Using a cohort study design with two-phase sampling ¹⁶, we included all mothers in the maternal ARV or infant NVP arms with ≥1 available plasma or breastmilk sample between 2-24 weeks postpartum and an HIV transmission event between 2-28 weeks postpartum (n=31, 94% of transmission events). We also included 15% of mothers randomized to the two treatment arms of BAN who did not transmit HIV to their infant by 28 weeks postpartum (n=232); sampling was primarily based on stored specimen availability. When multiple births occurred only first-born multiples were included (n=5). Our analysis included a total of 263 mothers at risk of transmitting HIV via breastmilk from 2-28 weeks postpartum.

Maternal adherence was calculated using pill counts taken by trained pharmacy staff on contiguous visits for the following postpartum time intervals: 2-4, 8-12, and 13-18 weeks. Specifically, adherence was calculated by dividing the difference in the number of pills distributed and returned by the number of pills expected to be consumed under perfect compliance.¹⁷ The observed distribution of maternal pill count adherence was skewed, with most women randomized to maternal ARVs having adherence values greater than 90%. The observed adherence distribution was anticipated due to mothers' high motivation to adhere to study drug in order to prevent HIV transmission to their infant. Mothers randomized to the infant NVP arm were assigned a maternal adherence value of zero for all intervals. In our main analyses, maternal adherence was categorized as poor (0-80%), partial (81-98%) and near perfect (>98%) using disjoint indicator variables. Categories were chosen based on previous studies and the observed adherence distribution, allowing for a relatively balanced number of observations across three adherence categories (poor, partial, and near perfect adherence). In sensitivity analyses, we treated adherence as a dichotomous variable (>90% versus ≤90%, and >80% versus ≤80%).

HIV RNA was quantified from blood plasma and whole breastmilk at enrollment, 2, 6, 12, 18, and 24 weeks postpartum. In addition, breastmilk HIV RNA was measured at 4 and 8 weeks postpartum. Plasma VL was quantified using the Abbott RealTime HIV assay (Abbott Molecular, Des Plaines, IL) by the 0.6ml protocol according to the package insert (lower limit of quantitation 40 copies/ml). Breastmilk VL was quantified from 0.6 ml whole breastmilk pre-treated with 209μl Abbott RNA sample prep lysis buffer and 60μl Abbott Proteinase K (53°C incubation for 20 min) using the Abbott RealTime HIV assay (lower limit of quantitation 56 copies/ml). HIV RNA concentrations that were detected but below the limit of quantitation were assigned a value of 39 for plasma and 55 for breastmilk, and RNA concentrations that were undetectable were assigned a value equal to 50% of the lower limit of quantitation (plasma: 20, breastmilk: 28).

Infant HIV status was determined at 2, 12, 28, and 48 weeks by Roche Amplicor 1.5 DNA PCR (Roche Molecular Systems, Pleasanton, CA, USA). PCR positive results were confirmed by testing an additional blood specimen. The window of infection was narrowed by testing infant dried blood-spots collected at 4, 6, 8, 18, 24, and 32 weeks. Breastmilk transmission was defined as first detection of HIV infection by PCR in infant blood between 2-28 weeks of age. Infants who tested HIV-1 DNA PCR positive at either birth or by two weeks of age were considered to be infected in utero or intrapartum and were excluded from all BAN analyses of breastmilk transmission. Sensitivity of HIV-1 DNA PCR for diagnosing intrapartum HIV transmission has been shown to reach 93% by 14 days of age, thus most perinatal infections will be detected within two weeks of birth.¹⁸

We used generalized linear mixed models with a random effect (i.e., random intercept) for each mother-infant pair. Identity link and Gaussian distribution were used to estimate differences in log₁₀ plasma VL by adherence category, and logit link with binomial distribution were used to estimate odds ratios for detectable breastmilk VL by adherence and plasma VL category. Adherence was paired with VL measured at the end of the adherence time interval. Therefore, only VL information from 6, 12, and 18 weeks postpartum was used when adherence was the exposure. The reciprocal of the study inclusion probability was used as a sampling weight in all regression analyses to account for the two-phase sampling.¹⁹

Single and multivariable Cox models were used to estimate the relative hazard of breastmilk HIV transmission by 28 weeks. Breastmilk VL was treated as a time-varying covariate in Cox models, lagged by one interval. Time until infant HIV infection was right censored at the first occurrence of the following: time of infant death, time of infant's last PCR negative test if lost to follow-up or at 28 weeks if the infant remained HIV-uninfected, or at reported cessation of breastfeeding unless HIV transmission occurred within 30 days of reported cessation.¹⁴

Effect measure modification was assessed by comparing unadjusted and adjusted estimates and 95% confidence intervals using an interaction term between the exposure and variable of interest. Study arm was identified as an effect measure modifier when assessing the association between breastmilk viral load and HIV transmission. No other effect measure modification was identified. A directed acyclic graph was used to identify potential confounders and a minimally sufficient adjustment set.²⁰ Potential confounding variables consisted of randomization assignment, demographic characteristics, and health status information collected at enrollment (hereafter referred to as baseline) (Table 1). Missing adherence and VL measures were accounted for using multiple imputation²¹ assuming data were missing at random.

Monte Carlo simulation was used to estimate the number of transmissions that would have occurred if all mothers randomized to maternal ARVs (n=848) had been 0%, 90%, and 100% adherent. Data sets with the same sample size as the maternal ARV arm of BAN were simulated under four adherence scenarios. Under the first scenario, longitudinal adherence values were generated by randomly sampling from the observed adherence values for all mothers in the maternal ARV arm. Under the remaining scenarios, all mothers were simulated to be 0%, 90%, and 100% adherent. Simulation results from the first scenario were compared with the actual observed data from BAN to ensure the simulation model was correctly calibrated. For all scenarios, time-varying breastmilk VLs were then randomly generated using the fitted mixed effects logistic regression model from the observed data conditional on simulated adherence values. A transmission time of infant HIV infection was then simulated based on a Cox model with constant baseline hazard and simulated time-varying breastmilk VL. Censoring times were simulated in accordance with the 12% observed dropout in BAN. Finally, the number of infant infections estimated to have occurred with 0%, 90%, and 100% adherence were compared.

All data analyses were conducted using SAS version 9.3 (SAS Institute, Cary, North Carolina, USA) and R version 3.0.2 (R Core Team, Vienna, Austria).

At least one paired maternal adherence and plasma HIV VL was available for 245 (93%) mothers meeting our study inclusion criteria. Among mothers in the maternal ARV arm, overall mean adherence was 88% [median 0.96, IQR 0.86-1.00] at 2-18 weeks postpartum and most received a boosted protease inhibitor regimen of zidovudine, lamivudine, and lopinavir/ritonavir (70%). Based on the unadjusted model, having partial adherence was associated with a 0.82 (95% confidence interval (CI) 0.55-1.09) log₁₀ reduction in plasma VL compared to having poor adherence (Table 3). Having near perfect adherence was associated with a 0.69 (95% CI 0.45-0.93) log₁₀ reduction in plasma VL compared to poor adherence. Adjusting for nutritional randomization and baseline maternal characteristics yielded similar results. Multiple imputation produced similar though attenuated associations between increased adherence and lower plasma VL.

Maternal adherence had an analogous association with suppression of virus in breastmilk. Among 211 (80%) mothers with paired adherence and breastmilk HIV VL data, partial adherence and near perfect adherence were associated with a 76% (95% CI 28-92%) and a 62% (95% CI 14-83%) relative reduction in the odds of having detectable breastmilk VL, respectively, compared to poor adherence (Table 4). Adjustment for potential confounding variables did not appreciably change the estimates or precision. The associations remained similar but less precise when using multiple imputation. Results are not presented for the strata of mothers randomized to infant NVP, as they did not contribute to the >80-98% or >98% adherence categories.

In sensitivity analyses, adherence was treated as a dichotomous variable (>90% versus ≤90%, and >80% versus ≤80%). Choice of adherence categorization did not change our overall conclusions, as higher adherence remained associated with reduced log10 plasma viral load and reduced odds of detectable breastmilk viral load, compared with lower adherence (data not shown).

Among the 221 (84%) mothers with ≥1 paired plasma and breastmilk HIV VL, mothers with detectable breastmilk VL typically had detectable plasma VL. Two (<1%) mothers had detectable breastmilk VL (56 and 77 copies/ml) and undetectable plasma VL at 6 weeks postpartum. Mothers with a detectable plasma VL had 59 (95% CI 21-169) times the odds of having a detectable breastmilk VL, compared with mothers with an undetectable plasma VL (Table 4). Adjusting for study arm, baseline maternal CD4+ count, and baseline maternal plasma VL resulted in a slightly attenuated association between detectable plasma and breastmilk VL (OR 40, 95% CI 15-107). Using multiple imputation produced similar though attenuated findings.

At least one plasma VL between 2-24 weeks postpartum was available for 27 of 33 (82%) mothers of HIV-infected infants. Among these, one transmission event occurred from a mother with undetectable baseline plasma VL. However, transmission occurred at 24 weeks postpartum and after three VL measures ranging from 8,000-108,000 copies/ml. The remaining 26 transmitting mothers had baseline plasma VL>3500 copies/ml.

Among 116 mothers with complete plasma VL information, five had undetectable plasma VL at all measured time points between 2-24 weeks. All five were in the maternal ARV arm, also had undetectable breastmilk VL at all measured time points, and none transmitted HIV to their infant. Additionally, plasma VL from the time point immediately preceding transmission was available for 21 mothers, with VL ranging from 102 to >636,000 copies/ml. No mother who consistently maintained a plasma viral load <100 copies/ml transmitted HIV to her infant during breastfeeding.

Complete time-dependent breastmilk VL and infant HIV status were available for 134 (51%) mothers. Of these, 15 experienced a HIV transmission event between 2-28 weeks postpartum (maternal ARV arm: 11, infant NVP arm: 4). Among these 15 mothers, 11 (73%) had at least one detectable breastmilk VL before transmission occurred, and 8 (53%) had a detectable breastmilk VL at the last measured time point before transmission.

Using multiply imputed data (n=31 transmission events), detectable breastmilk VL was associated with 3.8 (95% CI 1.2-12.1) times the rate of breastmilk HIV transmission, compared with undetectable breastmilk VL after adjusting for study arm and baseline maternal characteristics (Table 5). Breastmilk VL had a somewhat diminished association with HIV transmission in unadjusted analyses (HR 2.9, 95% CI 1.0-8.4). In complete case analyses, detectable breastmilk VL had a similar unadjusted and increased adjusted association with HIV transmission (unadjusted HR 2.7, 95% CI 1.4-5.4; adjusted HR 7.4, 95% CI 3.2-17.1). However, the adjusted complete case analysis results should be interpreted with caution due to the number of adjustment factors and limited number of transmission events.

Using Monte Carlo simulation, the calibration model predicted the same number of HIV transmissions between 2-28 weeks postpartum as observed in the maternal ARV arm of BAN (n=21). If all 848 mothers randomized to maternal ARVs had been 100% adherent, the model predicted 20.4 transmissions (95% prediction interval (PI) 12, 30). In contrast, if all of these mothers had been 0% adherent, the model predicted 42.3 transmissions (95% PI 30, 56), indicating an average of approximately 22 infections averted due to perfect maternal ARV adherence. Simulations where all 848 women were 90% adherent yielded similar results (20.5 transmission, 95% PI 12, 30) to the 100% adherence scenario.