We used machine learning models to identify distinct homogenous subgroups (profiles) in the largest dataset to date in VS-WWH and HIV-uninfected women. Separate patterns of cognitive performance, as well as associated factors of those patterns among each subgroup of women, were identified. The factors identified allow for screening and intervention, including potentially changing non-ART medications, as well as mental health and substance use screening and intervention.

In the context of viral suppression, we identified several profiles with distinct patterns of performance across 17 NP outcomes. While these profiles are statistically-derived, some of the profiles found here parallel commonly identified patterns in other neurological conditions or processes. Among the virally suppressed group, Profile 1-VS revealed a unique pattern reflecting exclusive weaknesses in cognitive sequencing (LNS Attention and Working Memory) and motor set-shifting (TMT-Part B). While to our knowledge, this combination of isolated deficits in cognitive sequencing and motor set-shifting has not been appreciated in other disease populations, specific deficits in cognitive sequencing/verbal working memory have been observed in individuals with schizophrenia and their first-degree relatives (25). Additionally, McDonald et al. (26) identified specific problems with motor set-shifting (TMT-Part B) in individuals with frontal lobe epilepsy. In contrast to the very specific weaknesses identified in Profile 1-VS, Profile 2-VS reflects general slowing, which is most often associated with typical (i.e., “healthy”) aging (27). Profile 3-VS, characterized by poor encoding and recognition with intact retention, is more of a typical HIV-associated profile (28) compared to mild cognitive impairment due to Alzheimer's disease (AD) (29). Profile 4-VS showed a mostly amnestic profile with some evidence of cognitive slowing, as can be observed in AD or in AD with vascular contributions (30). This profile is similar to Profile 4-UN, which reflected an amnestic profile that is often observed in typical AD (31). Profile 5-VS, showing intact memory storage and manual speed/dexterity, but weak or impaired attention, processing speed, learning, and executive functioning is similar to what is observed in individuals with diffuse frontal-subcortical small vessel disease (32). Interestingly, a profile did not emerge among VS-WWH reflecting specific motor slowing which has been linked to HIV infection. This is consistent with prior cross-sectional WIHS analyses, where motor slowing was not a prominent feature among WWH but rather verbal learning and memory (13).

Among the seronegative group, Profile1-UN was more likely to have diabetes, raising the possibility that their specific visual and motor deficits could be related to physical complications of diabetes, including diabetic retinopathy and neuropathy (33, 34). Profile2-UN reflects unique impairments on the most verbally mediated tasks (i.e., verbal learning and recall, and verbal fluencies). While we are unaware of any specific disease process or syndrome that shows the same pattern, this group of individuals has clear weakness in verbal skills, which could be due to many factors, including learning differences or damage to brain regions associated with verbally mediated tasks. Profile 3-UN, reflecting specific motor slowing, is commonly observed in individuals with basal ganglia dysfunction, such as Parkinson's disease (35). Profile 5-UN, revealing rather generalized cognitive weaknesses or impairments, but relatively preserved attention and visual processing, does not reflect any specific disease process or syndrome to our knowledge.

Even though the HIV group was virally suppressed, the dominant profiles did not fully align with the HIV-uninfected women, suggesting that HIV affects cognitive function even in the era of effective ART. There is a wealth of literature postulating neuronal damage as a result of ART agents (36), a viral reservoir that persists possibly due to poor CNS penetration of ART (37), or even legacy effects of damage occurring earlier during infection (38). Indeed, in the VS-WWH RF model where all impaired groups were grouped together, nadir CD4 was a top predictor of group membership. This also points to how existing studies that consider impairment to be a unidimensional construct may only be able to detect differences in these variables and miss those associated more strongly with some profiles than others.

Despite different cognitive profiles among VS-WWH, the most discriminative factors between each impaired profile vs. the unimpaired profile were similar and included a number of well-established sociodemographic cognitive correlates, such as years of education, age, and race/ethnicity (39). Clinic site location also emerged, a factor that we have also seen using more standard statistical approaches in the WIHS (13, 14). The factors underlying this rather robust association is unknown but may involve neighborhood factors, such as violence and food insecurity. Additionally, common behavioral correlates of cognition emerged including illicit substance use (40, 41), which in the case of marijuana was more likely to be used in the unimpaired profile compared to the impaired profile demonstrating weaknesses in learning, processing speed, and executive function (Profile 5-VS). This finding is consistent with some studies demonstrating the protective effects of marijuana use on cognition in PWH (42). We also found common clinical correlates of cognition that distinguished cognitive profiles among VS-WWH including BMI (43, 44) and PI use (45). Likely proxies of HIV disease burden, including nadir CD4 count and years of ART use, were also discriminators (38, 46, 47). In contrast, sociodemographic and medical variables were unable to distinguish cognitive profiles based on seven major cognitive domains (48).

Mental health factors also emerged as important profile discriminators among VS-WWH, including depressive and stress-related symptoms. Depressive symptoms differentiated a number of impaired profiles (4 of 5 profiles) compared to the unimpaired profile, whereas stress-related symptoms only emerged for two profiles including Profile 1-VS (sequencing) and Profile 4-VS (learning and memory). These findings align with our WIHS studies demonstrating numerous cognitive correlates of depressive symptoms (19, 49, 50), whereas stress-related symptoms related most strongly to learning and memory in the context of HIV (49, 51). Importantly, mental health factors are an unmet medical need and are modifiable targets to improve cognition in WWH (52).

INSTI use discriminated both Profile3-VS (learning and recognition) and Profile 4-VS (learning and memory) from the unimpaired profile. This finding is consistent with a number of recently published studies indicating INSTI use as a contributor to NP function. One study demonstrated an association between INSTI use and poorer learning and memory but not any other cognitive domains (53). A second study also demonstrated that switching or starting an INSTI was primarily associated with poorer learning among WWH (54). A third study demonstrated that long-term INSTI exposure distinguished two impaired profiles from an unimpaired profile (55).

Our study also allows us to investigate female-specific factors that are often ignored and identify the importance of oophorectomy and/or hysterectomy (Profile2-UN[Learning, Recall, Verbal Fluency], and Profile3-UN[Motor Speed]) and menopause status (Profile4-UN[Learning and Memory]). Interestingly, these female-specific factors only emerged as important profile discriminators among HIV-uninfected women. As the proportion of menopause-inducing and non-inducing oophorectomy and/or hysterectomy was similar across VS-WWH and HIV-uninfected women, one possible explanation is that the virus itself and clinical factors, such as ART may explain more of the variance in cognitive function in VS-WWH. However, in the absence of HIV, negative effects of oophorectomy and/or hysterectomy on cognition may become more apparent. Overall, these female-specific factors are potential contributions that are missed in other studies, which are predominately male. Future studies of women should evaluate these variables in a similar stratified form to identify potential mechanistic contributions.

The existence of distinctive patterns of cognitive performance, as well as distinct factors associated with those patterns, also adds to existing evidence of differing neuropathological mechanisms. The dominant profiles often contained patterns of weaknesses that were subclinical, yet still lower than the unimpaired profiles. In many cases, the associated factors are intervenable and should be followed up with mechanistic and longitudinal studies.

Differences between the profiles identified here and previous efforts to identify cognitive patterns can be attributed to both the methods used and the study population. To identify meaningful cognitive patterns, we used a combination of SOM and MClust, which is a slight deviation from tradition k-means clustering. The nature of k-means is that it yields clusters where the most dramatic differences are shown, which may ignore subtle differences in patterns. Even Molsberry et al. (7) and Amusan et al. (55) who used latent profile analysis using domain T-scores had their fit dominated by a high-performing and a low-performing group. Using SOM for dimension reduction on the T-scores for individual tests prevented us from following pre-conceived notions about the latent structures of cognitive domains, which have been shown to be different in HIV (56). Another reason that we may have found different profiles than prior studies is that we focused on a diverse sample of underserved lower-income, African-American and Hispanic WWH where social correlates of health are common (e.g., low educational attainment, poverty, food insecurity, etc.) (8, 57) and may lead to more heterogeneous patterns of cognitive function. Of importance, this demographic is a more accurate reflection of the HIV epidemic as opposed to the predominantly White populations evaluated in other cohorts.

The addition of machine learning models to traditional univariate statistics to identify dominating predictor variables is another distinguishing aspect of the current study. It is important to point out that RF modeling is a non-linear model, and that the variable importance measure does not take into account directionality. Therefore, it is possible to have a top predictor variable from RF that does not have P < 0.05 using a t-test. RF models are also multivariate; instead of, the predictive capabilities of variables are always observed within the context of other variables. This is important considering that none of these factors exist in isolation. This makes the model more powerful, but one limitation of this statistical approach is that it becomes more difficult to interpret and should be used as a springboard for more mechanistic studies and interventions, which is why machine learning models are often thought of as “hypothesis-generating” models.

In conclusion, in the largest sample of women to date in the United States, we have used a novel pipeline of machine learning methods to identify subgroups of patterns in NP performance and created predictive models to identify the factors that distinguish each pattern from an overall “unimpaired” group. We identified distinct patterns of cognitive weaknesses in VS-WWH that differed from the distinct patterns in HIV-uninfected women. We also identified factors that may contribute to these specific profiles as a springboard for mechanistic or interventional studies. Future studies should also investigate the stability of these profiles over time, and identify the ones, if any, that are prone to future decline.