is informative, but it does not tell us how much worse one grade is than another. In particular, grade numbers do not establish that moving from grade 1 to grade 2 has the same clinical consequence as moving from grade 3 to grade 4.

A utility assignment makes the consequences of those grades explicit for a decision. It attaches a value to each possible grade, with higher utility typically representing a more desirable outcome. The key modeling choice is the size of the utility differences between grades. Equal steps are convenient, but they are an assumption—not a property of an ordinal grading system.

The grades and utilities answer different questions. A grade records the category assigned under a toxicity classification system. A utility represents how that category is valued for the decision at hand. For example, grades 0 through 4 may be ordered categories, while a chosen utility scale might assign them values from 1 to 0. Those endpoint values are a convenient scale, not inherent properties of the grades.

To calculate an expected toxicity utility at dose \(d\), let \(p_g(d)\) be the probability of grade \(g\) at that dose. The grades must be mutually exclusive and collectively exhaustive for the outcome being represented, so \(\sum_{g=0}^K p_g(d)=1\). The expected utility is the probability-weighted average of the grade utilities.

This calculation assumes that the utility assigned to a grade is the same wherever that grade occurs in the modeled population and at the doses being compared. It also assumes that the specified grade distribution describes the same endpoint, assessment period, and target population for each dose. If duration, reversibility, or another feature changes the consequences of a grade, a grade-only utility may not capture the decision-relevant outcome.

Equal Spacing Is One Particular Assumption

Suppose the grades run from 0, no toxicity, to \(K\), the most severe grade under consideration. If utility is anchored at 1 for grade 0 and 0 for grade \(K\), an equal-spacing assignment is

This formula makes the utility decrement from every grade to the next equal to \(1/K\). It therefore says more than “higher grades are worse”: it assumes a constant utility loss per grade step. If that assumption is not clinically defensible, the scale should instead represent the intended differences in severity. For instance, a small loss between grades 0 and 1 and a much larger loss between grades 2 and 3 can be represented by unequal utility gaps.

A utility scale need not use 1 and 0 as its endpoints. What matters for a particular expected-utility calculation is that the values represent the intended relative consequences on a common scale. When utilities are specified on a 0-to-1 scale, 1 and 0 are simply transparent anchors for the best and worst grades included. A value such as 0.6 does not mean “grade 0.6”; it is a utility score.

Representing Unequal Clinical Severity Gaps

When adjacent grades do not have equal consequences, assign a utility to each grade that reflects the intended differences. For example, consider \(u(0)=1.00\), \(u(1)=0.95\), \(u(2)=0.80\), \(u(3)=0.30\), and \(u(4)=0.00\). The successive utility losses are 0.05, 0.15, 0.50, and 0.30. This assignment says that the move from grade 2 to grade 3 carries a much larger utility loss than the move from grade 0 to grade 1. The numerical values should be justified for the decision context; the grade labels alone cannot supply them.

Expected Utility Through Grade-Exceedance Probabilities

A useful way to see how each grade transition contributes to expected utility is to express the scale in terms of its successive losses. Define \(\delta_k=u(k-1)-u(k)\) for \(k=1,\ldots,K\). For a decreasing utility assignment, each \(\delta_k\) is nonnegative. Since \(u(g)=u(0)-\sum_{k=1}^{g}\delta_k\), expected utility can be rewritten using the probabilities of reaching or exceeding each grade threshold.

To derive this form, substitute the expression for \(u(g)\) into the probability-weighted average, then collect the terms associated with each grade threshold:

The second line uses the fact that the grade probabilities sum to 1. In the third line, a loss \(\delta_k\) applies to every grade at or above \(k\), which is why the inner sum is the exceedance probability \(P(G\ge k\mid d)\). Equal spacing makes all the \(\delta_k\) equal. Unequal spacing weights exceedance at different grade thresholds differently.

Assumptions, Diagnostics, and Alternatives

Before using a grade utility in a dose comparison, check that the grading categories have a consistent clinical interpretation and that the probabilities refer to the same grading rules and observation period. Verify that the probabilities are between 0 and 1 and sum to 1, and inspect whether the utility ordering matches the intended preference ordering. If a higher toxicity grade is intended to be worse, a utility assignment that increases with grade is a warning sign unless the scale has a different, explicitly defined meaning.

A second diagnostic is to recalculate the comparison under plausible alternative severity gaps. If the preferred dose changes when a small utility increment is altered, the choice is sensitive to the utility specification. Report that sensitivity rather than presenting the ranking as a consequence of the grade data alone. Uncertainty in predicted grade probabilities and uncertainty in the utility values are distinct: even stable utilities do not make an imprecise probability model certain, and precise probabilities do not remove uncertainty about how grades should be valued.

The main alternatives make different assumptions. Equal spacing is simple and reproducible, but imposes a constant penalty per grade step. Unequal, clinically specified utilities can represent different severity gaps, but require a defensible basis for those differences. A rank-only comparison preserves the ordering of grades without claiming numerical distances; however, ranks alone do not define a probability-weighted expected utility. Assigning the grade number itself as a score is not neutral: maximizing its negative, for example, is equivalent to imposing equal losses between adjacent grades.

A grade-only utility also has limitations. It treats all events assigned the same grade as having the same value, even if their duration or reversibility differs. It may not represent multiple toxicities or patient-specific valuations. If those features matter to the objective, the outcome representation must include them rather than expecting one ordinal grade score to stand in for information it does not contain.