Motor skills were assessed by the Grooved Pegboard Dominant and Non-dominant Hand tests

Leveraging comprehensive neuropsychological data from the large-scale cohort of the HIV Neurobehavioral Research Program at the University of California-San Diego, we used novel machine learning methods to identify differing profiles of cognitive function in PWH and to evaluate how these profiles differ between women and men in sex-stratified analyses. Rather, than using traditional cognitive domain scores, we used each of the NP test outcomes given that prior studies indicate that the correlation of NP test scores does not map to traditional domain scores in PWH. Furthermore, we determined how sociodemographic , clinical and biological factors related to cognitive profiles within women and men. Based on previous studies among PWH , we hypothesized that the machine learning approach would identify distinct subgroups of individuals with normal cognitive function, global cognitive impairment, and domain specific cognitive impairment. We further hypothesized that groups with domain-specific cognitive impairment would differ by sex, with WWH showing more consistent memory and processing speed impairment than MWH. Finally, we expected that similar sociodemographic/clinical/biological determinants would distinguish cognitive profiles among WWH and MWH; however, in line with previous research , we expected that depressive symptoms would be more strongly associated with cognitive impairment profiles among WWH than MWH. NP test performance was assessed through a comprehensive, standardized,indoor grow methods battery of tests that measure seven domains of cognition, including complex motor skills, executive function, attention/working memory, episodic learning, episodic memory , verbal fluency, and information processing speed.

Executive functioning was assessed by the Trail Making Test -Part B and the Stroop Color and Word Test interference score . Attention/working memory was assessed by the Paced Auditory Serial Addition Task . Episodic learning was assessed by the Total Learning scores of the Hopkins Verbal Learning Test-Revised  and the Brief Visuospatial Memory Test-Revised . Episodic memory was assessed by the Delayed Recall and Recognition scores of the HVLT-R and BVMT-R. Verbal Fluency was assessed by the “FAS” Letter Fluency test . Information processing speed was assessed by the WAIS-III Digit Symbol Test , the TMT-Part A, and the Stroop Color and Word Test color naming score. Raw test scores were transformed into age-, education-, sex-, and race/ethnicity-adjusted T-scores based on normative samples of HIV-uninfected persons . The use of demographically-adjusted T-scores are intended to control for these demographic effects as they occur in the general population.We examined sociodemographic, clinical, and biological factors associated with cognitive impairment in the literature and available with enough participants to be adequately powered in analyses. Sociodemographic factors included age, years of education, and race/ethnicity. Although these factors were used to create the T-scores, there can still be remaining demographic associations with cognition within clinical populations such as PWH. For example, there is considerable interest in the possibility of abnormal cognitive aging PWH; also, in general, older PWH tend to have had their infections longer, may have had longer periods without benefit of suppressive ART, and more history of worse immunosuppression.

Clinical factors included functional status as indicated by the number of daily activities with decreased independence from the Instrumental Activities of Daily Living questionnaire from the modified version of the Lawton and Brody Activities of Daily Living Questionnaire , reading level based on the Wide Range Achievement Test-4 Reading subtest , self-reported depressive symptoms on the Beck Depression Inventory versions I or II , and diagnosis of lifetime and current major depressive disorder as well as lifetime alcohol, cannabis, or other substance use disorder based on the Composite International Diagnostic Interview using DSM–IV criteria . Biological factors included HIV disease variables such as current CD4+ T-cell count, lowest CD4+ T-cell count ever recorded , plasma HIV viral load, estimated duration of HIV disease, current use of ART, current use of anticholinergic-based medications , Hepatitis C co-infection, and the cardiovascular comorbid conditions of hypertension, hyperlipidemia, and diabetes. All 13 NP tests were used to find groups of similar cognitive profiles within each participant subset and in the total sample using a pipeline that consisted of dimension reduction with Kohonen self-organizing maps followed by clustering to identify profiles based on those reduced dimensions. SOM was implemented using the Kohonen package in R . SOM is an unsupervised machine learning technique used to identify patterns in high-dimensional data by producing a two-dimensional representation consisting of multiple nodes where each node is a group of one or more individuals with similar cognitive profiles and the location of the nodes within the 2-D representation is also a metric of similarity. Unlike probabilistic models, each individual can only be assigned to one node.

The SOM grid consisted of a 10 × 10 hexagonal grid of nodes and the number of clusters for the final profiling was selected by looping over models created from 3 to 20 clusters and selecting the number that had the best fit based on entropy. Similar nodes were then clustered using the MClust package . MClust is an R Software package used for model-based clustering using finite normal mixture modeling that provides functions for parameter estimation via the Expectation-Maximization algorithm with an assortment of covariance structures which vary in distribution , volumes , shape , and orientation . This program identifies the best model based on entropy . Once the clustering of the nodes was completed, cluster profiles were assigned to the individuals associated with that node. By using SOM and MClust in sequence, we were able to achieve fine-tuned clustering based on patterns of performance in cognitive testing. Factors predicting profile membership between each impaired and unimpaired profile in the overall sample and within each group were explored by creating a predictive Random Forest model using the Caret package in R and then extracting variable importance . RF is an ensemble machine learning model based on classification trees that results in powerful prediction models based on non-linear combinations of subsets of input variables. Prior to model creation, the Synthetic Minority Over-sampling Technique with the DMwR package was used to control for bias due to any imbalance in the number of cases. RF models were created using internal validation using a 10-fold resampling method repeated 5 times. Pre-processing before RF creation involved removing variables as predictors if they had low variance or if they had >50% missing data. Any missing data in the remaining variables was imputed using the Multivariate Imputation by Chained Equations  package in R using random forest imputations. ROC confidence intervals were calculated using the pROC package in R with 2,000 stratified bootstrap replicates . Variable importance of all variables included in the RF models was used as the outcome metric of the predictive power of each variable. Variable importance is a scaled number [0–100] that indicates how important that variable is to the final predicted outcome in that model. For each tree in the RF model,cannabis dryer the out-of-bag portion of the data is recorded and repeated after permuting each predictor variable. The difference between the accuracy with and without each variable is averaged over all trees and then normalized by the standard error. For visualization, all variables were plotted by relative variable importance, and attention was given to the top 10 variables in each profile. Variable importance indicates how much that variable contributes to overall prediction accuracy, but as RF is non-linear model it does not indicate directionality. While the analysis pipeline and packages used along with the parameter inputs are stated above, we have added our code into a Supplementary Material to facilitate rigor and reproducibility. In this large-scale study using a novel pipeline combination of machine learning methods, we provide further evidence in support of heterogeneity in cognitive function among PWH. Our results do not negate the heterogeneity in cognitive function in HIV-uninfected individuals but rather highlights the heterogeneity among PWH that can often be masked by a dichotomous HAND categorization. In the total sample, we identified an unimpaired profile, a profile of relatively weak auditory attention and episodic memory, and a global weakness profile. As expected, given the relative sample sizes, the cognitive patterns in the total sample were in greater alignment with those found among MWH compared to WWH. Similar to results in the overall sample, we identified an unimpaired profile and a global weakness profile in MWH; however, unlike the overall sample and inconsistent with hypotheses of domain specific cognitive impairment profiles in both MWH and WWH, MWH demonstrated a profile with relative strengths in attention and processing speed.

Conversely, there were no unimpaired, cognitive strength or global weakness profiles among WWH. Rather, as hypothesized WWH demonstrated cognitive profiles reflecting a global weakness and domain-specific impairment including a weakness in learning and memory and motor skills. These findings suggest that sex and the sociodemographic factors associated with female sex within the HIV-infected population contribute to the heterogeneity in cognitive function among PWH. Studies examining cognitive function in combined samples of men and women may mask important sex differences in cognitive functioning among PWH, particularly in maledominant samples such as the current sample. These sex differences in cognitive profiles among PWH may result from biological sex differences and/or the psychosocial factors that tend to characterize WWH more than MWH . Biological sex differences include those seen in the general population such as sex steroid hormones , female-specific reproductive events and genetic factors or previously-reported sex differences specifically in HIV disease characteristics unmeasured herein . Regardless of the underlying mechanism, characterizing these sex differences in cognitive functioning among PWH can provide inroads to identifying mechanisms of cognitive dysfunction and optimizing risk assessments and diagnostic and therapeutic strategies for each sex. A notable sex difference in profiles was the lack of the unimpaired or cognitive strength profile among WWH that was observed among MWH. Our cognitive profile analyses are in line with prior studies that suggests that WWH are often but not always, more likely to demonstrate cognitive deficits than MWH . Our analysis suggests that the impairment manifests more often as domain-specific impairment in women than in men that may not be revealed in a more cross-domain summary measure like GDS or global Tscores. This female vulnerability to cognitive deficits is thought to reflect sociodemographic differences whereby low education and socioeconomic status and their associated psychosocial risk factors are more prevalent among WWH vs. MWH . These psychosocial risk factors can have adverse effects on the brain that lower cognitive reserve , suggesting that interventions geared toward addressing these psychosocial factors should be a priority for WWH and/or for women who are at increased risk of HIV. In support of these studies, Sundermann et al. found that the higher rates of cognitive impairment in WWH vs. MWH were eliminated after adjusting for the lower reading level that characterized WWH compared to MWH. Biological differences may also contribute to sex differences in the pattern and magnitude of cognitive impairment in PWH including disease characteristics, brain structure/function, sex steroid hormones and female-specific hormonal milieus . There is also evidence to suggest that WWH may be more cognitively susceptible than MWH to the effects of mental health factors . As mentioned, only women demonstrated more domain specific cognitive profiles including weakness in motor functioning and relative weakness in learning and memory. Similarly, previous studies report that learning, memory, and motor functioning are among the domains in which cognitive impairment is more common among WWH vs. MWH and these differences persisted after adjusting for HIV RNA and CD4 counts . These sex differences in domain-specific impairment may reflect psychosocial factors , biological factors , or interactions among them. Although women in general demonstrate relative advantages in verbal memory and fine motor function compared men likely due, at-least in part, to the effects of estrogen on the developing brain and the neuroprotective effects of circulating estradiol , the menopause transition has been associated with declines in verbal memory and motor function . The mean age of women in our study was 41 suggesting that a portion of women may be experiencing cognitive deficits associated with reproductive aging. Germane to the learning/memory impairment in WWH, women are more vulnerable to the negative effects of stress hormones on hippocampal-dependent tests compared to men . This finding may be particularly relevant to the current sample considering the high prevalence of psychosocial stressors among WWH including childhood trauma and domestic violence .