Utility of Different Adherence Measures for PrEP: Patterns and Incremental Value

According to UNAIDS, an estimated 2 million individuals acquired HIV in 2015 globally, with infection rates being highest in sub-Saharan Africa [1]. There is therefore an urgent need for effective methods to prevent ongoing transmission of HIV. A number of clinical trials examining the efficacy of oral pre-exposure prophylaxis (PrEP) [2‐6] with tenofovir disoproxil fumarate/emtricitabine (TDF/FTC) have been published in recent years. Results from these studies have been largely positive, indicate that PrEP prevents acquisition of HIV infection when users are adherent [2, 4], and have led to broad recommendations for PrEP use worldwide [7].

Accurately measuring adherence to PrEP remains a challenge. In the HIV treatment setting, adherence is equally difficult to measure but viral loads can serve as a surrogate of adherence, unlike in the context of PrEP. Although multiple metrics of PrEP adherence have been investigated, there is no gold standard. Non-pharmacokinetic (non-PK) measures include self-report, pharmacy refills, pill counts, and electronic monitoring (EM; pill bottles that record each opening). Pharmacokinetic (PK) metrics are also available, such as monitoring drug levels in plasma, peripheral blood mononuclear cells (PBMCs), dried blood spots (DBS), or hair. Various combinations of these measures have been used in the PrEP trials conducted to date, and recent papers have found different degrees of concordance among them [8‐13]. Self-report is generally found to over-estimate adherence and correlate poorly with adherence determined by objective measures, such as drug levels in plasma, EM, and pill count [5, 9, 10]. In the International AIDS Vaccine Initiative (IAVI) phase II trials of PrEP conducted among serodiscordant couples in Uganda [13] and men who have sex with men (MSM) and female sex workers (FSW) in Kenya [14], modest associations were observed between PK measures (plasma, PBMC, and scalp hair levels) and adherence assessed by EM, but there was little correlation between self-report and drug levels in any matrix.

Discrepant results with different adherence measures could imply bias and/or inaccuracy with one or more of the measures. For instance, social desirability bias and recall error commonly limit the utility of self-reported adherence. EM may be inaccurate if individuals remove more than one pill per opening or do not remove any pills with an opening. Moreover, removing pills from a container may not necessarily translate to drug ingestion. PK metrics are influenced by the half-life of the drug moiety and could be influenced by the timing of the sample draw relative to dosing (for short-term metrics), as well as biological and analytical variability in assay performance. Moreover, drug levels in different biomatrices such as hair and plasma measure adherence over different time periods and a patient may not be uniformly adherent over time. A significant limitation common to many comparisons of PK and non-PK measures is that the reporting periods are not always aligned. For example, a self-reporting period may be 1 month in duration and that period may be compared with a tenofovir concentration in plasma, which reflects short-term drug ingestion (1–7 days). Tenofovir drug levels in PBMCs reflect ingestion over the moderate-term (7–14 days) and levels in hair and dried blood spots reflect ingestion over the long term (weeks to months). Finally, analyses comparing adherence measures in PrEP trials to date have also not yet explored the incremental value of using multiple adherence measures, adjusting them to reflect the same duration of PrEP use. Guidance for adherence measurement in PrEP trials, demonstrations projects and real world roll-out is therefore needed.

In this analysis, we provide a comprehensive assessment of multiple non-PK and PK adherence measures adjusted to align over a set duration of time within a phase II PrEP trial in two distinct populations: MSM and serodiscordant couples. We use short and long-term PK measures to assess patterns and concordance of adherence to PrEP over time. We also determine the incremental value of assessing multiple adherence measures in combination (self-report and drug levels in plasma and hair) compared to EM.

Among participants taking daily PrEP in a clinical trial, we described patterns of adherence to PrEP by combining short and long-term pharmacokinetic measures and aligning them with concurrent electronic data monitoring metrics and self-reported adherence. We found evidence of consistently high adherence over different durations of time as assessed by multiple measures at both sites. In Kenya, the EM measure showed more moderate levels of adherence than in Uganda. Prior qualitative work with this population found that challenges, such as complexities of daily life, may have contributed to lower adherence for MSM [27].

We found mixed results in the concordance and correlation analyses. Concordance was high in comparisons of self-report, EM, and plasma and hair drug levels at the Ugandan site, although more moderate concordance was seen among these measures at the Kenyan site. Interestingly, the correlations between measures were higher in the Kenyan population compared to the Ugandan population, although correlations were generally lower than concordance. This finding may reflect the distribution of the data and the manner in which we created the PK categories for low, moderate, and high adherence. Our choice to use medians makes sense for this small study and the need to avoid the influence of outliers. However, values that fall just on either side of the medians may be misclassified, thus causing differences in analyses of concordance versus correlation. Furthermore, while the correlations are determined using data from these trials, the categorizations for concordance determination were based on median values from previous studies.

Importantly, limitations in the various adherence measures may have affected the accuracy of our comparisons. For instance, among the MSM at the Kenyan site, stigma [27] may have resulted in a reduced desire to use the EM device. Participants may have truly adhered to their medication through pocket dosing (e.g., opening the device and removing multiple pills on 1 day), thus resulting in high drug levels in plasma and hair, and decreased concordance of EM with other measures. Indeed, much of the discordance in this population reflected higher adherence as assessed by the PK measures and lower adherence as adjudicated by EM. Social desirability may have led to inaccuracies in self-reported adherence, as has been seen in multiple prior studies [25, 28]. Additionally, drug concentration determinations may have been limited by variation in sample processing, differences in metabolism within and between individuals, drug–drug interactions, and/or variable dose timing relative to sample collection (relevant only for plasma samples, in which the half-life of tenofovir is approximately 17 h) [8, 10, 19].

A ROC analysis was used to determine the incremental value of combining multiple adherence measures in order to discriminate among low, moderate, and high EM adherence. As expected, we found that self-report alone performed poorly (AROC 0.54). This finding is similar to an analysis in the Ancillary Adherence Study within the Partners PrEP Study where discrimination of steady-state daily dosing versus less than steady state dosing for plasma tenofovir was poor for self-report (AROC 0.52). Our results indicate that hair drug levels offered the best discriminant ability (AROC 0.85) as a single PK measure. This finding suggests that a long-term measure of adherence, as represented by hair levels, can help summarize the day-to-day variability captured by the EM device. Using one PK measure plus self-report improved the discriminant ability to about 90% and a combination of two PK metrics and self-report predicted the EM data even further. Differences in this analysis may exist by site, but the small sample size prevented site stratification.

In the regression analyses to predict variability in the EM adherence measure, hair drug levels explained the most variability at both sites, although plasma levels performed nearly as well in Kenya. The combination of hair and plasma drug concentrations explained nearly as much of the variability as combinations of three measures. This finding argues for combining short-term (e.g., plasma) and long-term (e.g., hair) measures to obtain a comprehensive picture of adherence over time. The incremental value of the other adherence measures either singularly or in combination was higher in the Kenya site than the Uganda site. Reasons for this finding are unclear, but could be related to the higher correlations between measures at the Kenya site. Both approaches showed generally increasing ability to discriminate adherence patterns with an increasing number of measures. However, the choice for the number of adherence measures used in a given study is dependent on resources and competing scientific priorities. Multiple measures may be most appropriate when an accurate assessment of adherence is needed for assessing biologic efficacy and/or triggering interventions [29]. If only one measure can be employed, our analysis suggests that either hair levels or EM monitoring can be used, with the former providing an overall measure adherence and the latter indicating patterns.

The major strength of this analysis includes the inclusion of multiple adherence methods, including concurrent collection of both non-PK and PK measures. This data provide a rare opportunity for aligning the adherence measures over the same time period. Although there were differences in the comparisons among measures by site, the data taken as a whole show that correlations were highest between the EM measure and each of the PK measures, with hair levels exhibiting the best performance.

There are important limitations with this analysis. First, because we were interested in only the active arm of the two studies and the daily dosing strategy, the sample size was small and in some instance the data analyses could not be stratified by site. The small sample size limits us making conclusive assertions about the hypothesised results. Therefore, the results of this analysis need to be interpreted with caution. Second, the sample was predominantly men, which could limit extrapolation of results to populations dominated by women. Third, we chose to use EM as the comparison standard for this analysis, as it has been previously shown to be an excellent measure for day-to-day adherence behaviour [25, 26]. It has also been shown to meaningfully discriminate between high rates of daily dosing compared to low rates [8]. Similar analyses could also be pursued to assess the ability to discriminate adherence per the PK measures. Finally, further investigations on PrEP adherence need to be carried out in routine care settings because adherence patterns in a placebo-controlled trial environment could differ from the real-world setting where participants are taking a product already known to work.

In sum, this analysis indicates that objective measures of adherence are more informative than self-report and that employing multiple measures of adherence is likely to increase accuracy in estimating adherence behaviour. Specifically, the combination of adherence measures that reveal behaviour over the long-term (hair) plus another short-term measure (plasma) may be useful to assess patterns of drug-taking. Important differences in measure performance, however, may arise in different populations and must be considered carefully for any study, most feasibly in the planning stage. Hair appears to be a good choice to include as an adherence measure where limited testing is available. Drug levels in dried blood spots also assess longer-term adherence [30], but were not collected in this particular study. Given the importance of adherence in achieving effective protection against HIV with PrEP, these data provide valuable insights in the design of future prevention studies.