Portal de Periódicos da CAPES

Monitoring of long-term toxicities of HIV treatments

2003; Lippincott Williams & Wilkins; Volume: 17; Issue: 17 Linguagem: Inglês

10.1097/00002030-200311210-00002

1473-5571

G. P. Bisson, Robert Gross, Veronica Miller, Ian Weller, Alexander M. Walker,

Pharmacological Effects and Toxicity Studies

Introduction The detection, characterization, and communication of information about drug toxicity are integral to medicine and drug regulation in all medical fields. Although the challenges we face in HIV may appear uniquely daunting, they are not essentially different. Over the first two decades of the HIV epidemic, the importance of drug safety monitoring was overshadowed by the need to develop potent therapies capable of arresting a fatal disease process. It is now clear, however, that the treatment successes achieved mainly through highly active antiretroviral therapy (HAART) are tempered significantly by drug toxicity [1–3]. These toxicities often occur in patients who have been exposed to multiple drugs for prolonged periods of time, thus the monitoring of long-term toxicities necessitates efforts extending beyond the usual follow-up of many cohorts and nearly all clinical trials. This review focuses on drug safety monitoring of HIV treatments and, in particular, the strengths and limitations of the available approaches for detecting and characterizing long-term toxicities. To understand the challenges specific to the monitoring of long-term toxicities, it is critical to understand the interconnected functions of both pharmacovigilance and pharmacoepidemiology. The purpose of pharmacovigilance is to detect previously unknown adverse drug effects. Pharmacovigilance sets the stage for formal pharmacoepidemiology studies, which involve control groups and are meant to refute or confirm and quantify drug safety risks. Risk management follows. Findings are communicated to patients and providers, while governments, manufacturers, and professional groups devise ways to change medical practice in order to avoid further toxicities. These processes are illustrated schematically in Figure 1. (A broader definition places all steps under the umbrella of pharmacovigilance, but in this review, we will refer to pharmacovigilance and pharmacoepidemiology as distinct concepts).Fig. 1.: Process of pharmacovigilance and pharmacoepidemiology.From a regulatory standpoint, safety monitoring of pharmaceuticals occurs in two phases – (1) before and (2) after a drug is approved. Leading up to approval, clinical trials provide data on predefined efficacy questions, frequent adverse events and immediate safety. Following approval, during postmarketing surveillance, safety information is derived primarily from spontaneous reporting systems supported by regulatory authorities in every industrialized country; further information may be generated by large post-marketing (Phase III and IV) clinical studies. The current system of drug safety assessment faces significant constraints, many of which are explained below. Despite these challenges, enhanced understanding of toxicities associated with use of antiretroviral medications can be expected to improve HIV care in several ways. Patients will be provided with information leading to more accurate expectations, perhaps decreasing disappointment and frustration when chronic, low-grade toxicities occur [4]. More realistic expectations may increase a feeling of trust and teamwork between patient and provider, which could lead to greater adherence [5]. Furthermore, better evidence regarding long-term toxicities will improve advice given to patients by clinicians about timing of initial therapy, choice of regimen, and drug substitutions or discontinuations. Better safety data may also lead to insights into the mechanism of toxicity and can eventually assist in drug development, in screening patients at high risk for toxicity, and in developing useful strategies for toxicity monitoring and management. For the purposes of this review, the term adverse event (AE) is defined as any untoward medical occurrence in a subject who has been administered a pharmaceutical product. The drug may or may not be causally related to the AE – in the cases where it is, the event is termed an adverse drug reaction (ADR). Indeed, sorting out the role of the drug in causing the AE is the goal of pharmacoepidemiology and drug safety monitoring. Available systems for monitoring the safety of HIV treatments – signal detection and hypothesis testing Signal detection Several systems currently are available for the detection of HIV-related AEs. They include pre- and post-approval clinical trials, spontaneous reporting systems such as the US Food and Drug Administration's (FDA) Adverse Event Reporting System and similar arrangements in the European Union and other industrialized countries, automated databases such as those compiled by Health Maintenance Organizations (HMOs), and observational cohorts. Randomized clinical trials Because individuals are first exposed to pharmaceutical products in the setting of clinical trials, these trials comprise a potentially valuable early source of toxicity data especially for relatively common ADRs with short-term onset. An example of one such toxicity is abacavir hypersensitivity reaction, which affects approximately 5% of individuals given the drug [6]. Randomized clinical trials (RCTs), however, have several limitations as signal generators. Pre-marketing studies usually recruit small, homogenous patient populations for study – typically fewer than 3000 exposed individuals. In this case, even if no serious ADRs are detected, one can only be confident that the rate of events is not greater than 3 per 3000, often referred to as the 'rule of 3' [7]. Thus, detection of relatively infrequent ADRs such as lactic acidosis is unlikely. Furthermore, individuals with significant comorbidities (e.g., hepatitis B and C) are commonly excluded, making detection of ADRs in these groups impossible. Similarly, ADRs may occur only in selected patient subsets that are insufficiently recruited and included in these trials. For example, female sex is associated with development of cutaneous rash caused by nevirapine [8]. Trials carried out largely in a male population could potentially miss or at least underestimate the occurrence of this reaction. Similarly, under-representation of ethnic minorities in clinical trials adds to this problem. Another limitation of RCTs (both pre- and post-approval) for the detection of long-term ADRs in HIV is the relatively short follow-up time of clinical trials. Indeed, increasing the follow-up of RCTs increases the cost and complexity of these studies. Moreover, the validity of the results often diminishes as losses to follow-up increase. However, several trials with extended follow-up are currently ongoing, including ACTG 384, the FIRST study, INITIO, and SMART, demonstrating that although difficult, long-term RCTs are feasible. Yet another limitation is the possibility that certain ADRs occurring in RCTs are under-reported. 'Minor' ADRs such as mood disorders or sleep disturbances may go undetected unless specifically targeted in data collection. These low-grade toxicities, however, may have substantial effects on adherence, which has been shown to affect several HIV-related outcomes [9,10]. Other ADRs, such as lipodystrophy, may not be classified in a standardized way, leading to poor detection in RCTs [11]. Homogenous patient populations, small sample sizes, and short durations of follow-up therefore significantly limit the use of RCTs for the detection of uncommon and late-onset toxicities. Spontaneous reporting systems After approval, the major mechanism of post-marketing surveillance is spontaneous reporting systems. This mechanism, used in both the United States and Europe, helps to identify ADRs that may not have been revealed during pre-approval trials. Completely passive in design, the identification of ADRs begins with the collection of spontaneous AE reports made by health professionals and patients and is typified by the FDA's MedWatch system [12]. Major strengths are the large scale (indeed, they potentially include all patients in clinical care in those countries with these systems) and the diversity of the population potentially included over long periods of time. Significant limitations do, however, exist. In the US, no federal laws or regulations require health care providers to report AEs related to pharmaceuticals, and it is estimated that the FDA receives reports on less than 1% of suspected serious drug-related events [13]. In many cases, physicians may feel an event is too trivial or too well known to report [14]. Other reasons for under-reporting may include physician guilt about harming a patient, fear of potential litigation, ambition to collect and publish cases, lack of awareness that an ADR has occurred, and lack of knowledge of how to report AEs via the available system [15]. A further limitation probably relates to lack of time and/or unwillingness to become involved in follow-up documentation or verification of reported AEs. Because of these limitations, the actual number of patients with a particular AE (the numerator of the AE rate) is unknown. Once reports are received, AEs are grouped into aggregate categories (i.e. a 'rash' may be further separated into 'maculopapular rash' or 'bullous eruption', etc.) based on standardized medical terminology dictionaries, such as MedDRA (Medical Dictionary for Drug Regulatory Activities) [16]. Some AE reports, however, may not be easily assigned to a single specific category. Moreover, terminology used in standardized medical dictionaries, despite ongoing efforts at harmonization, is often not consistent across international systems, adding further complexity to the process. After categorization, further challenges arise from the need to systematically identify and characterize those AEs that are observed to a greater rate than expected. The process of identification, based on pattern recognition, involves the use of prior knowledge and scientific inference to separate consistent, replicable 'signals' from a background of database 'noise' [17]. Given that any large surveillance system will produce many interesting but perhaps biased patterns of AEs and disease, it becomes vital that carefully reviewed associations be followed by formal pharmacoepidemiology studies in order to further evaluate cause and effect. This is particularly true in the case of HIV, where causal associations between drug and AE are often complicated by the multiplicity of treatments, any one of which might account for the toxicity, and by the possibility that the disease itself, apart from any treatment, may be the culprit. Indeed, the process of signal detection produces case reports and case series; observational and/or interventional studies that utilize control groups are critical in order to formally define risk. Automated databases Automated databases, originally developed to support computerized billing systems, are another potential resource for detecting antiretroviral toxicity. Automated databases provide large numbers of patients followed longitudinally through various health-care encounters and sequences of drug use. Necessarily smaller in their population coverage than nationwide spontaneous reporting systems, they nonetheless offer more complete ascertainment of serious events in the persons included in the database. Furthermore, some large systems (most notably in the UK, but also found in continental Europe and in the United States) were initially created as computerized medical records, and may, depending on local privacy laws, be used for individual record review or for population studies. The most common uses of these data historically have been to provide quantitative evaluation of signals generated elsewhere. They also provide a strong platform for building active surveillance systems. Another strength of automated databases is the relative heterogeneity of patients exposed to drug. Furthermore, prescription information may be available, providing one way of assessing duration of drug exposure. The longitudinal nature of the data is particularly valuable when the goal is to detect long-term toxicities. A partial list of automated databases available for the detection of ADRs in HIV is given in Table 1. Note that certain databases not included in the table, such as the General Practice Research Database in the UK and the Saskatchewan database in Canada, although large, currently have limited utility in HIV due to their relatively small numbers of HIV-infected patients included. A helpful discussion of the use of these specific databases for pharmacoepidmiologic purposes is contained in Part III of the book Pharmacoepidemiology [7].Table 1: Automated databases available for study of adverse events in HIV.However, in order for an AE to be coded in an automated database it must be recognized. The tendency of providers to recognize AEs as ADRs may in turn relate to diagnostic suspicion and/or other biases that over or underestimate the true association of a drug and a specific toxicity. In some cases, AEs may not be recognized at all. All of these issues may limit the ability to accurately identify and study toxicities using these sources. In some cases, billing codes (e.g. International Classification of Disease (ICD) codes) may capture the events with adequate sensitivity and specificity. However, the target event might be spread across many codes (e.g. upper gastrointestinal bleeding could be coded as upper gastrointestinal bleeding not otherwise specified, hematemesis, melena, or acute duodenal ulcer with bleeding), be buried under a rubric that contains numerous other entities, or correspond to an evolving syndrome for which no code yet exists, such as lipodystrophy syndrome [11]. Case ascertainment in this setting often requires multiple different aggregations of codes as different definitions of the same disease [18]. Chart review of suspected cases is almost always required. Alternatively, if linkage to the medical record is possible, some ADRs, such as anemia, may be ascertained via laboratory data. In general, those clinical entities that bring patients to the attention of caregivers, result in a quick and coded diagnosis, and can be supported or confirmed by a laboratory test (or chart review) are candidates for study using these sources. Specific ADRs, listed according to ease of study using these databases, are given in Table 2.Table 2: Adverse events according to ease of study using automated databases (aspects leading to ease or difficulty).Ad hoc cohort studies For epidemiologists, a cohort is simply a group of people followed over time during which health events are observed. In this sense of the term, the automated databases of the previous section can be used to form epidemiologic cohorts. The usable databases are circumscribed, however, by the kinds of data that are routinely captured. Ad hoc cohort studies, in which the data collection is specified in advance and implemented according to standard procedures, represent an improvement in detection of ADRs when the nomenclature and the diagnostic tests go beyond the routine (as occurs in automated databases). The follow-up time of cohorts often exceeds that of clinical trials conferring the benefits of identifying longer term and rarer toxicities. If data collection is prospective, 'low-grade' ADRs such as fatigue and mood disorders can be actively ascertained. Cohorts may also be useful for observation of specific patient subgroups, such as pregnant women and intravenous drug users – groups that are often excluded from clinical trials and underrepresented in automated databases. A partial list of cohorts relevant to HIV treatment is given in Table 3. Note that the cohorts' frequency and duration of follow-up, patient populations, data collected and number of participants vary greatly.Table 3: Partial list of cohorts available for study of adverse events (AEs) associated with antiretrovirals.Hypothesis testing After a signal is detected through pharmacovigilance efforts, further characterization is undertaken by way of hypothesis testing and formal epidemiological studies. In particular, it is important to clarify both causality and quantify risk. Understanding causality is critical to designing effective means to prevent or minimize the frequency of the ADR. Quantifying risk is critical to prioritizing the importance of the ADR relative to other ADRs and other causes of morbidity and mortality. Spontaneous reporting systems Spontaneous reporting systems offer very little basis for formal hypothesis testing. The significant under-reporting and lack of adequate data on number of patients exposed to a drug – as described in the section entitled 'Signal detection' above – does not allow ascertainment of the numerator (the total number of people with an event) nor the denominator, thus making the determination of incidence rates of AEs impossible. They are a valuable source of comparative information between drugs or drug classes only in the extreme circumstance of events that appear commonly in association with one drug or group, and almost never in their absence. Automated databases Automated databases have a potentially useful role in hypothesis testing. Once an ADR is suggested by spontaneous reporting, investigators can search an automated database for cases, controls, and exposures in order to explore associations. Planning for the use of automated databases for studies of ADRs in HIV requires several considerations. First, although database populations may be in the millions overall, only a small fraction of patients will have HIV, thereby limiting sample size and the power to detect rare events. Second, only outpatient prescription information is captured – data on over-the-counter drugs, alternative therapies, and inpatient prescriptions are not. Additionally, this information is based on prescription claims – it is not known if the medication was actually taken. Third, most automated databases are based on employee insurance information – if a patient changes insurance or employment, they may be lost from the database. Fourth, information on covariates such as diet, smoking, weight, and medication adherence patterns are often not readily available. This limitation may result in the presence of significant bias regarding patient selection, indications for treatment, and survival. Finally, death ascertainment is indirect (summarized usually from terminal care services), as only few databases are linked to vital records. One available supplementary source for this information is the National Death Index, but the index has a lag time of approximately 18 months. Ad hoc cohort studies Once an AE is identified, large, prospective patient cohorts can be used to assess associations of AEs with drugs or drug classes. A major advantage of prospective cohorts over automated databases is that specific treatment and covariate information can be more reliably ascertained by trained interviewers than can be obtained from medical records. For some HIV subgroups, such as intravenous drug users and pregnant women, observational cohort studies may be the only way to gain sufficient numbers of patients to adequately assess safety issues. Furthermore, the longitudinal nature of the exposure data in many cohorts is particularly useful in evaluating the effect of duration of treatment on development of toxicity. Some cohorts, such as the Multicenter AIDS Cohort Study (MACS) and the Women's Interagency HIV Study (WIHS), also provide valuable HIV-negative and HIV-positive treatment-naive control groups. Furthermore, biological samples relevant to toxicity may be adequately collected and stored – these data can aid in investigations into the pathophysiology of ADRs. One major problem with the study of HIV drug toxicity, even in ad hoc cohorts, is the lack of consensus definitions for various ADRs [19]. This is particularly true for certain complex ADRs, such as lipodystrophy syndrome [11]. Furthermore, non-uniformity of case definitions severely limits the ability of cohorts to increase power by pooling databases. Although individual cohorts may be large, for rare events, such as myocardial infarction (MI), the required number of case patients may indeed necessitate inter-cohort collaborations. This potentially useful method has been underutilized to date. One example of such collaboration is the D:A:D study – a prospective, multinational observational cohort study examining the association of antiretrovirals with cardiovascular disease. It involves 11 cohorts comprising 20 000 patients who are being followed for at least 2 years [20]. This study illustrates the way collaboration can increase sample sizes and power to detect infrequent but important ADRs that would otherwise be difficult or impossible to study. Another critical issue when considering cohorts as a source of toxicity data is that cohort studies are more likely to be affected by bias and confounding than are clinical trials, where randomization helps make exposed and unexposed groups more similar. Furthermore, issues of competing risks and the effect of co-morbidities may diminish the ability of investigators to identify less clinically apparent but nevertheless significant ADRs. All of these factors increase the time and effort needed for data collection, which contributes to costs and need for monetary commitments from both public and private funds. Randomized clinical trials Clearly, for the purposes of identifying valid associations between antiretroviral drugs and long-term toxicities, RCTs are the best research tool available. However, currently RCTs have a lesser focus on safety endpoints than on efficacy and this is true for both the pre and post-approval stages. This is unfortunate, considering that the search for specific causes of AEs could be greatly aided by the control of known and unknown confounders offered by randomization. Considering the situation of multiple treatments and co-morbidities common in HIV, the separate, blinded administration of unique regimens to two (or more) roughly comparable groups would likely afford valuable insights into HIV-drug safety. One excellent example of a clinical trial designed to follow patients for an extended period of time is CPCRA 065, or the SMART study. This study aims to compare the long-term clinical consequences of different strategies (drug conservation versus virologic suppression) of antiretroviral therapy, and follow up extends beyond 1 year. Other examples include the INITIO trial (efavirenz or nelfinavir plus didanosine and stavudine), with 3 years of follow-up, and the ACTG's ALLRT (Adult AIDS Clinical Trials Group Longitudinal Linked Randomized Trials) protocol, which will attempt to determine ways to maximize efficacy and minimize toxicity of drugs, and will follow patients for years. These studies demonstrate the multi-collaborative and resource intensive nature of the commitment required for these efforts. Improving the safe use of HIV medications – from reporting to risk management In order to improve the safe use of HIV treatments there needs to be commitment by all groups involved to make information on safety a major priority of HIV research. This commitment should result in a 'culture of excellence' dedicated to employing new talent and scientific rigor to the study of HIV-related AEs [21]. Commitment to this goal needs to translate into efforts specifically designed to improve all aspects of drug safety, including AE detection, analysis, and risk management. Overall, an agreement that safety measures constitute outcomes sufficiently important to merit their own dedicated research initiatives is needed. Critical to this agreement is an understanding that advances in the field of drug safety can yield measurable improvements in terms of public health. Also critical is an understanding of the limitations of our current knowledge of antiretroviral toxicities – what is known and what is not. This idea can be conceptualized as a drug safety map, as shown in Figure 2, derived from Waller and Evans [21]. The goal of improved pharmacovigilance in HIV should be to extend the boundaries of the small rectangle as much as possible.Fig. 2.: The drug-safety map. This figure aims to represent the 'dimensions' most relevant to the recognition and study of ADRs – mainly, duration of follow-up (the y-axis) and frequency of the event (the x-axis). The small box in the upper left represents the early, frequent area in which more effective, efficient study and risk management of ADRs occurs. The larger box comprising the rest of the map represents the unknown characteristics of those ADRs that occur later and less frequently.Reporting Improved post-marketing reporting of potential ADRs through spontaneous reporting systems is essential for enhanced understanding of long-term toxicities in HIV. Improving pattern recognition through provider education is one way reporting of AEs may be enhanced. In order to achieve this, education on ADRs should be prioritized in both medical school and continuing education curricula as well as via alternative, non-provider dependent mechanisms for AE reporting [22]. Given a greater understanding of pattern recognition, will providers report more AEs? One way to increase the likelihood of this may be to develop a team approach specifically directed at AE reporting. This approach should consider all members of the team, from patient to provider. Furthermore, an active approach emphasizing the exploration of new ways to improve reporting should be encouraged. A limited example of this active approach exists in the UK, where spontaneous reporting is based on the yellow card scheme. This system provides an opportunity for clinicians to report suspected ADRs to the UK Medicines Control Agency. Data from this system suggests that the inclusion of nurses into the pharmacovigilance team increased the number of reports received, which were of comparable quality to those received from doctors. In fact, nurses have officially been included as reporters of AEs in the UK [23]. Furthermore, a new blue card system (an extension of the yellow card scheme) targeting HIV drugs is an example of attempts at enhancing focused reporting. In other instances, as in France where reporting of AEs has been mandatory for physicians, midwives, and dentists since 1984, regulatory steps have been taken to increase the adequacy of reporting. Their impact will need to be evaluated, however. A critical, but often overlooked member of the surveillance team is the patient. Indeed, several long-term toxicities, including fat redistribution, were initially recognized by HIV-infected individuals. Although MedWatch currently accepts reports from patients, some data suggest that formalizing and encouraging patient involvement may further aid pharmacovigilance efforts. For example, in France (where patients currently are not allowed to report AEs) a pilot program allowing patients to self-report suspected ADRs using special reporting forms, is underway. The reports are communicated with a regional Pharmacovigilance Regional Centre (PVRC) which evaluates and stores the reports for further analyses later. Preliminary data suggest that the system significantly increases the number of reports received, especially lower-grade toxicities not requiring hospitalization or resulting in death (C Kreft-Jais, unpublished data). Moreover, the system exemplifies how such self-reporting systems can be efficiently channeled into existing surveillance systems. The team approach should be extended to include collaborations among groups of providers, such as clinics and hospitals, forming sentinel networks. The sentinel clinics would meet regularly and operate under a standard process for collecting data. Furthermore, the organizations could be queried to see if a particular AE signal found at one was also being received in others. Continued improvements are also needed in methods for differentiating AE signals from the vast expanse of background 'noise' (misreporting, duplication, coding errors) contained in the spontaneous reporting database. In this regard, there is considerable interest in data mining, a method used to detect higher-than-expected signals without using external exposure data or AE background information. The data mining method currently used by MedWatch, the multi-item gamma Poisson shrinker (MGPS), computes signal scores for pairs and higher-order (e.g. triplet, quadruplet) combinations of drugs and events that are significantly more frequent than their pair-wise associations would predict, and adjusts the size of estimates for the spuriously impressive associations that can appear more readily when the number of events is small [24]. Given the situation of multiple drug exposures and simultaneous co-morbidities common in HIV care, the technology may have significant benefit in the initial detection of ADRs related to HIV-treatments. Extending the reach of HIV pharmacovigilance also involves improving AE reporting from cohort studies and clinical trials. Perhaps most integral to this process is the formation of consensus definitions of all known ADRs, regardless of severity. Furthermore, regulatory emphasis should be placed on active surveillance through standardized forms of not only serious, acute AEs but also on potential long-term toxicities such as neuropathy, mood changes, fatigue, diarrhea and body morphology alterations. Surveillance should be active, and the frequency of these events should then be included in results in a standardized fashion. In order to detect long-term, low-grade ADRs in clinical trials, it will often be necessary to either increase the sample size, lengthen the duration of follow-up or both. Also critical is the inclusion of special patient groups in the study populations as mentioned previously. Finally, once studies are completed and submitted for presentation or publication, conferences and journals should not only reserve time and space for AE reporting but require it for studies that would otherwise be focused primarily on drug efficacy. Adoption of the CONSORT 2001 guidelines would help in this regard [25]. Analysis Methods of analysis also comprise an essential component to pharmacoepidemiology in HIV. The standard, primary intention-to-treat analysis of most clinical trials tends to overstate regimen tolerability and efficacy by not accounting for switches in antiretroviral backbone, a patient's ability to continue on study drug because of some intervention (e.g. loperamide for diarrhea), or ongoing ADRs that do not lead to changes in therapy [19]. This bias can be moderated by expressing of safety data as Kaplan–Meier plots to show cumulative risk, a method that explicitly incorporates and adjusts for the diminishing number of those at risk as the study duration increases. In addition, because ongoing cohort enrollment may have a dilution effect on the incidence of ADRs as time progresses, incidence rates of ADRs should be reported by calculating the number of events stratified by time of exposure. Elementary corrections for time on-study however do not account for other important confounding variables, such as length of time on drug and CD4 T-cell count, and minimizes the effect of competing risks (i.e. HCV causing liver dysfunction in a study of ART-induced hepatotoxicity). Both of these factors may require analysis of data using additional methods, such as multivariate Cox proportional-hazards models. Since a major concern of the effect of ADRs is their impact on tolerability of drug regimens, attempts to link ADRs with adherence measures should be made. Particular emphasis should be placed on the effect of ADRs on quality of life, as this association to date has been largely neglected. Finally, given the larger sample sizes needed for detection and analysis of more infrequent ADRs, further collaboration among cohorts is needed. In some cases, measures of association (i.e. incidence rates) for certain ADRs may be pooled and studied by formal meta-analysis, in this way increasing power. This type of collaboration will not only require investigators to cooperate and relinquish some degree of scientific 'ownership', but will also require additional funding. Funding agencies should be encouraged to support this type of initiative. An excellent summary of recommendations on ways to improve the analysis of ADRs, specifically within clinical trials may be found in a recent publication by Carr [26]. Health care researchers interested in the topic are referred there for a more extensive treatment of the topic. Risk management Once ADRs are characterized, the communication of this information needs to effectively reach not only health care providers but also patients. Regulatory agencies may act on the information in a variety of ways, including labeling changes, product warnings, 'Dear Doctor' letters, restricted use, and manufacturer withdrawal. Several examples of actions taken by the FDA on specific antiretrovirals are listed in Table 4. The efficacy of these interventions, however, has not been well studied. Such studies should include outcome measures of the interventions considering either clinical endpoints (morbidity and mortality) or process measures (effect on prescribing trends).Table 4: Examples of US Food and Drug Administration's regulatory actions in response to adverse event reports for anti-HIV drugs.Of course, an important caveat is that characterization of risk may only be clearly defined over time, and therefore some decisions regarding risk management have to be made prior to the full understanding of the phenomenon. Thus, expectations that all safety concerns be allayed and risk management strategies be finalized at their outset are clearly unrealistic. However, the goal should be to provide consumers and providers with as much evidence-based information as is available on the benefits and risks of HIV therapies as well as disclosure of the gaps in that information. This goal necessitates that risk assessments resulting from research findings be realistic and easily understood. The potential role of the Internet in offering data at the critical point of patient interaction, when drugs are prescribed, should be explored and developed further. Finally, considering the balance that exists between drug efficacy and toxicity, choices between drugs and drug classes should be subject to formal decision analyses. Decision analysis is a process which involves using the available evidence to create a model that defines predicted health outcomes associated with each option under consideration [21]. In terms of evaluating safety and the use of drugs, this process involves defining probabilities of ADRs, varying those probabilities within a reasonable range, and modeling the effects on various clinically meaningful outcomes (e.g. treatment discontinuation, non-adherence, change of therapy, clinical illness, death). This method makes clinical reasoning explicit, clearly identifies limitations of the available evidence, tests the impact of certain assumptions and may result in better justified decisions [21]. Ideally, use of this technique will help translate research findings into concrete clinical advances by facilitating more confident, rapid therapeutic choices at the critical level of patient care. Although it is recognized that certain dynamics shaping patient and provider preference cannot be easily included in these analyses, their results may aid in regimen choice after general treatment decisions (e.g. the decision to take HAART) have been made. Conclusion Overall, monitoring of long-term toxicities associated with HIV represents an area of research that is in urgent need of expansion. Indeed, to use medications effectively, we need to understand more precisely the realities of toxicity and the effect of these toxicities on clinical outcomes. Without this understanding, the success of our current therapies can, for a substantial number of individuals, be assumed to be short-lived. Acknowledgements This report is based on presentations and discussions during a 2-day workshop: 'Monitoring of Long-Term Toxicities of HIV Treatments' convened by the Forum for Collaborative HIV Research. The workshop brought together an international group of US and European drug regulators, HIV clinical researchers, HIV care providers, pharmacoepidemiologists, pharmaceutical company staff, and patient advocates. The Forum is a public/private partnership, which receives financial support from its governmental and industry members as well as in-kind support from its membership within the academic research, patient care, and advocacy communities. The workshop was sponsored by the Forum, with additional funding from the Division of AIDS, NIAID, NIH. We gratefully acknowledge travel support from Hoffman La Roche. We thank all members of the planning committee (see Appendix 2) and the co-cochairs Ian Weller and Alexander Walker for their investment in working with us to plan the workshop. We gratefully acknowledge the expert project management contribution from Bob Munk. June Bray, Paul Oh, Houtan Mova and Ipsita Das of the Forum Staff were responsible for making the project become a reality.