Determinants of discontinuation of initial highly active antiretroviral therapy regimens in a US HIV-infected patient cohort


  • *This work was presented as a poster at the 19th International Conference on Pharmacoepidemiology and the 1st International Conference on Therapeutic Risk Management, 21–24 August 2003, Philadelphia, PA, USA.

Dr Yong Yuan, Bristol-Myers Squibb Company, PO Box 4500, Princeton, NJ 08543-4000, USA. Tel: 609 897 2688; fax: 609 897 6319; e-mail:



Optimization of initial highly active antiretroviral therapy (HAART) for complete viral suppression and better tolerability is paramount for the prognosis of HIV-infected patients. Observational studies provide a better means than clinical trials of studying the determinants of discontinuation in actual practice.


A longitudinal cohort of US HIV-positive patients who initiated HAART for the first time from 1996 to 2003 were included in the analysis. Stratified Cox proportional hazards models, considering time-updated viral load and CD4 count data, were developed for analyzing time to first discontinuation.


A total of 3414 antiretroviral-naive HAART patients were identified. In a median follow-up period of 211 days (mean 324 days), 628 patients (18.4%) reportedly discontinued the HAART regimen because of drug toxicity, 456 (13.4%) because of non-compliance, and 257 (7.5%) because of treatment failure. In addition to the recorded reasons for discontinuation, black ethnicity [relative risk (RR) 1.28, 95% confidence interval (CI) 1.13–1.45], current smoking (RR 1.33, CI 1.18–1.50), high pill burden (RR 1.44, CI 1.22–1.70), and recent viral control (RR 0.63, CI 0.56–0.70) were all predictive of discontinuation. Only high pill burden (>15 pills/day), which is considered to be a surrogate for treatment regimen complexity, and the most recent poor viral control (HIV RNA) were found to be consistently associated with a higher likelihood of discontinuation.


Risk factors other than physician- or patient-reported reasons play a role in discontinuation of initial HAART regimens. Identification of these risk factors and simplification of treatment regimens in those at high risk for discontinuation appear to be necessary in order to maximize the effectiveness of HAART regimens.


Highly active antiretroviral therapy (HAART) has been shown significantly to reduce progression to AIDS-defining illnesses, AIDS-related hospitalization, and mortality in both clinical trials and a number of large observational studies [1–3]. However, discontinuation of HAART therapy may lead to suboptimal therapy and ultimately to treatment failure [4]. This underlines the importance of optimizing initial HAART, as success of salvage therapies after failure of initial HAART is limited. Recent data indicate that the overall discontinuation or modification rates of HAART regimens among patients range from 8 to 59% (median 33%) [5]. Antiretroviral toxicity or intolerance was reported as the main reason for discontinuation, accounting for up to 78% of all discontinuations, followed by treatment failure.

The advent of HAART as a treatment strategy marked a turning point in the battle against HIV infection. In particular, protease inhibitor (PI)-containing HAART regimens have led to a reduction in AIDS-related events, prolonged viral suppression and restoration of immunity, even in patients with advanced HIV disease. However, PI-containing HAART regimens are associated with two major problems: toxicity and regimen complexity. Side effects such as gastrointestinal intolerance, neuropathy, dyslipidaemia and glucose disorders have been described with various PI drugs [6–12]. Pill burden and regimen complexity may lead to problems with nonadherence, suboptimal therapy, discontinuation of therapy and, ultimately, treatment failure [4,13,14].

Observational studies in cohorts of HIV-infected patients have reported higher rates of discontinuation and failure than those found in clinical trials [5,15]. Observational studies are more likely to reflect what actually happens in practice settings. Clinical trials for assessing treatment discontinuation rates may not be representative of the overall HIV-infected population, mainly owing to inclusion/exclusion criteria which is too restrictive and mandated regimens and follow-up.

Furthermore, most analyses of reasons for HAART discontinuation have been limited to European studies [14,16,17] and we are not aware of any US observational study on reasons for HAART discontinuation. Most of the studies have been descriptive with limited time windows, available drug, and clinical experience. HAART pill burden and regimen complexity were not accounted for in the analyses, although a reduction in the number and frequency of pills to be taken every day has been suggested as the most significant treatment-related strategy to improve adherence [13].

To maximize the benefit of HAART for HIV-infected patients, understanding and modification of the predictors of nonadherence and drug discontinuation have become indispensable for effective patient management. The aim of this study was to define quantitatively the predictors of HAART discontinuation using a large prospective observational cohort of HIV-infected patients from across the USA.


Study database

HIV Insight™, a longitudinal database of out-patient clinical data collected from HIV clinics by Cerner (Vienna, VA, USA) and the HIV Outpatient Study (HOPS) on adult HIV/AIDS patients at physician sites throughout the USA, was used for this study. Approximately half of the physician sites are funded by the Center for Disease Control for HOPS. The other physician sites, collected in the exactly the same manner, are funded by the organization (Cerner) that manages the HOPS database.

HIV Insight™ and, in particular, the HOPS database have been described elsewhere [2,18–21]. Briefly, this is an ongoing prospective observational cohort study into which patients have been continuously recruited and followed since 1993. The participating physicians at these sites routinely care for over 2500 HIV-infected patients every year. Data from ongoing physician–patient interactions are collected through chart abstraction at each visit and entered electronically by trained staff. Diagnoses, drug treatments, symptoms, and laboratory results documented in each physician encounter are entered into the database, as well as patient demographics, patient risk factors, patient histories and reasons for ending drug therapies. The data are then compiled centrally and reviewed and edited before being analyzed. Each month, data quality control processing is undertaken to identify erroneous, suspicious or missing data and error reports are sent to sites for clarification or correction. Quality control processing programs are updated on a regular basis. Once a year, on-site audits are conducted to verify the completeness and accuracy of a sample of patient records. The ethical conduct of this study undergoes yearly review by federal (Centers for Disease Control and Prevention), contractor (Cerner Corporation), and local institutional research review boards.

Study patients

Data for this analysis are derived from a data extract containing information on 8792 adult HIV-infected patients with more than two visits from 1 January 1996 to 30 June 2003. An inclusion algorithm was used to identify the desired study population. Patients were included if (1) they used at least one HAART regimen (at least three antiretroviral drugs) from January 1996 to June 2003 (n=3813), and (2) they were ≥18 years old at HAART initiation (n=3811). Exclusion criteria were (1) use of any HAART regimen prior to 1996 (n=3429), and (2) missing values for any considered baseline variable (n=3414). A total of 3414 HIV-infected patients were included in the final study population. A patient's observation time lasted from the initial HAART (index date) to the first discontinuation or the end of follow-up (last clinical visit), whichever came first. Three types of HAART were considered in the study: (1) protease inhibitor-containing HAART (PI HAART), (2) non-nucleoside reverse transcriptase inhibitor-containing HAART (NNRTI HAART), and (3) HAART containing only nucleoside reverse transcriptase inhibitors (NRTI HAART).

Each reported start and stop date for antiretroviral drug use was collected, and all interruptions of therapy and their termination reasons (one primary and up to four secondary reasons) were also collected from clinicians in real time in the patient's medical records. Only primary reasons were considered in the study. The primary reasons for discontinuing a given drug recorded for patients remaining in the final analyses included: non-compliance attributable to patient preference (172; 12.83%), general non-compliance, including non-compliance on the physician's directive (284; 21.18%), general treatment failure (on the basis of viral load or CD4 count or other reasons: 184; 13.72%) immunological failure (73; 5.44%), general toxicity (363; 27.07%) and specific toxicity based on gastrointestinal intolerance, side effects (such as lactic acidosis), symptoms (nausea or fatigue), or laboratory data (abnormal liver function test or elevated cholesterol) (265; 19.76%). For this analysis, these crude reasons were further grouped as non-compliance, treatment failure and toxicities.

Statistical analyses

Not all drugs in a HAART regimen needed to be discontinued to be defined as HAART discontinuation in this study. Discontinuation was defined as stopping any antiretroviral in the initial HAART regimen and was attributed in the patient record to one of the following reasons: (1) treatment failure, (2) toxicity, or (3) non-compliance. Stopping treatment for more than 14 days was defined as ‘stopped’ in this study. Only the first discontinuation of HAART for each patient was included in the statistical analysis. Once the first HAART regimen had been modified in any way for a particular patient, that patient was no longer considered at risk for discontinuation and was therefore censored. Follow-up times of patients who never discontinued were censored at their last clinical visit. Patients who discontinued the initial HAART for reasons other than the endpoints of interest were censored at the time of discontinuation, under the assumption that the probability of discontinuation for other reasons was totally unrelated to that of discontinuing for treatment failure, toxicity or noncompliance. We found that the proportion of included patients censored did not differ amongst some key subgroups, such as HAART type (P=0.6192) and pill burden (P=0.7804) subgroups. The times to discontinuation broken down by reason were compared using the Wilcoxon rank-sum test. Univariate analyses were conducted to evaluate the association between the baseline risk factors and discontinuation of HAART. Stratified Cox proportional hazards models, considering time-updated viral control and CD4 count data, were developed for analyzing time to the first discontinuation for the three reasons of interest. Strata were formed according to type of initial HAART (PI HAART, NNRTI HAART, or NRTI HAART).

Baseline variables include age, gender, race, weight, illicit drug use and current and past smoking. In order to capture the pill burden of a HAART regimen, based on a pre-specified study plan, the total daily pill count for antiretrovirals was grouped into three categories: (1) low, < 10; (2) moderate, 11–15; (3) high, >15. We hypothesized that differences in discontinuation rates would be more sensitive to a five-pill increase than to a one-pill increase.

Two time-updated variables, CD4 cell count and viral load (HIV RNA), measured during follow up were considered in the analysis. These two covariates may change in value over the follow-up period, and therefore use of a time-updated specification of these variables is more appropriate and will avoid using future information to predict the discontinuation events [22]. These two variables were specified as dichotomous (undetectable viral load defined as either viral load <400 or <50 copies/mL, as the threshold for undetectable viral load for many HIV Insight patients was <400 copies/mL during the earlier phases of data collection, and <50 copies/mL during the late phases; high CD4 count defined as CD4 >350 cells/mm3), and updated every time a new event or censoring occurred to prepare the most recent data for the calculation.

All covariates considered in the model are specified and illustrated in Table 3 (see below). As there were many tied discontinuation events, the exact method [23] was used in the models to handle ties in the discontinuation time. We also tested whether the proportional hazard (PH) assumption was violated for each covariate used in the final Cox regression model [23]. All analyses were performed using Statistical Analyses Software (version 8.02; SAS Institute, Cary, NC, USA).

Table 3.   Estimated HAART type-stratified cox PH models of HAART discontinuation by reasons
VariableReference groupHazard ratio
All reasons with 95% CIFailureToxicitiesNon-compliance
  1. HAART, highly active antiretroviral therapy.
    *P<0.05; P<0.001.

MaleFemale0.926 (0.799, 1.075)1.0310.9830.826
Age (years)≤34    
 35–49 1.107 (0.979, 1.252)1.2471.214*0.924
 50–64 1.208 (0.992, 1.470)1.630*1.2400.975
 ≥65 0.981 (0.484, 1.987)0.5961.4720.644
 Black 1.283 (1.131, 1.455)1.467*1.0511.530
 Other races 0.886 (0.754, 1.053)0.8470.8181.023
 Current 1.329 (1.179, 1.498)1.0691.3661.435
 Past 1.209 (1.025, 1.426)*1.0641.270*1.222
Illicit drug useNone1.103 (0.996, 1.259)0.9791.0731.215
WeightContinuous0.999 (0.997, 1.001)1.0000.9981.000
Pill burdenLow    
 High 1.437 (1.216, 1.699)1.507*1.309*1.607
 Moderate 1.073 (0.935, 1.231)1.0371.0101.210
CD4 count≥350 cells/mm3<350 cells/mm30.934 (0.833, 1.047)0.5321.0631.063
Undetectable viral loadDetectable viral load0.626 (0.556, 0.704)0.4250.776*0.580

Sensitivity analyses, adjusting for calendar year of starting therapy and considering a continuous pill burden variable, were carried out to test the robustness of our findings. We also ran a subgroup analysis of PI HAART-only patients to investigate whether the overall findings were sustainable in this subgroup analysis.


A total of 3414 HIV-positive patients who had initiated HAART for the first time between 1 January 1996 and 30 June 2003 were identified (Table 1). Of these, 59.5% were white and 25.2% were African-American. Most of these patients were male (86.0%) and single (47.6%). A total of 34.2% were current smokers and 13.6% were past smokers. The mean age was 39.3 years (range 18–84 years) and 58.9% were 35–49 years old. As regards to pill burden, 1790 patients (52.5%) initiated HAART with a low pill burden, 1094 (32.0%) with a moderate pill burden, and 530 (15.5%) with a high pill burden. The majority of the study patients initiated a PI HAART (n=2597; 76%) from 1996 to 2003, followed by a NNRTI HAART (n=634; 19%) and a NRTI-only HAART (n=183; 5%). More than 92% of the study patients started their initial HAART between 1996 and 1999.

Table 1.   Study patient demographics and risk factors (N=3414)
Gender HAART type 
 Female479 (14.0%)PI2597 (76.1%)
 Male2935 (86.0%) NNRTI634 (18.6%)
Age (years)  NRTI183 (5.3%)
 ≤341021 (29.9%)Pill Burden 
 35–492009 (58.9%) Low1790 (52.5%)
 50–64358 (10.5%) Moderate1094 (32.0%)
 ≥6526 (0.7%) High530 (15.5%)
Race Index Calendar Year 
 White2030 (59.5%) 19961494 (43.8%)
 Black859 (25.2%) 1997936 (27.4%)
 Others525 (15.3%) 1998426 (12.5%)
Smoke  1999292 (8.6%)
 Current1166 (34.2%) 2000266 (7.7%)
 Past465 (13.6%)  
 Other1783 (52.2%)  
Weight (Ib)
 Mean (std)168 (31)  
Illicit drug use
 Yes673 (19.7%)  

In a median follow-up period of 211 days (mean 324 days), 628 patients (18.4%) reportedly discontinued the HAART regimen because of toxicity, 456 (13.4%) because of noncompliance, and 257 (7.5%) because of treatment failure. Toxicity accounted for 47% of these discontinuations, followed by noncompliance (34%) and treatment failure (19%). Discontinuations because of treatment failure (median 206 days) or toxicity (median 219 days) occurred earlier than discontinuation because of noncompliance (median 263 days) (P < 0.05).

Table 2 is a summary of a univariate analysis of baseline risk factors that might potentially impact the initial HAART discontinuation. We found that the characteristics that were associated with an increased likelihood of discontinuation of HAART were female gender [hazards ratio (HR) 1.22; 95% confidence interval (CI) 1.06–1.40; P=0.0057], black race (HR 1.46; 95% CI 1.30–1.64; P < 0.0001), current smoking (HR 1.37; 95% CI 1.23–1.53; P < 0.0001), and illicit drug use (HR 1.21; 95% CI 1.06–1.37; P=0.0039).

Table 2.   HAART discontinuation rates by baseline covariates (n(%))
VariableAll reasons
n (%)
n (%)
n (%)
n (%)
  1. HAART, highly active antiretroviral therapy.

 Female237 (49.4%)43 (9.0%)99 (20.7%)95 (19.8%)
 Male1104 (37.6%)214 (7.3%)529 (18.0%)361 (12.3%)
Age (years)
 ≤34384 (37.6%)65 (6.4%)168 (16.5%)151 (14.8%)
 35–49810 (40.3%)158 (7.9%)391 (19.5%)261 (13.0%)
 50–64139 (38.8%)33 (9.2%)64 (17.9%)42 (11.7%)
 ≥658 (30.8%)1 (3.9%)5 (19.2%)2 (7.7%)
 White756 (37.2%)136 (6.5%)390 (19.2%)234 (11.5%)
 Black425 (49.5%)96 (11.2%)165 (19.2%)164 (19.1%)
 Other races160 (30.5%)29 (5.5%)73 (13.9%)58 (11.1%)
 None579 (32.5%)124 (7.0%)267 (15.0%)188 (10.5%)
 Current567 (48.6%)97 (8.3%)264 (22.6%)206 (17.7%)
 Past195 (41.9%)36 (7.7%)97 (20.9%)62 (13.3%)
Illicit drug use
 Yes302 (44.9%)52 (7.7%)138 (20.5%)112 (16.6%)
 None1039 (37.9%)205 (7.5%)490 (17.9%)344 (12.6%)
Pill burden
 Low702 (39.4%)127 (7.1%)340 (19.0%)235 (13.1%)
 Moderate424 (38.8%)83 (7.6%)193 (17.6%)148 (13.5%)
 High215 (40.6%)47 (8.9%)95 (17.9%)73 (13.8%)

Table 3 shows the estimated stratified time-updated Cox PH models for time to the first discontinuation for the reasons of interest. In addition to the recorded reasons for discontinuation, black race [relative risk (RR) 1.28; 95% CI 1.13–1.45], current smoking (RR 1.33; 95% CI 1.18–1.50), past smoking (RR 1.21; 95% CI 1.03–1.50), high pill burden (RR 1.44; 95% CI 1.22–1.70) and the most recent viral control (HIV-RNA < 400 copies/mL) (RR 0.63; 95% CI 0.56–0.70) were all predictive of discontinuation (P<0.05). However, only high pill burden and the most recent poor viral control (in terms of HIV RNA) were found to be consistently associated with a higher likelihood of discontinuation across all discontinuation reasons. Tests of the PH assumption showed no evidence that the assumption of proportional hazards was violated.

We ran the model by replacing the pill burden variables with a continuous pill count variable. Although the estimated hazards ratio of the continuous pill burden variable was 1.013, it was not statistically significant (95% CI 0.997, 1.030), suggesting that each marginal daily pill count increase was not big enough to have an impact on discontinuation of HAART.

As most HAART regimens with high pill burdens were PI HAART regimens, in order to investigate whether these findings, interpreted as being attributable to low pill burden, may in fact be attributable to the avoidance of PIs, we ran a subgroup analysis for the PI HAART patients. The results confirmed that high pill burden was still a strong predictor of HAART discontinuation (HR 1.439; 95% CI 1.216–1.702; P<0.0001).


This study provides a quantitative analysis of the determinants of initial HAART regimen discontinuation in a large US HIV-infected patient cohort. This study differs from previous studies [14,16,17] in that it (1) was based on a database of US HIV-infected patients from multiple clinical sites; (2) contained very recent patient data for the latest regimens, including low pill burden regimens; (3) had a longer duration of follow-up.

A major finding of this study was that antiretroviral toxicity was the predominant reason for discontinuing the initial HAART, consistent with the findings of the majority of previous studies [5]. Non-compliance and treatment failure were the second and third most common reasons, respectively. It should be noted that the estimated rate of discontinuation because of non-compliance could have been higher as we only reported those discontinuation cases related to drug-taking behaviour and our study patients were HAART-naive patients who may be more likely to adhere to the initial HAART treatment than patients who are highly experienced with antiretroviral treatment.

Another key finding of this study was that discontinuation because of toxicity or treatment failure tended to occur earlier compared with discontinuation because of non-compliance. This finding differs from those of previous studies in that discontinuation because of toxicity occurred about 3 months earlier than discontinuation because of treatment failure [5]. The reason for the difference is unclear, but one explanation may be related to differences in the treatment regimens used in our studies compared with the earlier studies. A significant proportion of patients in our study were on NNRTI regimens, which are associated with a different side effect profile from those of the other classes of drugs, resulting in fewer discontinuations for toxicity reasons. Another possible explanation is that clinicians are more likely to switch earlier to an alternative antiretroviral regimen when patients are failing or developing side effects, while in contrast they try to improve treatment adherence in patients with low adherence using less drastic methods before switching medications.

Two risk factors, HAART pill burden and the most recent viral load control, were shown to be independently associated with a higher likelihood of discontinuing initial HAART across all reasons of interest. One of the differentiating factors in our analysis was the inclusion of total daily pill count, which can be viewed a surrogate for treatment complexity. Although treatment complexity has been consistently thought of as one of major reasons for discontinuing HAART regimens, few studies have modelled its independent impact on the discontinuation of HAART [5,14,16,17].

We found that HAART pill burden was an independent risk factor of regimen discontinuations, especially for discontinuation because of non-compliance. A previous study that defined HAART complexity as four or more antiretrovirals in a regimen also found that high pill burden was statistically significantly associated with discontinuation of HAART [5].

The most recent viral load was also found to be a major predictor of regimen discontinuation across all reasons of interest, consistent with an Italian study in which a time-updated variable for viral load and CD4 cell count was also used [17]. The achievement of low levels of viral replication is one of the key aims of antiretroviral therapy, and clinicians would consider modifying or discontinuing a HAART regimen after 16–24 weeks of treatment if a low viral load had not been achieved [17]. In our study, the CD4 cell count was not related to regimen discontinuation, suggesting that everyday clinical decisions might mostly be made on the basis of viral load measurements. Although the prognostic value of the CD4 cell count for clinical outcomes is well known [24–26], some studies found that the most recent CD4 cell count was not an important predictor of discontinuing the initial HAART regimen as a result of toxicity or treatment failure.

The study period spanned more than 7 years, during which time newer and simpler to use antiretroviral drugs were introduced and treatment guidelines were revised or updated. In order to adjust for the calendar year of starting therapy, we ran an analysis by adding dummy calendar year variables into the Cox model, and found the calendar year-adjusted model estimates to be similar to the unadjusted estimates, giving the same findings and conclusions.

Our study has some limitations, which should be considered in its interpretation. Only the primary reason for discontinuing the initial HAART was considered in the analyses. Although secondary reported reasons might also be contributory factors to stopping antiretrovirals, we believe that the primary reason should be primarily accountable for the HAART discontinuation. Patients who discontinued the initial HAART for other reasons outside our study interests were censored at the time of discontinuation, under the assumption that the probability of discontinuation of the initial HAART for one reason is completely unrelated to the probability of discontinuation for the reasons of interest. However, this assumption might not be valid as these censorings might be informative.

Another limitation of the study was that we were not able to account for the potential site-related variation which could confound our study findings, mainly because the site information was not disclosed in the database.

We are not sure whether, and to what degree, some of the discontinuations recorded as being primarily because of treatment failure or toxicity were classified as being because of non-compliance or other reasons, or vice versa.

While effective HAART regimens with very low pill burdens (pill counts as low as 3) may be currently available, adequate data on these regimens are not yet available. We tried to identify those patients that took 3–5 pills a day in this study sample: none took 3 pills per day, and only 163 (4.77%) took 5 or less. As these numbers were so small, we concluded that any finding based on more distinct categories for low pill burden would not be scientifically rigorous. However, we would like to suggest that future studies address this issue and investigate whether an even lower pill burden produces better adherence.

Lastly, as this was a retrospective observational study, use of the initial HAART therapy in the study might have introduced a selection bias, and comorbid factors and severity of illness may also affect the likelihood of HAART discontinuation. Although a Cox multivariate regression model was used to control for some confounding factors, hidden biases caused by unobservables may still exist.

The choice of initial HAART therapy has potential implications not only for the short-term success of therapy but also for the longer-term sustainability of antiretroviral activity, immune system restoration and adherence to the regimen, as well as the potential emergence of resistance to the individual drug components of the regimen [18].

The reported percentage of individuals who changed initial HAART as a result of toxicity, treatment failure or non-compliance is high and generates concern. The findings highlight the current need for new drugs with better tolerance, lower pill burden, and more convenient dosing regimens. Fortunately, the availability of newer once-daily antiretrovirals will offer clinicians and patients the opportunity to fashion PI-containing HAART regimens with a very low pill burden. Although the trend in favour of simplification of regimens cannot address all the issues that arise from exposure to lifelong therapy using complex regimens, at least patients' needs for lifestyle-friendly therapies could be fulfilled.

In conclusion, risk factors other than physician- or patient-specific factors were found to play an important role in determining discontinuation of initial HAART regimens. The study suggests that identification of these risk factors and simplification of treatment regimens for those at high risk for discontinuation are necessary to maximize HAART regimen effectiveness.


The authors acknowledge all HOPS investigators for their leadership and contributions in the HOPS study. We thank Rose Baker from Cerner for her unfailing support and her explanation of the database, and Scott Holmberg from CDC and Amanda Boening from Bristol-Myers Squibb for their useful suggestions about the conducting of the research. We are grateful to two anonymous referees for helpful comments.