- © 2002 by American Society of Clinical Oncology
Troxacitabine in Patients With Refractory Leukemia
To the Editor:
We read with interest the article by Giles et al1 reporting the results of a phase II study of troxacitabine in patients with refractory leukemia. Although we tentatively agree with the authors’ conclusions that “patients experienced oncolytic response at levels of systemic exposure of troxacitabine below those as-sociated with severe toxicity,” we have concerns about the manner in which the pharmacodynamic data were analyzed and presented. (These analyses actually used data from a previous phase I study.)
The authors determined the area under the troxacitabine concentration-time curve (AUC) on day 5 for each of 18 individuals with pharmacokinetic data available, divided these into quartiles, and plotted the percentage of patients with an oncolytic response (defined in terms of hematologic recovery) and the percent with grade 3 or higher nonhematologic toxicity against the quartile range in Fig 2. They also fit a curve (Emax model) to determine the AUCrange50, ie, the “AUC range (or quartile)” at which the effect is half-maximal.
The preferred method for evaluating the relationship between a binary outcome variable and a continuous predictor variable is logistic regression analysis, which does not require collapsing the predictor variable into categories. However, the data are much too sparse (only two of 18 patients with severe toxicity and only four without a response) to perform such an analysis, and we agree in principle with the approach taken by the authors to divide the AUC values into quartiles and report the percentage of patients responding (percent with toxicity) within each. However their attempt to quantify this relationship further, as depicted in Fig 2 and via the Emax model, is problematic for two reasons. First, they plot the AUC values as evenly spaced along the x-axis, which they are not. It would also have been more appropriate to plot the percentages at the midpoints of these intervals, although of course the final interval (≥ 1,420) presents a difficulty in this regard. Second, the determination of the half-maximal effect via the Emax model involves an extrapolation beyond the range of the (response) data, always a dangerous practice. Therefore, the assertion (if we are interpreting their results correctly) that the half-maximal effect occurs approximately midway through the first quartile (they obtain an estimate of 0.55 for AUCrange50) is open to question. The small sample size and the fact that the percentage for the second quartile (three of five) is lower than that for the first quartile (three of four) should also give one pause.
Finally, analysis of the association between AUC and toxicity or response in a dose-escalation study does not distinguish variability in dose from variability in clearance, the two components of AUC. Thus, the purported relationship between troxacitabine AUC and toxicity/response should be further tested in phase II trials, where the dose is fixed and clearance is the only determinant of AUC.
Response
In Reply:
We thank Dr Karrison and Dr Ratain for their interest and comments on our article.1 They principally had concerns regarding the analysis performed that related troxacitabine exposure to oncolytic response and recommended an alternative method for evaluating this relationship. As suggested, we applied a logistic regression model to the data to evaluate the association between troxacitabine day 5 area under the curve (AUC; ng/mL·hr) and oncolytic response. Relative to the analysis that was performed,1 this model is appropriate for a relatively small data set (N = 18) because the model uses only two df. In addition, this model assumes a monotonic relationship between day 5 AUC and response. In this analysis, six of 18 patients did not have a response, and two of 18 patients experienced severe toxicity. Results of the logistic regression model are shown in Fig 1. The relationship between day 5 AUC and response is positive. Figure 1 demonstrates that the probability of response increases between an AUC value of 250 and 1,250 ng/mL·hr, where it begins to level off, implying that exposure above AUC values of 1,250 ng/mL·hr are not increasingly effective. The 95% confidence interval for the response probability also suggests that there is no evidence that a day 5 AUC of 1,250 ng/mL·hr is any less effective than higher exposures. The model predicts that the odds ratio comparing two individuals whose day 5 AUC differ by 250 ng/mL·hr is 1.6 for probability of response (P = .11). Although the magnitude of this odds ratio is relatively large, it is not statistically significant because of the small sample size.
Fig 1. Predicted probability of oncolytic response as a function of troxacitabine day 5 AUC. Solid line represents fitted model; dotted lines are 95% confidence intervals for true response probabilities by day 5 AUC. The raw data are plotted; |P indicates response or nonresponse for individuals with grade < 3 nonhematologic toxicity, and X indicates response or nonresponse for individuals with a grade ≥ 3 nonhematologic toxicity. Symbols along fitted line represent different troxacitabine dose levels.
We agree that the association between AUC and toxicity or response in a dose-escalation study does not distinguish between variability in AUC that results from variability in the dose administered or variability in clearance between patients. We also agree that the best setting to test this relationship is in a phase II trial when the same dose is administered to all patients. Unfortunately, pharmacokinetic studies are not always incorporated into phase II trials, as they were not for troxacitabine. We hope that performing exploratory analysis of exposure-response relationships during phase I development will encourage the incorporation of pharmacokinetic studies into phase II trials to adequately define exposure-response relationships.
