Correspondence

Viral Load and Response to Treatment of HIV

N Engl J Med 1996; 334:1671-1673June 20, 1996

Article

To the Editor:

The interesting article by O'Brien et al. (Feb. 15 issue)1 has several shortcomings. When measurements corresponding to the same time points are not available for all patients in a study, patients with samples obtained early after the start of antiretroviral therapy tend to have higher observed average reductions in human immunodeficiency virus (HIV) RNA than patients from whom samples were mostly obtained later; their actual responses may be different.

Second, since data on plasma levels of HIV RNA were missing for more than half the patients in the immediate-treatment group for any point beyond four months, the claim that changes in HIV RNA are more sustained than changes in the CD4 lymphocyte count is uncertain. The patients with missing data may have had more (or less) favorable HIV RNA responses than the other patients. Another study has shown attrition among patients with less favorable responses.2

Third, because of regression to the mean,3 spuriously larger reductions in HIV RNA may be seen in patients with high observed viral loads at base line. The converse happens with low observed base-line values. Random errors in measurement associated with the polymerase-chain-reaction assay are not negligible in relation to the modest effect of zidovudine monotherapy; such errors may be larger with variously preserved samples, as opposed to fresh ones. There are methods of accounting for regression to the mean.4 Furthermore, when Cox models are adjusted for the base-line value of a marker, multicollinearity may result in imprecise estimates and wide confidence intervals if the base-line values and RNA changes are correlated.

Moreover, O'Brien et al. use different thresholds and mention models with continuous data. The philosophies behind these approaches differ. “When continuous changes in plasma HIV-1 RNA levels were analyzed,” the authors write, “the percentage of the treatment effect that was explained was higher than when thresholds of change were analyzed.” Why then insist on thresholds? This is important to clarify. Given the availability of potent antiretroviral agents, is a reduction of 0.5 log good enough? Or should maximal effects be sought aggressively, with combination therapy and frequent changes of regimen? Finally, although all the participants in the trial had symptomatic disease, the clinical response to treatment was not considered as a predictor in the analysis. Is clinical evaluation useless in a high-technology era?

The authors' conclusions may be valid and consistent with other data,5,6 and measurements of viral load may indeed be suitable guides for antiretroviral therapy. However, the above-mentioned problems must be addressed if we want to understand the extent of the information provided by RNA changes and its usefulness.

John P.A. Ioannidis, M.D.
Joseph C. Cappelleri, Ph.D., M.P.H.
Joseph Lau, M.D.
New England Medical Center Hospitals, Boston, MA 02111

6 References

1
O'Brien WA, Hartigan PM, Martin D, et al. Changes in plasma HIV-1 RNA and CD4+ lymphocyte counts and the risk of progression to AIDS. N Engl J Med 1996;334:426-431
Full Text | Web of Science | Medline

2
Eron JJ, Quinn JB, Hill-Price S, Rooney J, Rubin M. 52 week follow-up of NUCA 3001: 3TC, zidovudine or both in the treatment of HIV positive patients with CD4 cell counts of 200-500 cells/mm3. Presented at the Third Conference on Retroviruses and Opportunistic Infections, Washington, D.C., January 28–February 1, 1996. abstract.

3
Yudkin PL, Stratton IM. How to deal with regression to the mean in intervention studies. Lancet 1996;347:241-243
CrossRef | Web of Science | Medline

4
Fleiss JL. The design and analysis of clinical experiments. New York: John Wiley, 1986.

5
Wathen LK, Nickens DJ, Chuang-Stein CJ, et al. HIV-1 RNA viral burden at baseline or its reduction following antiretroviral therapy is highly correlated with reduced HIV-1 disease progression. Presented at the Third Conference on Retroviruses and Opportunistic Infections, Washington, D.C., January 28–February 1, 1996. abstract.

6
Hammer SM, Katzenstein DA, Hughes MD, Hirsch MS, Merigan TC, ACTG 175 Virology Substudy Team. Virologic markers and outcome in ACTG 175. Presented at the Third Conference on Retroviruses and Opportunistic Infections, Washington, D.C., January 28–February 1, 1996. abstract.

To the Editor:

O'Brien et al. report that treatment-induced changes in HIV RNA and the CD4 lymphocyte count could together explain 79 percent of the benefit of immediate, as compared with delayed, zidovudine therapy in a Veterans Affairs cooperative study and therefore “can be used to assess the efficacy of zidovudine and possibly other antiretroviral drugs as well.” There are serious flaws in this argument.

The percentage reported by O'Brien et al. pertains to the proportion, Po, of the net effect of treatment (i.e., the effect aggregated over all mechanisms of action of the treatment) on the clinical end point that is explained by the markers. Unfortunately, variability in the estimation of Po is generally excessive1 (and Lin et al.: unpublished data). According to the authors, the 95 percent confidence interval associated with the estimate of 79 percent ranges from 27 percent to 145 percent. Given the width of the interval and the lower limit of only 27 percent, one cannot conclude that the markers studied account for a substantial portion of the net treatment effect. Furthermore, the authors do not explain or justify the validity of their bootstrap method of calculating the confidence interval; this technical issue is discussed by Lin et al. (unpublished data).

Another problem with the assertion by O'Brien et al. is that a marker that captures only a small percentage of the change in the rate of clinical progression that is induced by beneficial effects of treatment on the targeted HIV mechanisms may yield a value for Po that approaches 100 percent. This phenomenon occurs because the net effect of treatment, which comprises intended beneficial and unintended adverse effects, may be substantially less than the benefit of the intended effects alone (unpublished data). For example, if plasma levels of HIV RNA truly accounted for 40 percent of the benefit of zidovudine from intended mechanisms but the adverse effects reduced the net benefit by half, this reduction would make RNA appear to explain 80 percent of the net treatment effect, because the denominator used to calculate Po would be reduced by half. Similarly, deferred treatment will tend to reduce the difference between the treatments actually received by the patients in the two study groups and hence reduce the net effects of treatment. The smaller the net effect (reduced both by a delayed benefit in the deferred-therapy group and by cumulative toxic effects), the larger and more misleading Po will be. Had zidovudine never been offered to the patients in the deferred-therapy group, the changes in markers would be the same, the differences in net effects of treatment larger, and Po smaller. Therefore, in addition to being highly imprecise, the estimate of Po reported by O'Brien et al. is likely to be artificially inflated. It is much less plausible that this estimate reflects the ability of a single measurement of two “noisy” markers to capture nearly all the past and future effect of zidovudine treatment.

Even had O'Brien et al. been able to show that the markers completely explained the difference attributable to treatment, this would not necessarily imply that they are good surrogates when zidovudine is administered differently to a different population of patients, let alone for different classes of agents. Only a detailed understanding of the pathogenesis of HIV disease and the mechanisms of drug action, developed from studies that have the power to elucidate these mechanisms, can offer hope for the use of biologic markers as substitutes for clinical end points in AIDS research and clinical practice.

Victor De Gruttola, D.Sc.
Harvard School of Public Health, Boston, MA 02115

Thomas Fleming, Ph.D.
Robert Coombs, M.D., Ph.D.
University of Washington, Seattle, WA 98195

1 References

1
Freedman LS, Graubard BI, Schatzkin A. Statistical validation of intermediate endpoints for chronic diseases. Stat Med 1992;11:167-178
CrossRef | Web of Science | Medline

To the Editor:

In the report by O'Brien et al., the risk of progression to AIDS was defined only about 10 percent better by changes in the HIV RNA level than by changes in the CD4+ lymphocyte count (Table 1 of the article). In the sequential Cox proportional-hazards model relating antiretroviral treatment to the progression to AIDS (Table 2 of the article), a treatment-associated decrease of 75 percent from base line in plasma HIV RNA levels (model 1 of O'Brien et al.) was associated with a relative risk of 0.56, as compared with 0.62 for a 10 percent increase from base line in the CD4+ lymphocyte count. Although multivariate analysis is needed to correct for base-line differences in the markers, it would be important to see a graph comparing the trajectories of the HIV RNA levels in the treated patients who had progression to AIDS and those who did not.

The statistical manipulations are impressive, but a direct demonstration that the patients whose treatment produced sustained decreases in HIV RNA levels had the best outcomes would be more convincing. Figure 2 of the article shows a Kaplan–Meier analysis of the time to progression to AIDS that does not take treatment into account in patients who had a six-month mean decrease of 75 percent in plasma HIV RNA, an increase of 10 percent in the CD4+ lymphocyte count, or both. At 18 months there was no difference in the AIDS-free rate between the patients who had both changes and those who had only an increased CD4+ lymphocyte count. In striking contrast, those whose HIV RNA levels decreased but whose CD4+ lymphocyte counts did not increase fared no better than the patients in whom neither the HIV RNA level nor the CD4+ lymphocyte count improved. This clinical information is more persuasive than the authors' suggestion that dividing the difference between the coefficient of model 3 and that of model 1 by the coefficient of model 1 has substantive meaning.

One cannot conclude from these data that the costly measurement of plasma HIV RNA levels is preferable to the measurement of CD4+ lymphocyte counts.

W. Jeffrey Fessel, M.D.
Permanente Medical Group, San Francisco, CA 94115

Author/Editor Response

The authors reply:

To the Editor: Ioannidis et al. are concerned about the analysis of samples obtained at different times from different patients during the first six months of treatment. Although it is true that the earliest samples, those obtained in the first month after the start of therapy, would have the greatest reductions in plasma HIV RNA, for 87 percent of patients at least one sample obtained during the first two months was available, and 14 percent had only one sample in addition to the base-line samples. The contention that plasma HIV RNA samples obtained beyond the fourth month were missing for more than half of the patients in the immediate-treatment group is not true; such samples were available for 76 percent of patients. Furthermore, it is unlikely that there was more attrition among the patients with less favorable responses, since there were no therapeutic options to zidovudine at the time of the study, and those patients were highly motivated to continue receiving therapy in a blinded fashion.

In analyzing the change from the base-line levels of the marker, a threshold of change was used in order to establish a measure of response that can be used in clinical practice. We agree that reductions of more than 0.5 log should be sought, but this is the minimal degree of change that is associated with a clinical benefit. In contrast to the studies cited concerning the relation between changes in viral RNA and outcome, the Veterans Affairs study analyzed a subgroup of patients in whom clinical benefit was demonstrated. Therefore, this report validates the use of changes in these markers to measure clinical responses to therapy.

The confidence intervals associated with our estimate of the treatment benefit explained by the markers are broad, as De Gruttola et al. point out, but they do not include zero. In measuring the effect of treatment, we related treatment and changes in markers to events. In both cases, there was a net benefit. Thus, adverse effects and deferred therapy should reduce both the numerator and the denominator in the calculation of Po. The concern that HIV markers cannot be considered good surrogates even if all the treatment difference is explained ignores the obvious importance of viral load and CD4 lymphocyte counts as indicators of HIV disease. Nevertheless, although antiretroviral therapy has various mechanisms of action, its overall purposes are to decrease levels of HIV replication (of which plasma HIV RNA is an excellent indicator) and to slow destruction of the immune system (which is measured well by CD4 lymphocyte counts).

Fessel helps make our point that both the plasma HIV RNA level and the CD4 lymphocyte count should be considered in monitoring HIV infection. Our Kaplan–Meier analysis of the time to the progression to AIDS in patients with favorable responses to treatment, as indicated by one marker, both, or neither, indicated that both markers were important prognostic indicators. However, the response to treatment was better demonstrated by plasma HIV RNA. The relative risk of progression to AIDS in the treated patients who did not have a decrease in plasma HIV RNA of at least 75 percent was similar to that in patients initially assigned to placebo. We agree that further studies, using other classes of antiretroviral agents, will be needed to hone guidelines for establishing a successful response to therapy.

William A. O'Brien, M.D.
West Los Angeles Veterans Affairs Medical Center, Los Angeles, CA 90073

Pamela Hartigan, Ph.D.
Veterans Affairs Cooperative Studies Program Coordinating Center, West Haven, CT 06516

John D. Hamilton, M.D.
Veterans Affairs Medical Center, Durham, NC 27705

Citing Articles (1)

Citing Articles

1
P Bycott. (1998) An Evaluation of a Measure of the Proportion of the Treatment Effect Explained by a Surrogate Marker. Controlled Clinical Trials 19:6, 555-568
CrossRef