Cognitive phenotyping of HIV in the era of antiretroviral therapy

Abstract

Purpose of review

The overarching goal of the current paper is to examine the empirical basis for neurocognitive phenotyping of HIV.

Recent findings

The pattern of cognitive symptoms associated with HIV has traditionally been referred to as a “subcortical” phenotype. Recent concern has been raised that the neurocognitive phenotype in the post-ART era has changed to reflect the addition of cortical features, suggestive of synergistic age-related neurodegeneration. Empirical evidence reviewed in this paper suggests that, when present, HIV-related cognitive impairment in the post-ART era remains “subcortical” in nature, regardless of advanced age or treatment status. Persistent neurocognitive impairment among virally suppressed individuals may reflect a combination of HIV disease factors, pre-existing risk factors, and/or emergent health comorbidities such as cerebrovascular disease in older people living with HIV.

Summary

An entrenchment of the “subcortical” neurocognitive phenotype of HIV appears to be unfolding in the post-ART era. Whether additional neurocognitive subtypes of HIV exist in the current era requires additional research utilizing harmonized test protocols and advanced computational methods capable of deep phenotyping. Recommendations from other neurological disorders are provided.

Introduction

HIV-associated neurocognitive disorders (HAND) are believed to affect nearly 1 in 2 persons living with HIV (PLWH) [1–3]. While HAND is not a universal manifestation of HIV, the massive scale of the HIV epidemic translates into an estimated point prevalence of 18 million PLWH who meet current research criteria for HAND [2–3]. In most cases, the symptoms are mild and do not disrupt daily living skills [3–5]. However, even mild cognitive symptoms undermine quality of life, and individuals with minor cognitive impairment are at risk of cognitive decline when co-morbid health conditions are present, despite treatment with antiretroviral therapy (ART) [6]. More profound cognitive impairment has been reported among PLWH who reside in countries where access to healthcare is suboptimal [7]. According to one estimate, HIV is the leading cause of significant brain dysfunction in young adults in resource-restricted regions of the world [8]. These trends are alarming given the absence of a cure for the disease, and a growing body of evidence that neurocognitive symptoms persist in the context of suppressive ART.

Numerous studies demonstrate that viral suppression does not fully reverse brain abnormalities in chronically infected individuals [9–10]. Additionally, initiation of ART may not be sufficient to prevent the development of new cerebral injury. Sanford et al. [11] reported neuroimaging abnormalities before ART, but no evidence of progressive injury 6 months after suppressive treatment. By contrast, preliminary results from adults with acute HIV revealed progressive brain volume loss over a two-year period despite initiation of suppressive ART during the first weeks of infection [13]. Long-term prospective studies are needed to determine whether individuals with either acute or more chronic infection are protected from de novo neurocognitive impairment. This is particularly true among PLWH on long-term suppressive ART, who are now expected to live as long as their uninfected peers. The life-saving benefits of ART introduce new health challenges including increased risk of synergistic or additive disease processes that have potential to alter the neurocognitive profile of HIV. This concern has been reported in the recent literature, noting that the neurocognitive phenotype of HIV has shifted towards a mixed subcortical-cortical pattern [14–*15]. If substantiated, the change would have high clinical significance because evidence of a cortical profile would implicate new disease mechanisms and neural substrates of neurocognitive symptoms, particularly Alzheimer’s disease (AD).

The current paper examines the empirical basis for neurocognitive phenotyping of HIV, with attention directed at the pattern of neurocognitive impairment before and after the start of the ART era. The paper concludes with recommendations to guide future studies and an example of novel research methods of deep phenotyping using unsupervised machine learning.

The subcortical cognitive phenotype of HIV

The cognitive phenotype of HIV has been described as “subcortical” since the beginning of the epidemic [16]. Independent of moderating factors (e.g., viral, host, cultural context), the cognitive symptoms in PLWH predominately include slow motor speed, bradyphrenia, reduced learning efficiency, and/or executive dysfunction, with sparing of core language and memory consolidation skills [17–20]. Impaired memory typically involves poor learning efficiency secondary to “upstream” abnormalities in attention, working memory, and/or executive function; amnestic memory dysfunction is uncommon [16–18].

The description of neurocognitive strengths and weaknesses align with the neuropathological features of HIV. Specifically, studies demonstrate high viral aggregation in the basal ganglia [21], and neuroimaging abnormalities in the white matter, caudate, and putamen in the early and later stages of HIV [22–26]. Neuroimaging abnormalities in the cortex and diffuse disruption to network connectivity have also been reported in PLWH [27], particularly in advanced or untreated disease [28]. These cortical abnormalities likely reflect remote signatures of damage to interconnected subcortical regions rather than isolated injury in the cortical mantle. Nevertheless, caution is warranted when applying “subcortical” or “cortical” phenotype labels because neither are defined by exclusive neuroanatomical models.

Cognitive signatures of HIV in the pre- and post-ART era

The pattern of neurocognitive symptoms associated with HIV before and after the introduction of ART were rigorously examined in two recent publications. In the first, Cysique et al. [29] reported more significant verbal learning impairment among PLWH in the post ART era compared to the prior period. Nearly identical results were described by Heaton et al. [30], with the additional observation that motor dysfunction appeared less common in the current era when compared to the prior period. However, neither study argued for the existence of a fundamental change to the phenotype of HAND in favor of a cortical phenotype. To the contrary, both studies emphasized that poor learning, rather than loss of retained information, contributed to the expression of worse memory. Indeed, close examination of the frequencies of delayed recall performance reveals high correspondence between the two treatment periods.

Outcomes from the studies above are noteworthy on two fronts. First, neuropsychology, perhaps more so than other disciplines, is riddled with inconsistent and nonspecific nomenclature. Memory, for example, is a complex phenomenon governed by multiple neural networks that harbor different affinities for disease processes. Evidence of impairment in one component of memory does not ensure impairment across other memory systems, or the involvement of a common neural substrate. This complexity necessitates a standardized nomenclature to facilitate interpretation across studies. Second, the neurocognitive phenotype described in both studies is highly consistent with the observations from Navia and Price [16] in the earliest days of the HIV epidemic, which included descriptions of abnormal memory function among individuals with neuropathology confined to the basal ganglia and white matter, without concomitant involvement of cortical regions.

Recommendations for future research

Unsupervised machine learning classification

Cluster analyses identify intrinsic patterns that help to organize large data matrices. These methods offer a less biased approach to explore hidden features of clinical phenotypes. A relative paucity of cluster analytic studies have examined neurocognitive symptoms of HIV, and results from the existing studies are difficult to interpret due to lack of standardized methods of neurocognitive assessment. Significant differences in test selection, nomenclature for defining cognitive domains (e.g., working memory vs. short-term memory vs. executive function), assignment of individual tests to specific domains (e.g., verbal fluency as a measure of Language vs. Executive Function), definitions of “impairment” (e.g., global deficit score, NPZ) are common [31–36]. Additionally, variability in methods to calculate global dysfunction contribute to the confusion because composite scores derived from cognitive tests that assess a restricted brain network (e.g., average scores on tests of frontal function) cannot be compared to composite scores comprised of tests that examined global brain networks. These challenges emphasize the need for a common methodology to quantify and describe neurocognitive impairment.

Specialty areas in neurology (AD, cerebrovascular disease, multiple sclerosis, etc.) and psychiatry (e.g., schizophrenia) benefit from common test batteries. By comparison, cognitive protocols in HIV studies are frequently designed and implemented on a study-by-study basis without pre-meditated efforts to support harmonization. Overcoming this challenge will require a consensus panel comprised of experts with input from multiple stakeholders to select tests based on psychometrics (reliability, validity, ceiling, floor), real-world relevance, opportunity for international application, brevity, availability of multiple forms, cost, and task-sharing opportunities (i.e., administration by non-neuropsychologists). Table 1 provides an example from the vascular cognitive impairment literature [37], which provides recommendations for longer and shorter protocols.

Table 1.

Harmonized Neurocognitive Assessment Protocols

Cognitive Domain	Vascular Harmonization	Modified for HIV
	Battery	Harmonization
60-Minute Battery
Premorbid Function	--	Wide Range Achievement Test-IV
Focused Attention	--	Digit Span Forward
		1-Back
Working Memory	Digit Span Forward	PASAT trial 1
	Digit Span Backward	LNS
Fine Motor Speed	--	Finger Tapping
		Grooved Pegboard
Psychomotor Speed	--	SDMT
		Color Trails 1
Executive Function^	Animal Fluency	Letter Fluency
	Letter Fluency	Action Fluency
	Color Trails 1 and 2	Color Trails 2
		Go/No-Go
		Stroop Interference
Visuomotor Speed	SDMT	--
Language	BNT-15	Animal Fluency
		BNT-15
Visuospatial	Rey-Osterrieth CFT Copy	--
Memory	Rey-Osterrieth CFT	CVLT–2
	HVLT-R	Alt: Rey AVLT
	Alt: CVLT-2
30-Minute Battery
Psychomotor-	Animal Fluency	Letter Fluency
Executive Function	Letter Fluency	Action Fluency
	SDMT	Color Trails 1 and 2
		Stroop
		Go/No-Go
Memory	HVLT-R	CVLT–2 learning and delay only
5-Minute Battery#
Montreal Cognitive	5-word registration, delayed	MOCA
Assessment (MOCA)	recall and recognition;	Color Trails 1 and 2
	orientation, and letter fluency

Italics represent tests that overlap with recommendations from the NIH Research Domain Criteria (RDoCS); PASAT: Paced Auditory Serial Addition Test (PASAT, trial 1 only); SDMT: Symbol Digit Modalities Test; BNT: Boston Naming Test (short form); RCFT: Rey Osterreith Complex Figure Test; HVLT-R: Hopkins Verbal Learning Test-Revised; CVLT-2: California Verbal Learning Test-2;

^{^}alternative computer-administered measures from the EXAMINER.

It is important that a harmonized test battery for HIV include domains/measures that are sensitive to diverse brain networks. While a central theme of the current paper is that available data point towards a persistent “subcortical” neurocognitive phenotype, there is high possibility that the pattern of neurocognitive symptoms will evolve in this population considering the expansive list of possible contributing factors. Recent studies identify potential causal pathways involving immune dysregulation (i.e., monocyte activation marker soluble CD163 [38–*40], increased levels of CD16 expressing monocytes [41], loss of CCR2 expressing non-classical monocytes [42], plasma and CSF levels of neopterin, MCP-1, lymphocyte markers such as IP-10 [43]), viral reservoirs, iatrogenic effects of medications, health co-morbidities (hepatitis C, cardio/cerebrovascular disease), and confounding effects of mental health symptoms. It is possible, therefore, that the neurocognitive phenotype of HIV will become more variable, and more nuanced based on the combinations of individual and interactive risk factors for each person.

Research is also needed to leverage advances in computational methods such as machine learning algorithms. An advantage of machine learning is the opportunity to explore the underlying data structure of complex clinical phenotypes by applying algorithms that iteratively learn patterns within high dimension data. The result is a much deeper level of clinical phenotyping. An example of outcomes using the deep clustering program CorEx [44] is provided in Figure 1. CorEx (correlational explanation) identifies hierarchical representations of data (i.e., phenotypes) using a mutual information criterion. When applied to neurocognitive performance, CorEx can help identify intrinsic patterns that are not evident using traditional correlational methods. An example is provided in Figure 1, which depicts two neurocognitive phenotypes of chronic HIV. Additional methods for phenotype analyses include graphical modeling [45] and Rasche analysis [46] have been applied successfully in other neurological populations (e.g., stroke), and represent important methodological tools for deep phenotyping of HIV.

Fig. 1. Deep phenotyping of neurocognitive performance in HIV.

Heat maps depicting two deep phenotypes of neurocognitive performance in HIV. Top: verbal-dominant factor; Bottom: visual-dominant factor. Performances are depicted as raw scores.

Finally, there is a significant need to enhance methods to quantify capacity to complete activities of daily living (ADLs) in PLWH. The degree to which cognitive symptoms interfere with ADLs is a critical component of current diagnostic criteria for neurocognitive impairment, including the Frascati criteria for HAND [3]. Despite the clinical relevance, accurate assessment of real-world function is challenging in both clinical and research settings. By default, the common approach in HIV involves the use of Western-based self-report scales that were developed for use in other populations, with insights provided by informants (spouses, adult children, life partners, etc.) familiar with the individual’s prior and current functional capacity.

Not surprisingly, self-reported ADL measures are not well suited for most adult populations of PLWH, with even less relevance to pediatric cohorts and individuals residing in resource-restricted settings. Objective measures of ADL capacity that simulate real-world tasks (e.g., balancing a checking account) are viable alternatives for select adult populations [47–48], but these methods require significant investment in assessment/scoring time and the tasks do not readily translate to international settings. More work is needed to develop brief yet sensitive objective measures of ADL capacity that have strong cultural relevance in both resource-enriched and resource-restricted settings. The latter may require a paradigm shift to address superordinate skills (e.g., transportation/navigation and bartering/trade) rather than the current focus on specific subordinate tasks (e.g., driving a car and balancing a checkbook).

Conclusion

The neurocognitive phenotype of HIV remains subcortical in the post-ART era. Evidence does not support a shift towards a mixed cortical-subcortical phenotype at the present time. The neurocognitive phenotype may shift in the future, reflective of numerous candidate risk factors that persist in the context of long-term ART. Development of a common test protocol and use of standardized nomenclature will enhance data harmonization and create opportunities for deep neurocognitive phenotyping using advanced computational methods.