Association of Immunosuppression and Viral Load With Subcortical Brain Volume in an International Sample of People Living With HIV

Key Points Question Are HIV plasma markers that are universally used to monitor immune function and treatment response associated with subcortical brain volumes in clinically and demographically heterogeneous HIV-infected individuals? Findings In this cross-sectional study of 1203 HIV-infected adults, lower current CD4+ T-cell counts were associated with smaller hippocampal and thalamic volumes independent of treatment status, although in the subset of participants not receiving treatment, they were associated with smaller putamen volumes. Across all participants, detectable viral load was associated with smaller hippocampal volumes, but in the subset of participants receiving HIV treatment, detectable viral load was also associated with smaller amygdala volumes. Meaning In a heterogeneous population of HIV-infected individuals, volumes of structures in the limbic system were consistently associated with plasma markers.


=
( 1 + 2 ) √ 1 2 * √ where 1 and 2 are the sample size for each group. For continuous variables, effect sizes were estimated by converting tvalues to partial correlation coefficients: For both equations, is the degrees of freedom as outputted by the 'nlme' package in R (version 3.2.3).

Age Feasibility Study Mega-Analysis
To ensure sufficient power to detect associations between brain measures and variables of interest from data pooled across studies, a preliminary analysis tested for associations between age and eight brain volumes. Random effects multivariable linear regressions were performed. Three fixed-effects covariates were included in the model: sex, the interaction between age and sex, and estimated total intracranial volume to adjust for variability in head size. To account for scanner effects, the data collection site was used as the random-effects grouping variable; there were 19 sites in all. Statistical analyses were conducted with the 'nlme' package in R (version 3.2.3). Effect sizes were estimated using the r-value (partial correlation coefficients) after accounting for all covariates. Significance was determined using the Bonferroni correction threshold (p ≤ 0.0063).
Across the entire sample (N=1,203), older age was associated with smaller volumes of all subcortical structures and larger ventricular volumes (r-value range: 0.13-0.30), consistent with studies in seronegative adults 1 . Sex was a significant predictor in models for associations between age and the thalamus, putamen, globus pallidus, hippocampus, and amygdala (p ≤ 0.0063). Age effects in the subset of participants taking cART, males, and females mirrored that of the full group. Participants not taking cART showed significant associations between age and caudate, putamen, globus pallidus, nucleus accumbens, and ventricular volumes. No significant age by sex interaction was detected.

HIV Subcortical Volume Comparisons with Healthy Lifespan Centiles
In a recent study from the ENIGMA lifespan working group, subcortical volume trajectories across the lifespan were derived from 18,605 healthy participants pooled from 88 international samples 2 . As there was a lack of site and scanner matched HIV-negative controls in our ENIGMA-HIV sample, we used the centile curves to extrapolate how subcortical volumes in HIV+ individuals compared to age-matched healthy individuals. Given the wide variability in population demographics and scanners from which the volumes were extracted and reported curves estimated, they may incorporate sufficient variability to serve as a surrogate for matched controls.
For each subcortical volume, we used the discrete centile values reported for every 10 years of age in Dima et al.
supplementary Table S6 to estimate whether the distribution of volumes in our HIV+ sample are similar to that expected for an age-matched healthy population.
We did this under the assumptions that: 1. for each volume, at each age provided, the distribution of volumes across the population was Gaussian, and 2. for all ages between those provided, (e.g, any age between 40 and 50), the distribution could be estimated by linearly interpolating the mean and variance of the data from the two closest ages. Note: In Dima et al. Table   S6, the ages for which centile volumes are reported span 20 to 90 years of age discretized at 10-year intervals.
We used the 50th centile (C50; 50th percentile) column as the mean of the distribution and estimated the standard deviation by assuming that the 75th centile (C75) is 0.674 standard deviations above the mean, as would be the case for a Gaussian distribution. Therefore, as z = (x-µ)/σ, we calculated the standard deviation for each structure, at each age provided, using the formula: As the centiles were calculated non-parametrically, the distribution was not exactly Gaussian; the standard deviation calculated from the 25th centile was (in nearly all cases) smaller than that calculated with the 75th centile. Assuming a wider distribution is a more conservative approach as it is harder to reject the null with greater variability or more noise; we therefore used the larger of the two estimates.
Next, for each age represented in our HIV+ sample, we estimated the distribution of each subcortical volume by linearly interpolating the volumes at the provided surrounding ages (e.g., 40 and 50 for someone 43 years of age). As each individual's volume is from a different distribution, we standardized the subject-level volumes by converting them to zscores and percentile estimates: (1) We first determined the z-score associated with the observed volume of that subject with respect to that expected distribution; (2) Then, for each individual observation, we randomly sampled 10,000 instances from the expected distribution and determined how many times the observed volume was less than or equal to the randomly drawn volumes. This resulted in an approximate percentile for the individual.
The probability density function (pdf) for the set of 1,203 paricipants' z-scores for each volume was then compared to the pdf of 1,203 randomly sampled draws from the standard normal distribution. The Kullback-Leibler divergence was then calculated using the KLD function from the "LaplacesDemon" library available in R, and the intrinsic discrepancy was used as a measure of deviation between the two distributions. These KLD values are provided in eTable 6. To determine whether this deviation was greater than would be expected by chance, and derive a non-parametric p-value estimate, we drew 1,203 random values from the normal distribution and used the KLD intrinsic discrepancy measure to compare their pdf to that obtained from the original random sample to which we compared the pdf of the HIV+ distribution; we then repeated this 10,000 times to determine how many times the observed was less than the random sample. The number of times our observation was less extreme (or one, whichever was greater) was then divided by total number of permutations to derive a p-value for the divergence.
A second method for determining whether the observed volumes were likely to come from the distributions derived from healthy controls made use of the percentiles derived as described above. If the set of HIV+ observations was derived from the healthy distribution, the percentiles are likely to be uniformly distributed. Therefore, we binned a uniform distribution [0,1] into 100 bins and performed a chi-squared goodness of fit test to determine whether the observed percentiles were approximately uniformly distributed. The chi-squared test was performed using the "chisq.test" function in R, with 10,000 simulated values. These p-values are also provided in eTable 6.
We note, all volumes for our HIV+ population appear to be from distributions that are not those described using healthy control data from multiple international imaging centers. Therefore, we anticipate significant case/control differences are likely across all structures, although the lack of HIV-negative controls from all sites prevented us from testing this directly.
For approximate visualization of distributions, we tabulated the number of HIV+ individuals in a given age range whose subcortical volumes fell between each pair of healthy centile values (e.g., for the hippocampus, how many HIV+ individuals ≤ 25 years of age and > 35 years fell between the C10 (10th percentile) and C25 (25th percentile) hippocampal volume values for age 30). The resulting histograms are shown in eFigures 3-6.

Validation Analyses
Two sets of validation analyses were performed: 1) dichotomizing CD4+ cell count based on the AIDS-defining threshold of 200/μL; and 2) defining a common dVL threshold across sites (400 copies/mL, the highest detection limit for any site). As in the primary analyses, multivariable random effects linear regressions were performed to evaluate associations between regional brain volumes and binary variables indicating 1) a CD4+ ≤ 200/μL (1)   for every year of age), standard errors (SE), and p-values from associations between age and regional brain volumes across all HIV+ participants, and by cART status and sex.   Effect sizes (d) for associations between regional brain volumes and dVL > 400 copies/mL across HIV+ individuals, and by cART status and sex. We note that dVL associations in those off treatment included only a limited number of individuals with undetectable VL (n=24), and should be interpreted with caution.