Survival in HIV-Infected Patients Who Have Received Zidovudine: Comparison of Combination Therapy with Sequential Monotherapy and Continued Zidovudine Monotherapy

Methods

The design of the Multicenter AIDS Cohort Study has been described previously [25]. Briefly, 4954 homosexual or bisexual men with no previous diagnosis of AIDS were enrolled in 1984 in Baltimore, Maryland; Washington, D.C.; Chicago, Illinois; Pittsburgh, Pennsylvania; and Los Angeles, California. Participants were seen semi-annually; at each visit, questionnaires were administered by an interviewer, a physical examination was done, and blood was drawn for laboratory tests and repository examination. An additional 625 participants, primarily under-represented minorities, were recruited from 1987 to 1991. A total of 2657 men were seropositive for human immunodeficiency virus type 1 (HIV-1) (shown by positive enzyme-linked immunosorbent assay and Western blot results) at study entry or seroconverted at some time during follow-up. Of the HIV-1-seropositive men, 1077 subsequently began receiving zidovudine before a clinical AIDS-defining condition was diagnosed (diagnoses were based on the 1993 Centers for Disease Control and Prevention [CDC] clinical AIDS definition, excluding the criterion for less than 200 CD4 cells/mm³ [26]). In this subcohort of patients receiving zidovudine, we compared the risk for progression to clinical AIDS diagnoses and survival according to whether the participants added didanosine or zalcitabine to the zidovudine they were receiving (combination therapy group, n = 318), switched from zidovudine monotherapy to monotherapy with either didanosine or zalcitabine (sequential monotherapy group, n = 263), or continued to receive zidovudine throughout the follow-up period (continued zidovudine monotherapy group, n = 496). The third group also included patients who stopped receiving all antiretroviral therapy. All models were run with and without a time-dependent covariate for men who stopped receiving all antiretroviral therapy. However, because estimates were almost identical, only the simpler models are reported.

Data on type, dose, and duration of antiretroviral therapy were collected during the routine study visit interview. Data were also collected on the use of other drugs (acyclovir and Pneumocystis carinii pneumonia prophylaxis including trimethoprim-sulfamethoxazole, aerosolized pentamidine, and dapsone) and clinical symptoms and signs (including diarrhea, oral candidiasis, or fatigue lasting more than 2 weeks; temperature more than 37.9 °C lasting more than 2 weeks; unintentional weight loss more than 4.54 kg; and herpes zoster). Complete blood counts and flow cytometry were done using quality-controlled, standard methods [27]. Zidovudine use was first reported in 1987, and use of didanosine and zalcitabine was first reported in 1990 [11, 28]. The follow-up period for our study was from 1987 to March 1994.

Clinical Outcomes

Clinical illnesses defined according to the 1993 CDC AIDS definition [26] were collected on a continual basis using self-reports and medical records. All self-reported diagnoses were validated from medical records according to CDC criteria [26]. Information on death was also collected, and records were continually updated. Because more than 95% of deaths among cohort members have been directly attributable to HIV infection or AIDS, we used all-cause mortality in all analyses. Clinical outcomes were grouped as 1) opportunistic infections, 2) cancers, 3) the wasting syndrome and dementia, 4) all AIDS-defining conditions [n = 494], and 5) all deaths (n = 431). Only 11% of the seroprevalent and seroconverted cohorts recruited for the Multicenter AIDS Cohort Study in 1984 had been lost to follow-up as of July 1994 [29].

Analysis

We compared baseline data and data at the visit before therapy was changed using the Wilcoxon rank-sum test, t-tests, and chi-square tests where appropriate [30]. The relative risk for developing an AIDS-defining clinical illness or dying, according to treatment group (combination therapy, sequential monotherapy, or continued zidovudine monotherapy), was estimated by separate Cox proportional-hazards models [30] for each outcome. Treatment group and covariates indicating HIV disease stages and other drug use were modeled as time-dependent covariates for two reasons. First, among the patients who began receiving zidovudine therapy, those who changed to combination therapy or sequential monotherapy did so at different subsequent time points. In addition, persons who changed from zidovudine monotherapy to other regimens had also been receiving zidovudine for different durations. Use of a time-dependent covariate for treatment group allowed comparison of patients who had been receiving zidovudine for the same amount of time and then changed to combination therapy, changed to sequential monotherapy, or continued to receive zidovudine. Such modeling eliminated differences in duration and time of changing therapy as sources of bias. Time-dependent covariates for changes in therapy also eliminated survival bias because only persons still surviving at each study visit were included in the treatment group comparisons at that time point. To further minimize any effect of changes in clinical or diagnostic practices over time, we adjusted analyses for the calendar year in which zidovudine monotherapy was first received (included in all models as a fixed covariate).

Second, time-dependent modeling of CD4 lymphocyte count, platelet count, hemoglobin level, HIV symptoms, and use of P. carinii pneumonia prophylaxis and acyclovir allowed adjustment of disease stage and use of other therapies at the same study visit during which treatment groups were being compared. This ensured that the model compared persons who were at the same disease stage and had been receiving zidovudine for the same period and that the model adjusted for the use of other therapies and temporal trends in access to therapy or diagnostic criteria. Finally, in models in which the outcome was survival, we added a time-dependent covariate for previous diagnosis of a clinical AIDS-defining illness.

The following are the time-dependent covariates used in the Cox models: 1) CD4 lymphocyte counts, modeled as a continuous variable [relative risk estimated per 100 cells/mm³]; 2) platelet counts, modeled as a continuous variable [relative risk estimated per 25 000 cells/mm³]; 3) hemoglobin levels, modeled as a continuous variable [relative risk estimated per 1 g/L]; and 4) symptoms of HIV infection, presence of AIDS, use of P. carinii pneumonia prophylaxis, and use of acyclovir, modeled as dichotomous variables (yes or no).

Time-dependent Cox models provide an unbiased estimate of the relative survival risk (or hazard) according to treatment group, but they do not allow an estimate of the duration of differences in survival. We did landmark analyses to address this question. In these analyses, participants must have begun receiving zidovudine and must have survived to one of three disease-stage landmarks: CD4 lymphocyte counts of 150 cells/mm³, 100 cells/mm³, and 50 cells/mm³. These landmarks were chosen because they represented a range of disease stages that included sufficient numbers of persons who changed from zidovudine therapy to either combination therapy or sequential monotherapy to allow meaningful comparisons of treatment groups. Because the survival benefit of combination therapy and sequential monotherapy did not significantly differ, we combined these groups to increase statistical power. Because patients who changed therapy and continued to receive zidovudine differ over time and because of potential selection biases, at each landmark we adjusted for the same prognostic factors (covariates) as in the time-dependent Cox models. However, no adjustment was made at subsequent time points (that is, after the landmark) for these variables to plot estimated absolute survival. To adjust for these covariates, we used mean values of continuous variables at the landmark to estimate survival curves by treatment group, and we set dichotomous covariates to discrete values. Thus, we compared persons who changed therapy at or before the landmark with those who continued to receive zidovudine monotherapy (before and after the landmark) while adjusting for other imbalances in disease-stage markers.

Finally, because this approach does not allow time-dependent analysis of covariates, patients who were receiving zidovudine at the landmark but who changed antiretroviral regimens after the landmark may introduce a potential source of bias. Classifying patients who changed therapy after the landmark with patients who changed therapy before the landmark tends to overestimate the survival effects of changing therapy, whereas including these patients in the control group (continued zidovudine monotherapy group) reduces group differences and over-attenuates survival estimates. In these circumstances, excluding this group gives the least biased survival estimate.

Results

The median duration of follow-up from the initiation of zidovudine therapy was 3.0 years (range, 0.1 to 6.7 years). The baseline CD4 lymphocyte counts of the three treatment groups significantly differed at the visit before the patients began receiving zidovudine. The sequential monotherapy group had a lower median CD4 cell count when zidovudine therapy was first started than did the zidovudine monotherapy and combination therapy groups (Table 1). However, the three groups did not differ for hemoglobin level, platelet count, percentage of patients who were symptomatic, percentage of patients who were receiving P. carinii prophylaxis, or percentage of patients who were receiving acyclovir.

The characteristics (at the visit before the change) of the group that changed to sequential monotherapy or to combination therapy are shown in Table 2. Patients who changed from zidovudine to sequential monotherapy rather than to combination therapy were more likely to do so at a more advanced disease stage. These data emphasize the importance of adjusting for disease stage not only at baseline but also at the time at which therapy is changed. Although persons who changed to combination therapy did so at an earlier disease stage than did persons who changed to sequential monotherapy (for example, median CD4 count of 246 cells/mm³ compared with 140 cells/mm³), they also did so at a later date (median year of change, 1992 and 1991, respectively). Both groups changed therapy as their risk for progression to AIDS increased: Forty-five percent and 47% were symptomatic, and 16% and 21%, respectively, had an AIDS-defining condition just before changing therapy.

Clinical Disease Progression

We examined the relative risk for developing four groups of clinical outcomes according to treatment group: presence or absence of any AIDS-defining condition, opportunistic infections, cancers, and the wasting syndrome and dementia. In the multivariate models, switching to sequential monotherapy was associated with a 29% reduction in the risk for developing an AIDS-defining condition (relative risk, 0.71 [95% CI, 0.52 to 0.98; P = 0.04]). Changing to combination therapy reduced the risk by 14%, but this difference was not statistically significant (relative risk, 0.86 [CI, 0.64 to 1.15; P > 0.2]).

Most of the effect of changing to sequential monotherapy could be explained by a reduction in the risk for opportunistic infections (relative risk, 0.76 [CI, 0.53 to 1.1; P = 0.13]), whereas the group that changed to combination therapy had a somewhat reduced risk for developing dementia or the wasting syndrome (relative risk, 0.57 [CI, 0.32 to 1.04; P = 0.05]). The effect of changing therapy on the risk for cancer was minor, regardless of regimen.

Survival

The survival effect of switching to combination therapy or to sequential monotherapy or of continuing to receive zidovudine is shown in Table 3. In this final multivariate Cox model, changing to combination therapy was associated with a 45% improvement in survival (relative risk, 0.55; P < 0.001) compared with continued zidovudine monotherapy. A 32% reduction in the risk for dying was seen in the sequential monotherapy group (P = 0.005). In the same model, higher CD4 cell count and hemoglobin level, use of P. carinii pneumonia prophylaxis, and use of acyclovir were associated with improved survival. Symptoms of HIV infection, a more recent year of starting zidovudine, and a previous AIDS diagnosis were associated with significantly worse survival.

Comparisons of Absolute Survival Times: Landmark Analyses

Specific disease-stage landmarks during follow-up were chosen as starting points from which duration of survival, by treatment group, could be estimated. Three disease-stage landmark points were chosen on the basis of CD4 lymphocyte counts of 150 cells/mm³, 100 cells/mm³, and 50 cells/mm³. Survival curves were estimated from Cox proportional-hazards models using mean CD4 lymphocyte counts, platelet counts, hemoglobin levels, percentage of symptomatic patients, percentage of patients receiving acyclovir, and percentage of patients receiving other prophylaxis at the landmark.

Survival curves for the 150 CD4 cells/mm³, 100 CD4 cells/mm³, and 50 CD4 cells/mm³ landmarks are presented in Figure 1 (top, middle, and bottom, respectively). The group that changed therapy had a longer median survival (about 3 to 6 months) than the group that continued to receive zidovudine. The relative risks for death associated with changing therapy were 0.58 (CI, 0.42 to 0.82) for the 150 CD4 cells/mm³ landmark, 0.66 (CI, 0.48 to 0.88) for the 100 CD4 cells/mm³ landmark, and 0.58 (CI, 0.44 to 0.77) for the 50 CD4 cells/mm³ landmark. As seen in Figure 1, the survival curves converged at 4.4 years for the 100 CD4 cells/mm³ landmark and at 3.5 years for the 50 cells/mm³ landmark. For the 150 CD4 cells/mm³ landmark, the duration of the difference in survival benefit was extended to 4.9 years of follow-up. Overall, survival was poor even for persons who changed therapy. Of patients who changed therapy before the CD4 count had decreased to 150 cells/mm³, none survived longer than 4.9 years (Figure 1, top). For patients who changed therapy before the CD4 count had decreased to 100 cells/mm³, mortality reached 100% at 4.3 years (Figure 1, middle); for patients who changed therapy before the CD4 count had decreased to 50 cells/mm³, the maximum survival was 3.5 years (Figure 1, bottom).

View larger version:

• In this window
• In a new window

Figure 1. Survival curves are based on the risk estimates for a patient who has a CD4 count of 100 cells/mm (first landmark), 60 cells/mm (second landmark), or 30 cells/mm (third landmark); a platelet count of 195 000 cells/mm (first landmark) or 190 000 cells/mm (second and third landmarks); a hemoglobin level of 127 g/L (first landmark), 123 g/L (second landmark), or 119 g/L (third landmark); and symptoms of human immunodeficiency virus infection and who is receiving acyclovir and pneumonia prophylaxis. Survival curves derived from a Cox proportional-hazards model comparing persons who changed from zidovudine monotherapy to alternate antiretroviral therapy (either combination therapy or sequential monotherapy) before reaching one of three landmarks: a CD4 lymphocyte count of 150 cells/mm³ (top), a CD4 count of 100 cells/mm³ (middle), and a CD4 count of 50 cells/mm³ (bottom).33333Pneumocystis carinii

Discussion

In our study, from changing zidovudine monotherapy to either combination therapy (with didanosine and zalcitabine) or sequential monotherapy was associated with significantly improved survival. However, the absolute extension of survival was modest, at best ranging from 3 to 6 months [1]. The significant improvement in survival seen for the groups that changed therapy was consistent with findings from several previous studies that have shown that in patients in whom antiretroviral regimens were altered, antiviral activity and increases in CD4 lymphocyte counts are greater than in patients who continued to receive zidovudine monotherapy [13-16, 24]. Previous published clinical trials of sequential monotherapy [13-15] or combination therapy [24] have not shown an effect on survival, probably because of few end points, interventions occurring late in the disease course, short follow-up, and participant drop out. A preliminary analysis of the ACTG 175 trial, presented after we completed our analysis, suggested a survival benefit with changing to or adding didanosine to zidovudine [31, 32]; this is consistent with our findings.

The European/Australian Delta trial [33] and the CPCRA (Terry Beirn Community Program for Research on AIDS) Nucombo trial [34] studied patients already receiving zidovudine who began receiving combination antiretroviral therapy at CD4 counts substantially lower than those in either the ACTG 175 study or our study. Unlike ACTG 175 or the Multicenter AIDS Cohort Study, neither study showed any advantage of combination therapy over continued zidovudine monotherapy. This further supports initiating and changing therapy at higher CD4 cell counts.

Several studies [8, 14, 35-38] suggest that the effect of subsequent nucleoside therapy is significantly affected by the CD4 cell count and duration of previous therapy, such that changing therapy at higher CD4 cell counts seemed to enhance the duration of therapeutic benefit compared with later changes in therapy. In our study, participants who changed therapy before reaching a CD4 cell count of 150 cells/mm³ had a longer survival (maximum, 4.9 years), whereas analyses at 50 cells/mm³ showed convergence of the survival curves at 3.5 years. However, because the median survival differences among treatment groups varied little by disease stage, it is unclear from our data whether there was a true interaction between changing therapy and disease stage. Overall long-term survival was grim, even among patients who changed therapy; this finding indicates the continued need for newer, more active antiretroviral regimens.

The small effect of changing therapy (to combination or sequential monotherapy) on the risk for other AIDS-defining conditions is, at first glance, surprising. Potential explanations include the universal use of prophylaxis for opportunistic infections that affect AIDS onset and the fact that zidovudine monotherapy was changed to combination therapy or sequential monotherapy just before or at the point of disease progression (21% and 16% of patients, respectively, had an AIDS-defining condition, and half were symptomatic). These data emphasize that AIDS outcomes are becoming increasingly soft end points for the assessment of antiretroviral therapy.

The Cox proportional-hazards model (with time-dependent covariates indicating change in treatment regimen, disease-stage markers, and use of other prophylactic regimens) allows an estimate of the overall survival benefit attributable to changing antiretroviral therapy while eliminating survival bias and minimizing disease-stage bias. To estimate the absolute increase in duration of survival, we used three disease-stage landmarks because treatment assignments changed over time. Unlike the Cox models with time-dependent covariates, the Cox models used at each landmark include only fixed covariates to adjust for imbalances among treatment groups [39]. This is less biased than using unadjusted analyses such as the Kaplan-Meier method, but it falls short of time-dependent modeling because 1) patients who change therapy after the landmark are misclassified as having continued zidovudine monotherapy at the landmark and 2) any differences in survival are attenuated. To address this problem, we derived estimates closest to the time-dependent models by excluding persons who changed therapy after the landmark. The sensitivity analyses we did to check our assumptions gave expected results. First, including persons who changed therapy after the landmark with persons who changed therapy at or before the landmark tended to give a longer survival benefit for changing therapy (as long as 12 months), but this estimate includes the possibility of a survival bias that overestimates the treatment effect. Second, including persons who changed therapy after the landmark in the continued zidovudine group completely attenuated the survival benefit and was inconsistent with the time-dependent models. Finally, we censored persons who changed therapy after the landmark; this yielded results similar to those of our original analysis, albeit with slightly higher relative risk estimates at the two higher CD4 cell counts.

Observational studies can yield important information on the effectiveness of therapies in clinical practice and can provide data on many clinical end points, which is often not feasible in trials of therapy initiated in intermediate- or early-stage disease. However, because participants in observational studies of these types are not randomly assigned to therapeutic regimens, these studies have important potential biases that must be addressed if the data are to provide valid inferences. The first of these is a survival bias (a form of lead-time biased sampling) [40, 41]. In cohort studies of HIV disease progression such as the Multicenter AIDS Cohort Study, new therapies or regimens are made available over time; long-term survivors are more likely to live long enough to gain access to these therapies. This bias leads to an overestimation of treatment effect because persons progressing more slowly will probably be in the treatment group, whereas those who progress more quickly will probably be in the non-treatment or comparison group. Strategies to adjust for this problem include stratification by the year in which therapy was started or changed so that only persons surviving to the same year are compared in each stratum [5, 42, 43]. A more elegant solution is to use a Cox model that allows not only for time-dependent variables indicating treatment group but also for adjustment of disease-stage variables in a time-dependent fashion [44]. This approach, which we [4, 39] and others [9] have used in previous studies, eliminates survival bias because the model compares only persons in the different treatment groups who survive to the same time point. This approach also has the advantage of allowing adjustment for disease stage at the same time the effect of the treatment group is being assessed. This helps minimize the second major bias of observational treatment studies, which occurs because of differential disease stage at the time of changing therapy. Finally, to test the proportionality assumption of the Cox model, we included a time interaction with main effects [44], which confirmed the validity of assumption.

Although we are confident that these approaches to modeling limited or eliminated the most important biases in our study, it is impossible to adjust for all variables. For example, in order not to include too many variables, we did not adjust for the use of antifungal agents or rifabutin. However, we believe that these variables probably had an insignificant effect on our results. Neither intervention has been shown to increase survival in clinical trials [45, 46], and neither is used by most patients in the Multicenter AIDS Cohort Study [28].

Appendix

The following are members of the Multicenter AIDS Cohort Study. The Johns Hopkins University School of Public Health, Baltimore, Maryland: Alfred J. Saah, MD, MPH, Principal Investigator; Ellen Taylor, MS; Joseph B. Margolick, MD, PhD; Richard Markham, MD; Justin McArthur, MBBS; Homayoon Farzadegan, PhD; Haroutune Armenian, MD, DrPH; and Neil M.H. Graham, MBBS, MD. Northwestern University Medical School, Chicago, Illinois: John P. Phair, MD, Principal Investigator; Joan S. Chmiel, PhD; Bruce Cohen, MD; Jerry Wesch, PhD; and Steven Wolinsky, MD. University of California, Los Angeles, Los Angeles, California: Roger Detels, MD, MS, Principal Investigator; Barbara R. Visscher, MD, DrPH; John L. Fahey, MD, MS; Janis V. Giorgi, PhD; Jan Dudley, MPH; Moon Lee, PhD; and Pari Nishanian, PhD. University of Pittsburgh Graduate School of Public Health and School of Medicine, Pittsburgh, Pennsylvania: Charles R. Rinaldo Jr., PhD, Principal Investigator; Monto Ho, MD; Lawrence A. Kingsley, DrPH; Phalguni Gupta, PhD; John Mellors, MD; and Allan Winkelstein, MD. Data Coordinating Center (Johns Hopkins School of Public Health): Alvaro Munoz, PhD, Principal Investigator; Lawrence Park, MSE; Stephen Gange, PhD; Donald R. Hoover, PhD; Kenrad Nelson, MD; Lisa P. Jacobson, PhD; and Sol Su, PhD.

From the Johns Hopkins University School of Hygiene and Public Health and School of Medicine, Baltimore, Maryland; Albany Medical College, Albany, New York; Northwestern University, Chicago, Illinois; the University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania; and the University of California, Los Angeles, Los Angeles, California.

Mr. Park: Department of Public Health Sciences, Bowman Gray School of Medicine, Medical Center Boulevard, Winston-Salem, NC 27157.

Dr. Stein: Departments of Medicine and Pharmacology, Albany Medical College, 47 New Scotland Avenue, MC A-142, Albany, NY 12208.

Dr. Phair: Department of Medicine, Northwestern University, 680 North Lake Shore Drive, Suite 1106, Chicago, IL 60611.

Dr. Mellors: Graduate School of Public Health, University of Pittsburgh, 403 Parran Hall, Pittsburgh, PA 15261.

Dr. Detels: Department of Epidemiology, University of California, Los Angeles, Center for the Health Sciences, 10833 Le Conte Avenue, Los Angeles, CA 90095-1772.