Metabolic Risk Factors as Differential Predictors of Profiles of Neurocognitive Impairment Among Older HIV+ and HIV− Adults: An Observational Study

Abstract

Objective

Neurocognitive performance among older persons, including those living with HIV (people living with HIV [PLWH]), exhibits significant heterogeneity, suggesting subpopulations with differing profiles of neurocognitive impairment (NCI). Metabolic factors are associated with NCI, but their relationships to cluster-derived NCI profiles are unknown.

Method

Participants (144 PLWH and 102 HIV uninfected) aged 50+ years completed a neuropsychological battery assessing seven cognitive domains. Latent class analysis (LCA) identified NCI profiles separately by HIV serostatus and in a combined sample. Obtained classes were examined against the Montreal Cognitive Assessment (MoCA) and diagnoses of HIV-associated neurocognitive disorders (HAND). Multinomial regression identified metabolic predictors of classification.

Results

LCA identified three latent classes in each participant sample: Class1_{Multidomain NCI} (high probability of impairment across multiple domains), Class 2_{Learning & Recall NCI} (high probability of impairment in learning and recall), and Class 3_{NC Unimpaired} (low probability of NCI across all domains). Severity of NCI implied by classes corresponded with MoCA scores and HAND diagnoses. In analyses on the combined sample, compared to HIV-uninfected individuals, PLWH were more likely to be in Class1_{Multidomain NCI}. Among PLWH, those with dyslipidemia and hypertension had greater odds of classification in Class 1_{Multidomain NCI} while those with central obesity had higher odds of classification in Class 2_{Learning & Recall NCI}; metabolic syndrome approached significance as a differential predictor. Regardless of HIV status, individuals with diabetes were more likely to be in Class 1_{Multidomain NCI}.

Conclusions

Metabolic risk factors confer heightened risk of NCI in HIV infection. Interventions to reduce metabolic risk may improve neurocognitive outcomes among PLWH.

Introduction

Despite the advent of combination antiretroviral therapy (cART), neurocognitive impairment (NCI) remains common among people living with HIV (PLWH). NCI among PLWH, also termed HIV-associated neurocognitive disorder (HAND), continues to be observed in up to half of PLWH in the cART era (Blackstone et al., 2012; Heaton et al., 2010). The neurocognitive profile of HAND reflects considerable interindividual variability, with a wide range of performance in terms of severity and patterns of domain-specific cognitive dysfunction (Butters et al., 1990; Sacktor & Robertson, 2014; Woods, Moore, Weber, & Grant, 2009). Although impairments in learning and executive functioning tend to be most prevalent in the cART era, many PLWH are also impaired in delayed recall, motor functioning, speed of information processing, and verbal fluency domains (Heaton et al., 2010). Thus, NCI among PLWH exhibits significant heterogeneity, and this variability in cognitive performance suggests that subpopulations of PLWH may differ in risk factors and mechanisms contributing to NCI.

NCI among PLWH may be mediated by a variety of neuropathological processes including chronic inflammation, incomplete HIV suppression in the central nervous system, neurotoxicity of antiretroviral drugs, and legacy effects following severe immunosuppression (i.e., nadir CD4 levels, <200) (Bryant et al., 2016; Gannon, Khan, & Kolson, 2011; Heaton et al., 2011; Letendre, 2011). Additionally, with significant reductions in morbidity and mortality on cART, PLWH have improved longevity and are now subject to experiencing age-related declines in cognition (The Antiretroviral Therapy Cohort Collaboration, 2008; Samji et al., 2013). For example, among HIV-uninfected adults, normative developmental changes to cognition in late life typically entail declining fluid cognitive abilities in the context of stable crystallized abilities. In particular, fluid abilities, which include processing speed, problem solving, and reasoning, are independent of learned knowledge and decline by about −0.02 SD per year. In contrast, crystallized abilities, or learned skills and general knowledge that are acquired over a lifetime, remain intact or improve by 0.003–0.02 SD per year as age progresses (Salthouse, 2012). Furthermore, with the accumulation of disease burden over the lifespan, a substantial proportion of older adults subsequently develop age-related declines in cognition that may be due to a wide variety of neuropathological processes, including neurodegenerative (such as Alzheimer’s and Parkinson’s diseases) and cerebrovascular diseases (Gorelick et al., 2011; Harada, Natelson Love, & Triebel, 2013; van Hooren et al., 2007), each of which are associated with a different neurocognitive profile. Thus, HIV infection and age may have contributing and synergistic effects on the cognitive functioning of older PLWH (Valcour, Shikuma, Watters, & Sacktor, 2004).

A number of studies have implicated metabolic syndrome (MetS), a constellation of related cardiometabolic conditions (i.e., central obesity, hyperglycemia, dyslipidemia, and hypertension), and its constituent risk factors with worse neurocognitive performance. In particular, MetS has been identified as a predictor of accelerated cognitive decline, with increased risk of progression of cognitive impairment and higher incidences of dementia among individuals with diagnoses of the conglomeration of metabolic conditions (Atti et al., 2019; Yaffe et al., 2004). Robust associations have also been made between the individual components of MetS and cognitive impairment, with studies linking them to mild cognitive impairment, Alzheimer’s disease, and vascular dementia (Beydoun, Beydoun, & Wang, 2008; Kanaya, Barrett-Connor, Gildengorin, & Yaffe, 2004; Knopman et al., 2001; Reitz, Tang, Manly, Mayeux, & Luchsinger, 2007; Whitmer et al., 2008). Similar to the effects among individuals of the general population, there is evidence to suggest that cardiometabolic risk factors are associated with worse neurocognitive performance in PLWH (Gorelick et al., 2011; Gustafson et al., 2013; McCutchan et al., 2012; Valcour et al., 2005; Wright et al., 2010; Yaffe et al., 2004). However, older PLWH are at greater risk for developing metabolic conditions, compared to age- and sex-matched controls (De Wit et al., 2008; Samaras et al., 2007; Vance et al., 2014). Thus, MetS and its constituent conditions may promote neuronal injury and could account for the higher risk of neurocognitive deficits among older PLWH (Canizares, Cherner, & Ellis, 2014). This notion of MetS-associated neurocognitive vulnerability among PLWH is supported by studies that have demonstrated that components of MetS heighten the risk for HAND and correlate with imaging markers of neurochemical abnormalities and neuroinflammation among older PLWH (Cysique et al., 2013; McCutchan et al., 2012; Sattler et al., 2015; Saylor et al., 2016; Valcour et al., 2005; Yu et al., 2019).

Given variability in neurocognitive performance among PLWH and older adults, as well as individual differences in the presence of comorbidities that may impact neurocognitive performance, identifying neurocognitive profiles and their predictors is an important area of research. One approach that has been used to identify subgroups of individuals with similar patterns of neurocognitive performance is latent class analysis (LCA). LCA is an analytic method allowing the characterization of categorical unobserved (latent) variables from analyses of the structure of relationships among several observed variables. LCA has been utilized as an exploratory technique to uncover latent classes associated with Alzheimer’s and Parkinson’s diseases (Brennan et al., 2017; Hagenaars & McCutcheon, 2002; Libon et al., 2014); however, little prior research has been conducted to compare the neurocognitive profiles among older PLWH and HIV-uninfected adults. Additionally, although metabolic factors have been associated with NCI among PLWH and HIV-uninfected individuals, the relationships of MetS and metabolic risk factors to cluster-derived profiles of NCI have not been examined. Thus, this project aimed to (a) separately determine and compare profiles of NCI among PLWH and HIV-uninfected older adults using LCA, (b) determine the effect of HIV infection on NCI profile classification in a combined sample of all participants, and (c) examine the synergistic effect of HIV infection and metabolic conditions on NCI profile classification. We hypothesized that LCA would identify subgroups of PLWH and HIV-uninfected individuals that differ in their profiles of NCI and that a diagnosis of HIV infection would be associated with higher odds of classification in classes characterized by cognitive impairment. We additionally hypothesized that the presence of individual metabolic risk factors and the diagnosis of MetS would have greater negative effects on cognition among PLWH. Specifically, we hypothesized that among PLWH, metabolic risk factors and MetS would be associated with higher odds of classification in subgroups characterized by more extensive NCI.

Methods

Participants

Data from 246 participants (144 PLWH and 102 HIV uninfected) were obtained from two California studies of HIV and aging; all participant visits occurred between December 2010 and February 2016. Study eligibility criteria were similar for both PLWH and HIV-uninfected individuals. Specifically, to enroll in these studies, participants had to be English-speaking and possessing the capacity to provide informed consent; an additional inclusion criterion for our analyses was age 50 years or older. Candidates were excluded from study participation if they had a history of conditions that might confound analyses, such as head injury with loss of consciousness for longer than 30 min, neurological disease with neurological or neuropsychiatric sequelae (e.g., stroke and seizure disorders), psychotic disorder, non-HIV neurodegenerative diseases (e.g., self-disclosed non-HIV dementias, as diagnosed by the participant’s healthcare provider), or a severe learning disability (e.g., Wide Range Achievement Test-Fourth Edition [WRAT4] Word Reading score < 70).

Measures

Neurocognitive functioning

Participants completed a comprehensive neurocognitive test battery covering seven ability domains (verbal fluency, executive functioning, speed of information processing, verbal and visual learning, verbal and visual delayed recall/memory, working memory, and motor skills; Table 1). Using published normative data, raw scores were converted to T-scores demographically corrected for the effects of age, education, sex, and race/ethnicity (Heaton, Miller, et al., 2004; Norman et al., 2011). Consistent with established procedures, individual T-scores were then converted to deficit scores ranging from 0 (no impairment) to 5 (severe impairment) and averaged within domains to create domain deficit scores (DDS). Previous studies have shown a balance of sensitivity and specificity in identifying NCI using DDS cutoffs of >0.5 (Heaton, Miller, et al., 2004); thus, participants were classified with NCI versus no NCI in each of the seven domains using this DDS threshold. Additionally, deficit scores were averaged across all tests in the battery to create a global deficit score (GDS) for each individual, which was similarly dichotomized into ratings of impairment (GDS ≥ 0.5) versus no impairment (GDS < 0.5) (Carey et al., 2004; Heaton et al., 2004).

Table 1.

Measures within the comprehensive neuropsychological test battery

Cognitive domain	Measures
Verbal fluency	Controlled Oral Word Association Test
	Category Fluency (Animals)
	Category Fluency (Actions)
Executive functioning	Wisconsin Card Sorting Test (64-item)
	Trail Making Test, Part B
	Stroop Color Word Trial
Speed of information processing	WAIS-III Digit Symbol
	WAIS-III Symbol Search
	Trail Making Test, Part A
	Stroop Color Trial
Learning and memory (two domains—learning and recall)	Hopkins Verbal Learning Test-Revised
	Brief Visuospatial Memory Test-Revised
Working memory/attention	WAIS-III Letter-Number Sequencing
	PASAT (1st channel only)
Motor	Grooved Pegboard Test (dominant and non-dominant hands)

Note: WAIS-III = Wechsler Adult Intelligence Scale 3rd Edition; PASAT = Paced Auditory Serial Addition Task.

Metabolic comorbidities

Metabolic risk factors (i.e., dyslipidemia [elevated triglycerides and/or low HDL], central obesity [i.e., large waist circumference], diabetes, and hypertension) were determined from a combination of self-report (e.g., history of diabetes and/or hypertension) and objective laboratory (i.e., phlebotomy and anthropomorphic) assessments. Metabolic risk factors examined were consistent with the National Cholesterol Education Program Adult Treatment Panel III (NCEP ATP III) criteria for metabolic risk (Grundy et al., 2005; Huang, 2009). The NCEP ATP III panel defines metabolic risk as (a) large waist circumference (>102 cm [40 in] for men, 88 cm [35 in] for women), (b) elevated triglycerides (≥150 mg/dl) or prescription, (c) low HDL cholesterol (<40 mg/dl in men, < 50 mg/dl in women) or prescription, (d) hypertension (>130 mmHg systolic or > 85 mmHg diastolic) or prescription, and (e) elevated fasting glucose (≥100 mg/dl) or prescription. Thus, if participants reported receiving a prior medical diagnosis, reported prescription of drug treatment for a condition, or if they had laboratory measurements for any of the conditions at or above NCEP ATP III thresholds, they were considered positive for that metabolic risk factor. Finally, to obtain a diagnosis of MetS, indicators of MetS were summed into a composite variable, which ranged from 0 to 5, and an NCEP ATP III diagnosis of MetS was applied if they had 3 or more of the factors.

Alternative measures of cognitive functioning

MoCA

In addition to completing the standard neuropsychological battery, participants were also assessed using the Montreal Cognitive Assessment (MoCA), a well-validated neurocognitive screener for clinical geriatric populations (Nasreddine et al., 2005).

Diagnoses of HAND

Among PLWH, determinations of the presence and categories of HAND were made following Frascati criteria (Antinori et al., 2007). Specifically, PLWH were assigned diagnoses of asymptomatic neurocognitive impairment (ANI), mild neurocognitive disorder (MND), and HIV-associated dementia (HAD) based on the degree of neurocognitive and functional impairment. Determinations of functional decline were made using at least two types of evidence. Specifically, a combination of the Patient’s Assessment of Own Functioning Inventory, the modified Lawton and Brody Instrumental Activities of Daily Living scale, and (among individuals currently employed) an employment questionnaire (asking about any decreases in work productivity, accuracy/quality of work, increased effort required to do one’s usual job, and increased fatigue in association with the usual workload) was used to make determinations of declines in everyday functioning (Chelune, Heaton, & Lehman, 1986; Heaton et al., 2010; Heaton, Marcotte, et al., 2004).

Potential Covariates

Sociodemographic variables that differed across PLWH and HIV-uninfected participants, as well as the following variables that have previously been associated with neurocognitive performance or MetS, were considered as potential covariates affecting latent classification:

Premorbid functioning

All participants completed the reading subtest of the WRAT4 (Wilkinson & Robertson, 2006). Single-word reading tests, such as the WRAT4 Word Reading test, are common performance-based tools for estimating premorbid verbal IQ, and the WRAT4 Word Reading has shown strong test–retest reliability as a measure of premorbid functioning (Franzen, Burgess, & Smith-Seemiller, 1997; Rabin, Paolillo, & Barr, 2016). The WRAT4 Word Reading test has been well validated as a measure of premorbid functioning among PLWH (Casaletto et al., 2014).

Lifetime depression

The presence or absence of lifetime major depressive disorder was evaluated with the Composite International Diagnostic Interview (CIDI v.2.1), using diagnostic criteria based on the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (APA, 2000).

Lifetime substance use disorder

The presence or absence of a lifetime substance use disorder (i.e., lifetime substance abuse and/or lifetime substance dependence) was also evaluated CIDI v.2.1 using criteria based on the DSM-IV (Association).

Hepatitis-C infection

Diagnoses of hepatitis-C virus (HCV) infection were determined across all participants using enzyme-linked immunosorbent assay.

History of smoking

Lifetime history of smoking tobacco was obtained by participant self-report.

HIV disease and treatment characteristics

HIV-related disease and treatment characteristics were obtained to further characterize the sample of PLWH. In particular, we gathered information about their (a) estimated duration of HIV disease, (b) duration on cART, (c) nadir CD4+ T-cell counts, (d) current CD4+ T-cell counts, and (e) HIV viral load. With the exception of current CD4 count and viral load, all HIV disease characteristics were obtained from participants by self-report. For laboratory-based measures (current CD4 count and HIV viral load), blood samples were collected via phlebotomy and clinical assays were performed by a Clinical Laboratory Improvement Amendments-certified laboratory. Current CD4+ T-cell count was measured by flow cytometry, and HIV RNA levels in blood were measured by reverse transcriptase-polymerase chain reaction with a lower limit of quantitation (LLQ) at 50 viral copies/ml.

Data Analysis

LCA was used to estimate the number of classes best-capturing subgroups of individuals with similar patterns of DDS across the seven domains of neurocognitive performance. LCA is a statistical method for identifying underlying class membership when true class membership is unknown by inference from categorically measured variables (Hagenaars & McCutcheon, 2002). In order to determine the number of categorical latent classes underlying the observed DDS across all measured domains of cognitive functioning, several models with increasing numbers of latent classes were estimated separately by HIV status. With the goal of maximizing statistical power, another LCA was completed on the entire sample, due to the high degree of correspondence between LCA results among HIV status groups. For each participant sample that was examined, relative indices of model fit were utilized to select the model with the number of classes best fitting the data. In particular, the best-fitting model was selected based on four indices: the Bayesian information criteria (BIC), Akaike information criteria (AIC), entropy, and the bootstrapped likelihood ratio test (BLRT). For each model with k number of classes, the best log-likelihood was replicated in order to avoid convergence at a local maximum. Additionally, in log-likelihood ratio tests to examine model fit, the log-likelihood for the null model with k-1 classes was verified as being equal to the best log-likelihood value of the previously tested model with one fewer class (Berlin, Williams, & Parra, 2014).

Once the optimum number of classes was identified on the combined participant sample, the final LCA model identified was adjusted for relevant covariates (i.e., any of the previously-described potential covariates that were associated with latent classification at p < .10). Convergent validity of these covariate-adjusted classes was examined by comparisons of the means of scores on the MoCA across each of the identified latent classes using Wald χ²–tests; corrections for Type I errors in multiple comparisons were made using the Benjamini–Hochberg procedure (Benjamini & Hochberg, 1995). Among PLWH, convergent validity of latent classes was examined using a two-sided rank-based nonparametric test for ordered alternatives (Jonckheere–Terpstra) to determine changes in the severity of HAND diagnoses by latent classification. Finally, the extent to which HIV infection moderated the association between metabolic disease characteristics and group membership in the identified latent classes was determined using multinomial regression.

In all regressions and Wald tests, a ``3-step” approach, in which class membership was assigned using observed DDS categorizations and relevant covariates prior to the inclusion in the model of outcome or predictor variables, was used in order to prevent these auxiliary factors from altering the structure of latent classes and influencing final class membership (Asparouhov & Muthén, 2014; Vermunt, 2010). All LCA analyses were conducted using Mplus version 7.4 using robust maximum likelihood estimation that is robust to missing data and non-normal distributions of outcomes (Muthén & Muthén, 1998–2015). Descriptive statistics, correlational analyses, and nonparametric tests for ordered alternatives were carried out in SPSS v. 25 (IBM Corporation, 2017).

Results

Descriptive Statistics

By the combined self-report and/or NCEP ATP III criteria, 162 (65.9%) participants had hypertension, 91 (37%) had central obesity, 173 (70.3%) had dyslipidemia, 42 (17.1%) had diabetes, and 93 (37.8%) had MetS. Of these, 46.3% of participants who were categorized as hypertensive reported prescription of antihypertensive medication, 69.0% of those categorized as having diabetes reported prescription of diabetes medication, and 48.6% of those with dyslipidemia were on lipid-lowering medication. Among PLWH, 126 were evaluated for HAND and 62.1% of these individuals met criteria for a HAND diagnosis; 56 (42.7%) met criteria for ANI, 17 (13.0%) met criteria for MND, and 6 (0.5%) met criteria for HAD. With regards to demographics, participants were on average 57.96 years old (SD = 6.10 years), with 14.43 years of education (SD = 2.55 years), and were predominantly non-Hispanic white (72.4%) and male (79.3%). Further detailed descriptive statistics of the sample and summaries of study variables, broken down by HIV status, are presented in Table 2.

Table 2.

Characteristics of Study Participants by HIV Serostatus

	PLWH (n = 144)	HIV uninfected (n = 102)	p-value
Demographic characteristics
Age; mean (SD)	57.85 (6.17)	58.10 (6.03)	.76
Years of education; mean (SD)	14.36 (2.62)	14.53 (2.46)	.61
WRAT4 Word Reading; mean (SD)	103.08 (14.71)	106.35 (14.5)	.085
MoCA; mean (SD)	24.66 (3.43)	25.67 (2.96)	.015
Male, n (%)	129 (89.6%)	66 (64.7%)	<.001
Race/Ethnicity			.54
Non-Hispanic White, n (%)	105 (72.9%)	73 (71.6%)	—
Black, n (%)	18 (12.5%)	17 (16.7%)	—
Hispanic, n (%)	12 (8.3%)	9 (8.8%)	—
Other, n (%)	9 (6.3%)	3 (2.9%)	—
Ever smoked, n (%)	62 (43.1%)	27 (26.5%)	.004
Lifetime substance use disorder, n (%)	72 (50.0%)	43 (42.2%)	.24
Lifetime major depression, n (%)	83 (57.6%)	25 (24.5%)	<.001
Hepatitis C virus-infected, n (%)	25 (17.4%)	8 (7.8%)	.024
Metabolic conditions
Diabetes, n (%)	31 (21.5%)	11 (10.8%)	.020
Hypertension, n (%)	98 (68.1%)	64 (62.7%)	.35
Dyslipidemia, n (%)	115 (79.9%)	58 (56.9%)	<.001
Central obesity, n (%)	40 (27.8%)	51 (50.0%)	<.001
Metabolic syndrome, n (%)	59 (40.1%)	34 (33.3%)	.22
Cognitive impairment (domain deficits)
Verbal, n (%)	42 (29.2%)	20 (19.6%)	.090
Executive functioning, n (%)	48 (33.3%)	21 (20.6%)	.025
Speed of information processing, n (%)	47 (32.6%)	15 (14.7%)	.001
Learning, n (%)	77 (53.5%)	38 (37.3%)	.014
Recall, n (%)	73 (50.7%)	40 (39.2%)	.086
Working memory, n (%)	34 (23.6%)	16 (15.7%)	.119
Motor, n (%)	39 (27.1%)	14 (13.7%)	.011
Global, n (%)	66 (45.8%)	30 (29.4%)	.008
HIV disease characteristics
HIV-associated neurocognitive disorder, n (%)
Normal	52 (36.1%)	—	—
Asymptomatic neurocognitive disorder	56 (38.9%)	—	—
Mild neurocognitive disorder	17 (11.8%)	—	—
HIV-associated dementia	6 (4.2%)	—	—
Not assessed	13 (9.0%)
Undetectable viral load, n (%)	130 (90.3%)	—	—
Estimated duration of HIV (years), median (IQR)	19.2 (13.1, 25.8)	—	—
On cART, n (%)	138 (97.2%)	—	—
Months on current cART regimen, median (IQR)	37.6 (11.1, 72.3)	—	—
Months on any cART regimen, median (IQR)	141.3 (77.0, 207.5)	—	—
Nadir CD4+ T-cells, median (IQR)	145 (39, 300)	—	—
Current CD4+ T-cells, median (IQR)	602 (414, 785)	—	—

Note: IQR = interquartile range; WRAT = Wide Range Achievement Test; MoCA = Montreal Cognitive Assessment Test. Nadir CD4 count is the lowest of self-reported or laboratory-obtained values.

Identification of Number of Latent Classes

Table 3 reports numbers of individuals assigned to latent classes and the indices of fit for LCA models that were tested separately among PLWH and HIV-uninfected participants. Through comparisons of indices of fit of LCA models with increasing number of classes, which ranged from 1 to 4, a 3-class model was found to best fit the observed DDS profiles in both the PLWH (BIC = 1057.50, entropy = 0.91, and BLRT = 56.67, p < .001) and HIV-uninfected samples (BIC = 656.40, entropy = 0.92, and BLRT = 20.84, p < .001). Panel A of Fig. 1 depicts the DDS profiles of the three classes among PLWH while Panel B depicts the profiles among HIV-uninfected individuals. As can be observed from the pattern of DDS categories across domains, the three classes in both the PLWH and HIV-uninfected samples can generally be described as Class 1_{Multidomain NCI} (high probability of impairment across multiple domains), Class 2_{Learning & Recall NCI} (high probability of impairment in learning and recall), and Class 3_{NC Unimpaired} (having low probability of NCI across all domains).

Table 3.

Results of LCAs among PLWH and HIV-uninfected adults

Number of classes	Log-likelihood (No. of free parameters)	AIC	BIC	Entropy	Number (%) per class	BLRT
PLWH
1	−631.61 (7)	1277.22	1275.86	1.00	144 (100%)	N/A
2	−536.321 (15)	1102.65	1099.73	0.83	47 (32.6%)	190.57,
					97 (67.4%)	p < .001
3	−507.989 (23)	1061.98	1057.53	0.91	30 (20.8%)	56.67,
					52 (36.1%)	p < .001
					62 (43.1%)
4	−500.37 (31)	1062.75	1056.72	0.92	5 (3.5%)	15.23,
					28 (19.4%)	p = .29
					50 (34.7%)
					61 (42.4%)
HIV uninfected
1	−364.22 (7)	742.43	738.69	1.00	120 (100%)	N/A
2	−321.76 (15)	673.16	665.51	0.83	37 (36.3%)	84.91,
					65 (63.7%)	p < .001
3	−311.34 (23)	668.67	656.40	0.92	15 (14.7%)	20.84,
					27 (26.5%)	p < .001
					60 (58.8%)
4	−303.24 (31)	668.48	651.94	0.91	6 (5.8%)	16.19,
					9 (8.8%)	p =.17
					32 (32.4%)
					55 (53.9%)

Note: AIC = Akaike Information Criterion; BIC = Bayesian Information Criterion; LMR-LT = Lo Mendell Rubin Likelihood Ratio Test; BLRT = Bootstrapped Likelihood Ratio Test.

Fig. 1.

Panels depicting the 3-class LCAs of cognitive impairment among PLWH, HIV-uninfected participants, and in a combined sample.

A series of Wald chi-square tests (with adjustments for multiple comparisons) were carried out to examine differences in the estimated probabilities of impairment across all domains of cognitive performance within the three similar pairs of classes by HIV status. Comparing across all domains for Class 1_{Multidomain NCI}, the estimated probability of impairment in speed of information processing was found to be significantly greater among PLWH than among HIV-uninfected individuals (Wald χ²[1] = 155.14, adjusted p < .001). Among individuals classified as Class 2_{Learning & Recall NCI}, learning and working memory were more likely to be impaired among PLWH (Wald χ²[1] = 155.11, p < .001 and Wald χ²[1] = 164.72, p < .001, respectively). Additionally, despite the overall low probabilities of impairment, among individuals in Class 3_{NC Unimpaired}, delayed recall was more likely to be impaired among PLWH (Wald χ²[1] = 137.59, p < .001).

As the number of latent classes was similar and the overall patterns of impairment were analogous for PLWH and HIV-uninfected individuals, both samples were combined and a new best-fitting LCA model that accounted for HIV (through the regression of class on HIV serostatus) was estimated. Consistent with previous results, a 3-class model in the combined sample (with HIV serostatus as a covariate) was found to best describe the DDS categories across the seven neurocognitive domains (BIC = 1714.99, entropy = 0.90, and BLRT = 83.78, p < .001).

Multinomial regression analyses were utilized to identify other potential covariates for inclusion as variables affecting class formation (Table 4). WRAT4 Word Reading level and HCV infection were identified as being both associated with class membership in the HIV-adjusted LCA (p < .10). Specifically, individuals with higher WRAT4 Word Reading scores were marginally less likely to be classified as Class 1_{Multidomain NCI} (OR = 0.97, p = .093) while individuals with diagnoses of HCV were marginally more likely to be classified as Class2_{Learning&Recall NCI} (OR = 2.13, p = .081), relative to Class 3_{NC Unimpaired}. Thus, WRAT4 and HCV infection were subsequently included as covariates affecting class formation. Table 5 reports fit indices of the series of LCAs on the combined sample for models that adjusted for HIV infection only and for the final 3-class LCA model that adjusted for HIV infection, WRAT4 Word Reading, and HCV infection (BIC = 1777.90, entropy = 0.93, and BLRT = 89.88, p < .001).

Table 4.

Regression estimates obtained from multinomial models examining potential covariates of classification

Outcome: classification in Class 1Multidomain NCI					Outcome: classification in Class 2Learning & Recall NCI
Predictor	Estimate	SE	Odds ratio	p-value	Predictor	Estimate	SE	Odds ratio	p-value
WRAT4	−0.04	0.02	0.97	.093	WRAT4	−0.01	0.02	0.99	.55
HIV status	1.41	0.51	4.11	.005	HIV status	0.42	.32	1.53	.18
Lifetime depression	0.41	0.46	1.51	.37	Lifetime depression	−0.16	0.35	0.85	.64
HIV status	1.36	0.52	3.89	.009	HIV status	0.42	0.36	1.52	.24
Lifetime substance use disorder	0.29	0.45	1.33	.52	Lifetime substance use disorder	0.32	.37	1.37	.39
HIV status	1.40	0.50	4.08	.005	HIV status	0.36	.33	1.44	.27
Hep C infection	−0.45	0.78	0.64	.56	Hep C infection	0.758	.434	2.13	.081
HIV status	1.60	0.49	4.96	.001	HIV status	0.45	0.32	1.57	.15
History of smoking	−0.10	0.44	0.90	.82	History of smoking	−0.43	0.41	0.66	.30
HIV status	1.55	0.51	4.73	.002	HIV status	0.57	0.34	1.77	.092
Gender	−0.91	0.78	0.41	.25	Gender	−0.02	0.42	0.98	.96
HIV status	1.24	0.65	3.44	.057	HIV status	0.41	0.34	1.50	.22
Age	0.05	0.03	1.05	.15	Age	0.02	0.03	1.02	.52
HIV status	1.54	0.54	4.66	.004	HIV status	0.42	0.33	1.52	.20
Race/ethnicity	−0.05	0.46	0.95	.92	Race/ethnicity	0.06	0.47	1.56	.91
HIV status	1.48	0.50	4.41	.003	HIV status	0.42	0.32	1.53	.92

Note: Provided estimates are relative to classification in Class 3_{NC Unimpaired}.

Table 5.

Results for LCAs in the combined sample (PLWH and HIV-uninfected individuals), adjusted for relevant covariates

	Log-likelihood (no. of free parameters)	AIC	BIC	Entropy	Number (%) per class	BLRT
HIV adjusted
1	−1013.80 (7)	2041.59	2043.94	1.00	246 (100%)	N/A
2	−870.19 (16)	1772.39	1777.75	0.79	82 (33.3%)	287.20,
					164 (66.7%)	p < .001
3	−828.30 (25)	1706.61	1714.99	0.90	39 (15.9%)	83.78,
					74 (30.0%)	p < .001
					133 (54.1%)
4	−823.26 (34)	1714.52	1725.93	0.87	31 (12.6%)	10.09,
					36 (14.6%)	p > 0.99
					53 (21.5%)
					126 (51.2%)
HIV, HCV, and WRAT4 Word Reading-adjusteda
3	−809.418 (29)	1676.84	1777.90	0.93	39 (16.2%)	89.88,
					83 (34.4%)	p < .001
					119 (49.4%)

Note: AIC = Akaike Information Criterion; BIC = Bayesian Information Criterion; LMR-LT = Lo Mendell Rubin Likelihood Ratio Test; BLRT = Bootstrapped Likelihood Ratio Test.

^{a a}Due to missing data on a covariate of classification (HCV diagnosis), the numbers of individuals included in combined analyses decreased to 241.

Panel C of Fig. 1 depicts the profiles of NCI of the final model (adjusted for HIV serostatus, HCV infection, and WRAT4 Word Reading) estimated on the combined sample. The interpretation of latent classes remained that of an impaired class across multiple domains (Class 1_{Multidomain NCI}; N_Total = 39, N_PLWH = 31, N_{HIV-uninfected} = 8), an impaired class in learning and recall (Class 2_{Learning & Recall NCI}; N_Total= 83, N_PLWH = 32, N_{HIV-uninfected} = 51), and an NC unimpaired class (Class 3_{NC Unimpaired}; N_Total = 119, N_PLWH = 57, N_{HIV-uninfected} = 62).

Examinations of the frequencies of global NCI (GDS ≥ 0.5) across the three latent classes revealed that all participants in Class 1_{Multidomain NCI} and 60.2% of participants in Class 2_{Learning & Recall NCI} were also categorized as globally neuropsychologically impaired. Despite low overall estimated probabilities of impairment among individuals in Class 3_{NC Unimpaired}, 5.9% of participants in Class 3_{NC Unimpaired} were classified as impaired based on GDS. For these individuals, their overall pattern of performance conformed more closely in the probabilistic LCA model to that of other individuals in Class 3_{NC Unimpaired} than to individuals in either Class 1_{Multidomain NCI} or Class 2_{Learning & Recall NCI}.

Convergent Validity of Latent Classes

Convergent validity of classification was examined by comparison of MoCA scores across the latent classes. Individuals in Class 3_{NC Unimpaired} had the highest MoCA scores (mean = 26.30, SE = 0.27), followed by those in Class 2_{Learning & Recall NCI} (mean = 24.19, SE = 0.34), and finally by individuals in Class 1_{Multidomain NCI} (mean = 22.70, SE = 0.79). The omnibus χ²-test was statistically significant (χ²[3] = 35.98, p < .001) and, relative to those in Class 3_{NC Unimpaired}, individuals in Class 2_{Learning & Recall NCI} and in Class 1_{Multidomain NCI} had significantly lower scores on the MoCA (χ²[1] = 23.37, adjusted p = .001 and χ²(1) = 18.04, adjusted p = .001, respectively). MoCA scores between individuals in Class 1_{Multidomain NCI} and in Class 2_{Learning & Recall NCI} differed only with marginally, (χ²[1] = 2.80, adjusted P = .094).

Among PLWH, a rank-based parametric analysis was carried out to examine whether the severity of HAND diagnoses differed by latent classification. A Jonckheere–Terpstra test for ordered alternatives indicated statistically significant decreases in HAND severity among individuals with progression through the ordered latent class categories (i.e., with increasing HAND severity from Class 1_{Multidomain NCI}, to Class 2_{Learning & Recall NCI}, to Class 3_{NC Unimpaired}), TJT = 2572.00, z = −6.80, p < .001. Pairwise comparisons indicated more severe diagnoses of HAND among individuals in Class 1_{Multidomain NCI} compared to Class 3_{NC Unimpaired} (TJT = 156.50, z = −6.19, adjusted p < .001) and among individuals in Class 2_{Learning & Recall NCI} compared to Class 3_{NC Unimpaired} (TJT = 504.00, z = −5.42, adjusted p < .001). Lastly, the greater degree of HAND severity among individuals in Class 1_{Multidomain NCI} compared to individuals in Class 2_{Learning & Recall NCI} approached statistical significance (TJT = 518.50, z = −1.92, adjusted p = .082).

Predictors of Classification

Table 6 reports results of the multinomial regression of latent classes on metabolic risk factors. In all analyses, the interaction between metabolic risk and HIV serostatus was included in order to examine the differential effects by HIV serostatus. When an interaction was not significant, that term was discarded in order to interpret the main effects. PLWH with dyslipidemia (OR = 4.03, p = .021) and hypertension (OR = 3.72, p = .023) had greater odds of classification in Class1_{Multidomain NCI}, relative to Class 3_{NC Unimpaired}. PLWH with MetS had marginally greater odds of classification in Class1_{Multidomain NCI}, relative to Class 3_{NC Unimpaired} (OR = 5.88, p = .082). Regarding classification as Class 2_{Learning & Recall NCI}, only central obesity increased the odds that PLWH would be classified as Class 2_{Learning & Recall NCI}, relative to Class 3_{NC Unimpaired} (OR = 2.80, p = .035). The interaction of HIV serostatus and diabetes was not significant; however, in a model of only main effects, diabetes was associated with greater odds of classification in Class1_{Multidomain NCI}, relative to Class 3_{NC Unimpaired} (OR = 2.52, p = .040).

Table 6.

Regression estimates obtained from multinomial models examining predictors of classification

Outcome: classification in Class 1Multidomain NCI					Outcome: classification in Class 2Learning & Recall NCI
Predictor	Estimate	SE	Odds ratio	p-value	Predictor	Estimate	SE	Odds ratio	p-value
Central obesity	−1.13	0.70	0.32	.11	Central obesity	−0.54	0.40	0.58	.18
Central obesity*HIV status	1.22	0.82	3.37	.14	Central obesity*HIV status	1.01	0.48	2.75	.035
HCV infection	−0.67	0.90	0.51	.46	HCV infection	0.68	0.42	1.97	.11
WRAT4	−0.04	0.02	0.96	.085	WRAT4	−0.01	0.02	0.99	.41
HIV status	1.56	0.51	4.74	.002	HIV status	0.44	0.33	1.56	.18
Dyslipidemia	−0.81	0.67	0.44	.23	Dyslipidemia	−0.66	0.41	0.51	.11
Dyslipidemia*HIV status	1.39	0.61	4.03	.021	Dyslipidemia*HIV status	0.63	0.38	1.88	.10
HCV infection	−0.67	0.90	0.51	.46	HCV infection	0.68	0.42	1.97	.11
WRAT4	−0.04	0.02	0.96	.085	WRAT4	−0.01	0.02	0.99	.41
HIV status	1.56	0.51	4.74	.002	HIV status	0.44	0.33	1.56	.18
Hypertension	−0.53	0.63	0.59	.40	Hypertension	−0.12	0.38	0.89	.75
Hypertension*HIV status	1.31	0.58	3.72	.023	Hypertension*HIV status	0.43	0.37	1.53	.25
HCV infection	−0.67	0.90	0.51	.46	HCV infection	0.68	0.42	1.97	.11
WRAT4	−0.04	0.02	0.96	.085	WRAT4	−0.01	0.02	0.99	.41
HIV status	1.56	0.51	4.74	.002	HIV status	0.44	0.33	1.56	.18
Diabetes	0.12	0.87	1.12	.89	Diabetes	−0.98	0.88	0.37	.27
Diabetes*HIV status	1.09	0.95	2.97	.25	Diabetes*HIV status	0.99	0.98	2.69	.32
HCV infection	−0.67	0.90	0.51	.46	HCV infection	0.68	0.42	1.97	.11
WRAT4	−0.04	0.02	0.96	.085	WRAT4	−0.01	0.02	0.99	.41
HIV status	1.56	0.51	4.74	.002	HIV status	0.44	0.33	1.56	.18
MetS	−0.97	0.96	0.38	.31	MetS	−0.41	0.47	0.66	.38
MetS*HIV status	1.77	1.02	5.88	.082	MetS*HIV status	0.83	0.56	2.29	.14
HCV infection	−0.67	0.90	0.51	.46	HCV infection	0.68	0.42	1.97	.11
WRAT4	−0.04	0.02	0.96	.085	WRAT4	−0.01	0.02	0.99	.41
HIV status	1.56	0.51	4.74	.002	HIV status	0.44	0.33	1.56	.18

Note: Provided estimates are relative to classification in Class 3_{NC Unimpaired.}

Post hoc analyses

Additional post hoc analyses examining the effect of HIV disease characteristics on classification were carried out to further examine effects among PLWH.

No significant relationships in the regression of the latent classes on nadir CD4, current CD4 count, estimated duration of HIV disease, detectability of HIV viral proteins above the LLQ of 50 copies/ml, or duration on cART were found. However, higher nadir CD4 was negatively associated with diagnosis of dyslipidemia (ρ = −0.18, p = .035) while longer duration on cART was associated with MetS (ρ = 0.20, p = .019) and diabetes (ρ = 0.30, p < .001).

Discussion

Multiple studies examining NCI have observed an association between aging, metabolic risk factors, and HIV infection; however, relationships between cluster-derived profiles of NCI, metabolic risk factors, and HIV serostatus have not been explored. Thus, in a sample of older adults, we aimed to characterize profiles of NCI and to examine HIV serostatus as a moderator of the link between metabolic risk factors and profiles of impairment.

We utilized LCA, a person-centered statistical classification technique of examining heterogeneity to identify distinct neuropsychological profiles of NCI among PLWH and HIV-uninfected individuals. In our sample of older adults, approximately 7.8% of HIV uninfected and 21.5% of PLWH were classified as having overall high rates of impairment across all neurocognitive domains (Class 1_{Multidomain NCI}), while 31.4% of HIV uninfected and 35.4% of PLWH were classified as primarily impaired in learning and delayed recall (Class 2_{Learning & Recall NCI}). A sizeable proportion of individuals in both groups (~60.8% of HIV-uninfected individuals and 39.6% of PLWH) were classified as neurocognitively intact (Class 3_{NC Unimpaired}). Thus, PLWH were more heavily represented in the latent class characterized by appreciable multidomain NCI while impairments to learning and recall were more common across the participant sample, regardless of HIV status.

The convergent validity of latent classes was supported by findings of the degree of NCI within latent classes tracking with worsening neurocognitive performance on an independent measure of neurocognitive performance, the MoCA. In particular, individuals in Class 1_{Multidomain NCI} and Class 2_{Learning & Recall NCI} had average MoCA scores below the clinical cutoff of 26/30 that corresponds to a preliminary diagnosis of mild NCI (Nasreddine et al., 2005). Furthermore, the average MoCA scores of individuals in Class 1_{Multidomain NCI} were lower than that of individuals in Class 2_{Learning & Recall NCI}, suggesting that the widespread pattern of NCI observed in Class 1_{Multidomain NCI} paralleled a greater degree of clinical impairment than in the other two classes. Similarly, among PLWH, the severity of NCI implied by the latent classes followed the severity of HAND denoted by diagnoses based on the Frascati criteria (Antinori et al., 2007).

Previous studies have described the domain-specific patterns of NCI in PLWH as ``spotty” (Blackstone et al., 2012). We found that three distinct patterns of NCI predominate among PLWH and that they are similar to profiles of NCI among HIV-uninfected individuals; thus, NCI phenotypes observed among PLWH may not be unique to HIV infection. Although the patterns of impairment and interpretations of classes were similar across PLWH and HIV-uninfected samples, some domains within similar pairs of classes appeared to be more vulnerable to impairment among PLWH. In particular, in Class 1_{Multidomain NCI}, impaired speed of information processing tended to be more prevalent among PLWH, while learning and working memory were more greatly affected in Class 2_{Learning & Recall NCI}, and delayed recall appeared more affected in Class 3_{NC Unimpaired}. This greater impairment among PLWH in speed of information processing, learning, recall, and working memory is consistent with preferential involvement of fronto-striatal brain regions (Woods et al., 2009).

Previous work has also used clustering approaches to identify profiles of neurocognitive performance among PLWH and HIV-uninfected individuals and identified varying numbers of classes (Brennan et al., 2017; Dawes et al., 2008; Fazeli et al., 2014; Libon et al., 2014; Lojek & Bornstein, 2005; van Gorp et al., 1993). Some of the differences in these studies may be a function of the clustering method (e.g., cluster analysis vs. LCA), sample size, the nature of indicators of neuropsychological performance, and populations at risk for NCI. Thus, the number of classes obtained in this study may differ from those previously identified in other studies. However, despite variability in the number of estimated classes underlying neurocognitive performance, clusters of NC performance similar to the three classes obtained in this study have previously been observed. Indeed, other studies have found evidence of classes of individuals with few deficits, classes with some exhibiting global NCI, and others with NCI that is primarily amnestic, suggesting a high likelihood that similar patterns of NCI exist across different clinical samples (Brennan et al., 2017; Libon et al., 2014; Lojek & Bornstein, 2005; van Gorp et al., 1993).

When metabolic predictors were examined, diabetes was found to be associated with multidomain NCI, regardless of HIV status. Additionally, PLWH with hypertension and dyslipidemia were more likely to have widespread multidomain NCI while PLWH with central obesity were more likely have cognitive impairment in the domains of learning and recall. PLWH with MetS were also marginally more likely to have global NCI. Taken together, these findings that HIV infection can amplify the detrimental effects of metabolic risk factors on cognition are consistent with those of other studies that have similarly noted the potentiation of neurocognitive insults in the context of HIV infection (Canizares et al., 2014; Vance et al., 2014). Given this, there may be a need for interventions that focus on supporting the cardiometabolic well-being of PLWH that may limit or reverse declines in cognition.

Our study has several important limitations. First, following the inclusion of HIV serostatus as a covariate in our classification, only a small number of HIV-uninfected individuals were classified in Class 1_{Multidomain NCI}. Thus, comparisons of the effects of metabolic risk factors and MetS by HIV status are among a small number of HIV-uninfected participants in this class. Additionally, although we found a significant main effect of diabetes on latent classification, suggesting no differential impact of diabetes on cognition by HIV status, our study had relatively few individuals with diabetes (compared to the numbers of individuals with other metabolic risk conditions). Thus, our results are tentative with respect to the independence of the effects of HIV and diabetes on cognition. Future studies should attempt to replicate our findings in a larger sample of individuals living with and without HIV to enable more robust determinations of the specific effects of metabolic risk factors by HIV serostatus. We also found that MetS only marginally increased the odds of global impairment among PLWH, potentially suggesting that meeting criteria for MetS was less informative about risk for global NCI in the context of HIV infection than having individual diagnoses of the component risk factors. However, because some of the individual components of MetS had differential effects on cognition (with central obesity emerging as a predictor of learning and recall impairment while other metabolic risk factors were associated with greater global impairment among PLWH), the predictive value of MetS for global cognitive impairment may have been attenuated. Thus, taken altogether, further research on larger sample sizes is warranted to examine the separate and conjoint influences of metabolic conditions on cognition.

A further limitation of our work is related to the utilization of a variety of sources of information to derive diagnoses of metabolic risk conditions (i.e., a combination of self-report of medical history, prescription of pharmacological treatment, and measured laboratory values). Although the diversity of sources allowed for a fuller characterization of our sample, their heterogeneity and various levels of objectivity introduce error in the determinations of metabolic risk conditions. Additionally, related to measurement and the chosen neurocognitive domains, we emphasized domains commonly affected in HIV. Regarding memory specifically, our battery included learning and delayed recall domains (as opposed to indices such as recognition and/or retention) because of their sensitivity to subcortical types of NCI (Salmon & Filoteo, 2007). Thus, our work may not generalize to other studies in which additional indices of memory are used.

We also examined the convergent validity of latent classification against the MoCA and found evidence to suggest that differences in the degree of severity of NCI implied by the three latent classes corresponded to differences in cognitive dysfunction captured by this standardized external measure. However, as the MoCA is a brief screener, it is less comprehensive and almost certainly less sensitive to HAND and other forms of NCI than the standard neuropsychological test battery that was utilized in LCA analyses. Given this, more comprehensive measures should be considered in future work to assess construct validity.

Furthermore, our analyses are cross-sectional; thus, we are unable to determine whether these latent classes have implications for the future course of neurocognitive health. Longitudinal studies are needed to examine if metabolic risk factors predict future cognitive impairment, including the development of dementia, and whether they may speak to the disentanglement of HAND from Alzheimer’s disease among PLWH (Milanini & Valcour, 2017). Additionally, although our participant sample ranged widely in age (50–65 years) and represented an older demographic of PLWH than many studies examining MetS have previously considered, individuals represented in this work were on average younger than those typically examined in studies of metabolic risk factors and cognitive decline in the general population. Thus, future work should also be inclusive of substantially older geriatric PLWH to more thoroughly examine the relations between cardiometabolic conditions and NCI. Lastly, future studies should examine the effects of sex and racial/ethnic differences on patterns of performance on neuropsychological tests, include measures that may be able to distinguish Alzheimer’s disease from HAND among PLWH, and combine these studies with neuroimaging and the assessment of biomarkers.

In summary, this study utilized LCA to model heterogeneity in the neurocognitive performance of older adults living with and without HIV infection and related profiles of NCI to metabolic risk factors. We found evidence of the same numbers of classes and analogous profiles of impairment, but with somewhat differing predictors of impairment, among PLWH and HIV-uninfected individuals, suggesting that pathways toward corresponding profiles of NCI may be multiply determined. PLWH with metabolic risk factors were at increased risk of NCI; thus, early intervention to reduce metabolic risk factors may improve the neurocognitive outcomes of PLWH. However, additional longitudinal observational studies, coupled with imaging and biomarkers, could help to inform and target interventional approaches.

Funding

This study was supported by the following grants: Multi-Dimensional Successful Aging Among HIV-Infected Adults (NIH/NIMH R01MH099987; PIs: Drs. Dilip V. Jeste and David J. Moore) and Successfully Aging Seniors with HIV (California HIV/AIDS Research Program IDEA Award ID10-SD-057; PI: Dr. David J. Moore). This work was further supported by salary support for Dr. Pasipanodya from the Interdisciplinary Research Fellowship in NeuroAIDS (R25MH081482). Dr. Montoya, Ms. Campbell, Ms. Hussain, and Ms. Paolillo were supported by the National Institute on Drug Abuse (T32 DA031098) while Mr. Saloner was supported by the National Institute on Alcohol Abuse and Alcoholism (T32AA013525).