This study was a secondary cross-sectional analysis of stored baseline serum samples and neuropsychological test data obtained from a subset of participants who were enrolled in the Johannesburg, South Africa, and Harare, Zimbabwe, sites of the AIDS Clinical Trials Group (ACTG) A5199 study from 2006 to 2010. The A5199 study was a neurological sub-study of the PEARLS study (A5175), which was a multinational randomized trial across resource-limited settings focused on treatment strategies for people living with HIV starting their first ART regimen. The methodologies and outcomes of A5199 (ClinicalTrials.gov NCT00096824) and A5175 (ClinicalTrials.gov NCT00084136) have been detailed in prior publications [21, 22]. Key inclusion criteria for the A5199 study were an age > 18 years, a documented HIV-1 infection, a CD4 count < 300 cells/mm3, a Karnofsky performance score > 70, limited prior ART exposure (less than 7 days), absence of significant psychiatric pathologies, and no active substance abuse or dependence at study entry determined by site clinician.

For this analysis, participants were chosen through criterion-based purposive sampling: enrollment at the Johannesburg or Harare sites of the A5199 study, availability of neuropsychological test results, presence of pre-stored serum samples, and consent for specimen utilization. Original participant consent was acquired during the parent study, adhering to ethical protocols. This analysis received approval from the Human Research Ethics Board of each participating site and the University of KwaZulu-Natal Biomedical Research Ethics Committee.

Serum concentrations of mature brain-derived neurotrophic factor (mBDNF) and its precursor proBDNF were quantitatively assessed using commercially available enzyme-linked immunosorbent assay (ELISA) kits (Biosensis Pty Ltd., Thebarton, SA, Australia). The serum specimens, previously stored at − 70 °C, were processed in accordance with the manufacturer’s protocol. To minimize assay variance, serum levels of proBDNF and mBDNF from each subject were measured on the same plate to reduce the potential variation between plates/kits. All samples were tested in duplicate, and the two tests were averaged. The intra-assay coefficients of variation for the mBDNF and proBDNF were 2.7% and 2.2%, respectively. Analysts performing the assays were blinded to the clinical characteristics and neuropsychological test results of participants. Serum mBDNF and proBDNF were log-transformed for analysis due to right skewness.

Participants at both research sites underwent a battery of standardized neurocognitive assessments designed to evaluate a range of cognitive and motor functions. Fine motor dexterity and speed were assessed using the Grooved Pegboard Test for both dominant and non-dominant hands, as well as the Finger Tapping test. Gross motor skills and gait velocity were measured using the Timed Gait test, and verbal fluency was assessed via the semantic verbal fluency test [23,24,25]. These specific neurocognitive tests were selected to ensure brevity and practicability across multicenter trials in various resource-constrained settings, while also minimizing the influence of linguistic and cultural biases [26]. All test scores were standardized into z-scores adjusted for site, age (above and below 35 years), sex (male/female), and education (above and below 10 years) using normative data from the International Neurocognitive Normative Study (INNS ACTG A5271), which drew participants from HIV testing centers and local clinics to reflect the at-risk population [27]. Z-scores of Grooved Pegboard and Timed Gait tests were computed to let higher scores reflect better performance. The primary outcome was the NPZ-6 composite score, representing overall neurocognitive function, calculated as the mean of individual test z-scores; lower scores denoted reduced cognitive ability. As secondary outcomes, z-scores for each test were analyzed independently, with scores above the mean indicating better than average cognitive performance.

Covariates were selected a priori based on prior empirical evidence highlighting their confounding potential in the relationship between serum brain-derived neurotrophic factor (BDNF) levels and cognitive performance. These covariates included age (quantified in years) at the initial study visit, years of educational attainment, biological sex, body mass index (BMI, calculated as kg/m2), the CD4 to CD8 cell count ratio at study entry, log-transformed plasma HIV RNA levels, and the specific study site from which the data were collected.

Baseline characteristics were detailed using descriptive statistics. For continuous variables, differences between sites were evaluated using the t-test or Mann–Whitney U test, depending on the data distribution. Categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate.

Three linear regression models were developed. Model 1 evaluated the association of NPZ-6 (composite measure of cognitive function) with log-transformed mBDNF levels, adjusting for confounding factors including age, sex, education, study site, CD4:CD8 ratio, HIV viral load, and BMI. Model 2 assessed the relationship between NPZ-6 and log-transformed proBDNF levels, and model 3 examined the association with the log-transformed mature BDNF to proBDNF ratio. Forest plots illustrated these associations visually. Each model’s assumptions were rigorously tested, including linearity (assessed via scatterplots of residuals against fitted values), independence (Durbin-Watson test), homoscedasticity (Breusch-Pagan test and residuals vs. fitted values plot), and normality of residuals (histogram, QQ-plot, and Shapiro–Wilk test). Additionally, multicollinearity was assessed using variance inflation factors (VIFs). Outliers were identified using the standard deviation and interquartile range methods, complemented by Cook’s distance to detect influential observations. We conducted sensitivity analyses by excluding these outliers.

Quantile regression analyses at the 25%, 50% (median), and 75% quantiles were conducted for both mBDNF and proBDNF. This approach allowed for an understanding of the relationship across the distribution spectrum of cognitive function.

In addition, we employed structural equation modeling (SEM) to further examine the complex interrelationships between neurotrophic factors, demographic variables, and individual cognitive test scores. The SEM was estimated using the weighted least squares mean and variance adjusted (WLSMV) method to accommodate the non-normal distribution of some variables. Model fit was evaluated using various fit indices including Comparative Fit Index (CFI ≥ 0.95), Tucker-Lewis Index (TLI ≥ 0.95), root mean square error of approximation (RMSEA ≤ 0.05), and standardized root mean square residual (SRMR ≤ 0.08). To improve model fit, modification indices were inspected, suggesting potential paths for inclusion in a revised model.

All tests were two-sided, and the level of statistical significance was set at p < 0.05. All statistical analysis was conducted using R software version 4.3.1 (R Core Team, 2023, Vienna, Austria).