Semantic memory navigation in HIV: Conceptual associations and word selection patterns

Abstract

Objective:

This proof-of-concept study aimed to characterize semantic memory profiles in individuals with human immunodeficiency virus (HIV) and mild neurocognitive impairment.

Method:

Using a semantic relatedness task, we explored conceptual association and word selection patterns in people living with HIV PLWH; n = 50) relative to people living without HIV (n = 46). We also studied whether word selection patterns in the PLWH group were associated with working memory capacity, cognitive flexibility and inhibitory control.

Results:

While accuracy did not differ between groups, PLWH produced significantly longer responses than controls (r = .32), with fewer hypernyms (d = .47), more troponyms (r = .37), and words that were more frequent (r = .39) and had more phonological neighbors (r = .22). These patterns survived covariation with participants’ cognitive status. None of these patterns correlated with measures of working memory, cognitive flexibility, inhibitory control or viral load (all correlation coefficients < .36).

Conclusions:

Together, these results suggest that PLWH might use alternative word finding strategies during semantic memory navigation, irrespective of the severity of other cognitive symptoms. Such findings contribute to the characterization of cognitive deficits in HIV and to the search for novel markers of the condition.

1. Introduction

Accruing research indicates that human immunodeficiency virus (HIV) can compromise diverse cognitive skills (Wang et al., 2020; Winston & Spudich, 2020). In addition to executive problems, deficits have been detected in semantic memory tasks, pointing to abnormal concept processing (Tierney et al., 2018). Yet, the evidence is scant and no study has examined which associations and word selection patterns characterize semantic memory search in the disorder. To bridge this gap, we examined conceptual and psycholinguistic features of responses to a semantic relatedness task in people living with and without HIV (PLWH, PLWoH).

Affecting nearly 40 million people (Nygren-Krug, 2018), HIV compromises the immune system, increasing vulnerability to opportunistic infections, specific cancers (World Health Organization, 2023), and brain anomalies (Schouten et al., 2011). The latter often lead to HIV-associated neurocognitive disorders (HAND), a spectrum of varyingly severe dysfunctions (Clifford & Ances, 2013). Early research has linked HIV-associated dementia (HAD) to deficits in orientation, concentration, memory, and calculation (Jordan et al., 1985; Navia et al., 1986). Current studies, however, have targeted executive functions, revealing impairments of working memory (Harrison et al., 2017; Janssen et al., 2015), cognitive flexibility (Dawes et al., 2008; Schouten et al., 2011), and inhibitory control (Chang et al., 2002; Walker & Brown, 2018). Yet, HAND also extends to other domains.

In particular, several studies point to difficulties with semantic memory, a complex knowledge system mediating the processing of words, pictures, objects, sounds, faces, and events (Lambon Ralph et al., 2012). Unlike early studies suggesting that semantic memory was unaffected by HAD (White et al., 1997), contemporary reports on PLWH have documented deficits in semantic event sequencing (Melrose et al., 2008), semantic clustering (Gongvatana et al., 2007), word learning (Robbins et al., 2013), and conceptual recognition (Chan et al., 2012; Milanini et al., 2016; Robbins et al., 2013). Semantic memory assessments, then, can illuminate the scope of HAND and contribute to patient characterization and phenotyping (Dastgheyb et al., 2019).

However, this domain remains underexamined and poorly understood in HIV research. Most studies restrict their analyses to valid response counts or response time measures, overlooking which features typify patients’ choices semantic memory search. Suggestively, emerging evidence indicates that HIV may affect two relevant aspects, namely: conceptual association and word selection patterns.

Cognitive studies on HIV have revealed abnormal conceptual association –i.e., non- standard semantic relations between specific stimuli and participants’ responses. For example, a picture naming study showed that, relative to PLWoH and cognitively preserved PLWH, persons with HAND made more associative errors, often replacing the items’ correct names by hierarchically less abstract terms (e.g., ‘Vesuvius’ for ‘volcano’) or words denoting the item’s function (e.g., ‘painting’ for ‘easel’), in addition to other semantic features (Tierney et al., 2018). The preference for less abstract associations aligns with evidence of abstract thinking deficits in HAND (Chan et al., 2012, Woods et al., 2009), indicating suboptimal access to high-level semantic categories and prompting the hypothesis of distinct difficulties with retrieving coarse-grained lexical units (i.e., hypernyms). In the same vein, categorization deficits in HIV have been revealed via nonverbal materials, as seen in tasks requiring picture-to-concept mappings (Grant et al., 1987; Heaton et al., 1995; Spies et al., 2012; Woods et al., 2009). Taken together, prior studies suggest that HAND may be typified by atypical concept association patterns.

Also, HAND may entail a preference for psycholinguistically simple words during vocabulary search. For instance, a verbal fluency study showed that, compared with normative data (Ferreres et al., 2007), words produced by PLWH were more frequent, familiar, and concrete, as well as shorter and with more common phonological structures (Rofes et al., 2021). Moreover, specific response properties, such as frequency, emerged as relevant predictors of patients’ overall fluency performance (Rofes et al., 2021). Suggestively, too, a preference for shorter, more frequent, more concrete, and more phonologically common words has been reported in other populations exhibiting cognitive deficits, including persons living with Alzheimer’s disease (Ferrante et al., 2024) and Parkinson’s disease (Toro-Hernández et al., 2024). Accordingly, psycholinguistic measures may also represent useful markers of semantic association anomalies in HIV.

Of note, word retrieval processes hinge on executive skills affected in HIV (Heaton et al., 2010, 2011), such as working memory (Acheson & MacDonald, 2009; Rosen & Engle, 1997), cognitive flexibility (Hedman et al., 2022), and inhibitory control (Johnson & Anderson, 2004) –but see Witten et al. (2015). More particularly, research on other conditions typified by semantic memory deficits has revealed associations between specific word properties (e.g., frequency) and executive abilities (Ferrante et al., 2024). Therefore, correlations between conceptual/lexical features and executive outcomes could reveal whether predicted semantic memory deficits in PLWH are influenced by broader cognitive difficulties.

Here, using a validated task (Joanette et al., 2004), we presented PLWH and PLWoH with pairs of concepts and asked them to state what they had in common. Instead of measuring whether responses matched predefined correct answers, we combined qualitative and quantitative analyses to capture the type of association and the psycholinguistic properties of each response. Based on previous findings, we predicted that, irrespective of general accuracy, PLWH would (i) favor less abstract and more function-based associations, manifested by (ii) short, highly frequent, and phonologically common words. Also, we explored whether discriminatory features were associated to patients’ outcomes in executive domains typically affected by HIV. With this approach, we aim to increase current understanding of semantic memory deterioration in persons with HIV.

2. Methods

The study’s methods are diagrammed in Figure 1.

Figure 1.

Study design. (A) Participants and cognitive assessments. (B) Summary of the semantic relatedness task. (C) Dependent variables and analytical methods. PLWH: people living with HIV; PLWoH: people living without HIV.

2.1. Participants

The study involved 96 native Spanish speakers from Argentina, namely, 50 PLWH and 46 PLWoH. This sample size reaches a power of 0.95 (Supplementary material, section S1). All participants were Hispanic/Latinos and none identified with any other ethno-racial group. All participants were functionally preserved and reported an absence of developmental language problems, no history of neurological or psychiatric disease, no history of substance use, and normal scores on the Mini-Mental State Examination (⩾ 27/30, Butman et al., 2001). No PLWH had AIDS at the time of testing. Thirty-five were on antiretroviral treatment (not efavirenz), with a viral load below 50 copies/mL. The remaining 15 were untreated, with a CD4+ count above 350 cells/mL and a mean viral load of 22,735 (log₁₀: 4.36). All patients were further evaluated with tests of working memory, via a direct and backward digit span task (Wechsler et al., 2012); cognitive flexibility, with the Trail-Making Test Subsets A and B (Strauss et al., 2006); and inhibitory control, through the Hayling Test (Abusamra et al., 2006) and the Stroop Test (Golden, 1994). No participant had a formal diagnosis of dementia based on current criteria (Antinori et al., 2007). Neuropsychological outcomes showed that 12 exhibited normal cognitive profiles (within 1 SD of all screening tasks), 28 met criteria for asymptomatic neurocognitive impairment (ANI) and mild neurocognitive disorder (MND) (at least 1 SD below adjusted norms in at least two domains), and 10 exhibited outcomes compatible with HAD (at least 2 SDs below adjusted norms in two or more domains). PLWH and PLWoH were matched on sex, age, and years of education (Table 1). Demographic characteristics, viral load, CD4 levels, neuropsychological outcomes, and cognitive status of participants in PLWH group are offered in the Supplementary material (section 2).

Table 1.

Participants’ sociodemographic and cognitive profiles.

	PLWoH (n = 46)	PLWH (n = 50)	Statistics	p-value
Sociodemographic profile
Sex (F:M)	17:29	15:35	X2 = .52	.47a
Years of age	41.8 (8.9)	40.5 (9.2)	t = .75	.45b
Years of education	13.9 (4.1)	13 (3.5)	t = 1.16	.25b
Cognitive profile
Backward digit span (score)	---	4.43 (.25)	---	---
Trail-Making Test B (score)	---	66.3 (6.2)	---	---
Hayling Test (score)	---	.59 (.09)	---	---

Data presented as mean (SD), except for sex.

^ap-values calculated via a chi-squared test (χ²).

^bp-values calculated with independent samples t-tests. PLWH: people living with HIV; PLWoH: people living without HIV.

All participants signed written informed consent in accordance with the Declaration of Helsinki. The project was approved by the Central Ethics Committee of the Ministry of Health (record 18/2010).

2.2. Materials and procedure

The semantic relatedness task was taken from the validated Spanish version of the Protocole Montréal d’Évaluation de la Communication –MEC (Ferreres et al., 2007), a standardized test previously used in other populations with cognitive impairment (Guinjoan et al., 2015; Lomlomdjian et al., 2017). This test taps on participants’ ability to identify and explain semantic relationships between word pairs. It comprises 24 two-word trials. Half of these were used as target trials, featuring words from the same semantic category. The remaining half were presented as fillers to ensure that relatedness judgments were not being made randomly. All words are used in two different pairs, one featuring a traceable to the same semantic category via hypernymy (e.g., ‘apple’ and ‘plum’, traceable to ‘fruit’) and one lacking such hypernymic relation (e.g., ‘plum’ and ‘ship’). Each word pair is presented on a separate sheet, simultaneously in oral and written form. Participants are asked to determine whether the words are conceptually related and, if so, what their relationship is. The task lasts between 3 and 5 minutes. Its identification score (yes/no response) has an internal consistency of α = 0.73, while its decision score (the provided explanation of the words’ relation) has an α = 0.77 and an interrater reliability of 0.92 (Fonseca et al., 2008).

2.3. Performance measures

2.3.1. Accuracy

Accuracy was established following the MEC’s standard procedures (Joanette et al., 2004), elaborated from the normative data of 180 healthy participants. Responses were given 1 point whenever participants correctly established whether the two items mentioned were semantically related, yielding a possible maximum of 24 points. An additional point was assigned if they could correctly identify their prototypical semantic relation, with a possible maximum of 12 points, as established in MEC norms.

2.3.2. Conceptual association

To establish the conceptual association expressed by each response, we employed a validated lexico-semantic categorization scheme following Khoo & Na (2006). We considered six categories: hypernymy, referring to concepts which are hierarchically more abstract than both stimuli (e.g., for lion and dog: ‘They are animals’); troponymy, indicating function, manner or result (e.g., cigar and pipe: ‘They are both used to smoke’); meronymy, highlighting a part in a part-whole relation (e.g., bomb and rifle: ‘Both carry gunpowder’); cause-effect, indicating a consequence of the elements (e.g., cigar and pipe: ‘They cause cancer’); synonymy, highlighting semantic overlap between the items (e.g., ‘They are two cigarettes’); and holonymy, highlighting the whole in a part-whole relation (e.g., bomb and rifle: ‘They are part of the war’). After setting the guidelines for response coding, two expert raters (a psycholinguist and a neuropsychologist) independently coded all responses. Following Syed and Nelson (2015), a third rater (a neurolinguist) was invited to discuss discrepant classifications and responses which did not clearly fit any of the categories. Interrater reliability between the first two raters was 96.52 % (κ = .94, 95% CI = 0.90 – 0.97). Most discrepant cases were successfully settled with the third rater, as only 4 trials (less than .01% of the data) failed to elicit consensus and were thus labeled as ‘unclassifiable’.

2.3.3. Psycholinguistic properties

Responses were further analyzed in terms of their psycholinguistic properties, via automated pipelines within the TELL app (García et al., 2023). As in recent works (Ferrante et al., 2024; Toro-Hernández et al., 2024), we tagged every word in each response, retained all content words (nouns, verbs, adverbs, adjectives), and extracted the following normative features from the EsPal database (Duchon et al., 2013): frequency (occurrences in daily language use), familiarity (recognizability of the referent), length (number of phonemes), phonological neighbors (number of words with similar sublexical structure), and concreteness (sensorimotor apprehensibility). Each of these variables was averaged across all content words in each response, yielding one mean value per response.

2.4. Data analysis

First, we analyzed between-group differences in the percentage of correct answers. Due to non-normal distribution of the data, these analyses were based on the Wilcoxon rank-sum test (Wilcoxon, 1992). Effect size was estimated via the rank-biserial correlation coefficient (r).

Second, between-group differences in the proportion of each type of conceptual association were analyzed with Pearson’s chi-squared test (Sachs, 2012). Unclassifiable and null responses (i.e., those lacking a description from the participant) were excluded from analysis. Post-hoc analyses were performed with Bonferroni correction. Effect sizes were estimated through Cohen’s d (Cohen, 2013).

Third, between-group differences for each psycholinguistic feature were assessed with Wilcoxon’s rank-sum test, due to non-normal distribution of the data. Effect sizes were estimated via rank-biserial correlation coefficient (r). Analyses were conducted on target trials for which an explanation was provided. This trimming resulted in removal of 11 data points (1.01%).

Responses that exceeded 3 SDs in word length were removed. This data trimming resulted in the removal of 7 data points from the control group (1.31%) and 6 observations in the HIV patient group (1.24%), resulting in the inclusion of 1003 observations for final analysis. To ensure that results were not driven by specific data trimming procedures, we replicated all analyses upon excluding (i) missing norming data (as defined above) and (ii) outlier responses (beyond 3 SDs from the mean length of the participant’s group (Supplementary material, sections 3 and 4). We further ran analyses considering ‘cognitive status’ as a covariate (Supplementary material, section 5).

Finally, to examine whether atypical semantic memory traits were associated to patients’ general cognitive dysfunctions, we performed correlations between (a) significantly altered psycholinguistic features and (b) cognitive tests tapping on working memory (backward digit span), cognitive flexibility (Trail-Making Test B), and inhibitory control (Hayling test). Additional correlations were performed with the participants’ viral load. To favor comparability across tasks, patients’ scores on each cognitive test were transformed into z-scores using normative means and standard deviations from PLWoH from Ferreres et al. (2007). Given data distribution, we used Pearson correlation tests, adjusting p-values adjustment for multiple comparisons via the False Discovery Rate (FDR) measure.

All analyses were conducted on R (R Core Team, 2013), with alpha levels set at p < .05. Visualizations were produced with Ggplot’s R library (Wickham et al., 2016).

2.5. Data availability

All data and the scripts used for analysis are publicly available at https://osf.io/fx6ck/?view_ only=b5ab9136d07941c4ae070655e4d5967e.

3. Results

3.1. Accuracy

Accuracy was high for both PLWH (97.6% ±.44) and PLWoH (97.8% ± .44). A Wilcoxon rank-sum test revealed no significant differences between groups (Z = −0.38, p = .69, r = 0.039).

3.2. Conceptual associations

Approximately 95% of the responses provided by PLWH and PLWoH were hypernyms and troponyms. The distribution of response types across all six categories differed between groups (χ² (5, 91) = 20.87, p < .001, V = .14). Post-hoc comparisons revealed that PLWH produced fewer hypernyms (p < .001, d = .43) and more troponyms (p = .028, d = .37) than PLWoH. Differences in the distribution of other types of association were not significant (all p-values > .26). For details, see Table 2 and Figure 2.

Table 2.

Distribution of conceptual associations in the semantic relatedness task.

Group	Hypernym	Troponym	Meronym	Cause-Effect	Synonym	Holonym	Unclassifiable	Null response
PLWoH (n = 46)	91.64	4.91	1.45	0.91	0.00	0.00	0.36	0.73
PLWH (n = 50)	83.36	9.75	2.35	2.86	0.17	0.34	0.34	0.84

Values expressed as percentages over the total set of responses per group. PLWH: people living with HIV; PLWoH: people living without HIV.

Figure 2.

Significant effects in conceptual association analyses. Proportion of hypernyms (A) and troponyms (B) provided by people living without HIV (PLWoH; n = 46) and people living with HIV (PLWH; n = 50) in the semantic relatedness task. Black circles represent the group’s mean and black bars represent standard errors. The boxplot represents the distribution of responses including the median (dark line), interquartile range, and outliers (colored circles).

3.3. Psycholinguistic properties

PLWH used significantly more content words than PLWoH (Z = −3.09, p < .01, r = .32). Also, compared with PLWoH, PLWH produced words that were significantly shorter (Z = −2.87, p < .01, r = .29), more frequent (Z = −3.87, p < .001, r = .39), and with more phonological neighbors (Z = −2.092, p = .03, r = .22). No significant between-group differences emerged for familiarity (Z = −1.09, p = .27, r = .11) or concreteness (Z = −1.30, p = .19, r = .13). See Figure 3 and Table 3. All significant effects, except for ‘familiarity’, remained so upon rerunning analyses without removing outliers (Supplementary material, section 4) and when adding ‘cognitive status’ as a covariate (Supplementary material, section 5). Also, ‘cognitive status’ did not yield significant effects on any of the psycholinguistic measures (all p-values > .07).

Figure 3.

Significant effects in psycholinguistic feature analyses. Mean values for (A) number of words (B) phonemes per word, (C), frequency, and (D) number of phonological neighbors. Black circles represent each group’s mean and black bars represent standard errors. The boxplot represents the distribution of responses including the median (dark line), interquartile range, and outliers (colored circles). PLWH: people living without HIV (n = 46); PLWH: people living with HIV (n = 50).

Table 3.

Length and psycholinguistic properties of responses to the semantic relatedness task.

Psycholinguistic feature	Mean (SE)		z	p	r
	PLWoH (n = 46)	PLWH (n = 50)
Response length (in content words)	2.39 (.06)	2.62 (.06)	−3.09	<.01	.32
Log frequency	1.86 (.03)	2.01 (.03)	−3.86	<.001	.39
Number of phonemes	5.24 (.07)	4.96 (.07)	−2.87	<.05	.29
Familiarity	5.97 (.02)	6.05 (.02)	−1.97	<.05	.20
Phonological neighbors	26.92 (.75)	28.58 (.65)	−2.09	<.05	.22
Concreteness	4.87 (.04)	4.94 (.04)	−1.22	.22	.12

Values expressed as percentages over the total set of responses per group. PLWH: people living with HIV; PLWoH: people living without HIV.

3.4. Correlations between psycholinguistic properties and cognitive outcomes

In the PLWH group, no discriminatory psycholinguistic property (frequency, length, phonological neighbors) significantly correlated with scores on any cognitive test (backward digit span, Trail-Making Test B, Hayling Test) or with viral load –all p-values > .07; all effect size estimates (r) < ± .36 (Table 4). No significant correlations were found upon rerunning analyses without removing outliers (Supplementary material, section 6).

Table 4.

Correlations between psycholinguistic properties of responses in semantic relatedness task and performance in cognitive tests in people living with HIV (n = 50).

	Backwards digit span		Hayling (subset B)		Trail-Making Test (subset B)		Viral load
Psycholinguistic property	r	p	r	p	R	P	r	p
Log frequency	−.16	.53	.22	.32	−.01	.97	−.02	.97
Length (number of Phonemes)	.35	.10	−.34	.09	−.12	.68	−.01	.97
Phonological neighbors	−.31	.10	.29	.12	.07	.96	.01	.97

4. Discussion

We examined semantic memory alterations in HIV via a semantic relatedness task. Relative to PLWoH, PLWH established fewer hypernymic and more troponymic relations between concepts, using words that were shorter, more frequent, and with more phonological neighbors. These psycholinguistic patterns were not associated with patients’ general cognitive dysfunctions, speaking to their generalizability across neuropsychological profiles. Below we address these findings.

When establishing conceptual associations, PLWH produced fewer hypernyms (d = .47) and more troponyms (d = .37) than PLWoH. This mirrors findings from a picture naming task in which participants with HAND made more associative semantic errors than cognitively preserved PLWH and PLWoH (Tierney et al., 2018), replacing the items’ names with less abstract subordinates (e.g., proper names) and words denoting the item’s function, among other inadequate labels. Difficulties accessing hierarchically higher associations of a stimulus, such as hypernyms, might reflect reduced abstraction capacity, as shown in non-verbal tasks requiring picture-to-concept mappings in symptomatic PLWH (Heaton et al., 1995) and even seropositive PLWH without a diagnosis or signs of HIV-related dementia (Grant et al., 1987), irrespective of sex (Spies et al., 2012). Indeed, hypernym activation involves recognizing that both words in a trial share an overarching (i.e., more abstract) category and retrieving the label of such a category. As proposed for other populations showing reduced capacity to access abstract concepts (Toro-Hernández et al., 2024), such patterns indicate difficulties with accessing semantic memory information that is removed from direct sensorimotor motor experience. Conceivably, too, increased reliance on troponyms might reflect a compensatory strategy driven by sensorimotor proximity. Such associations involve functional features of the items, including their affordances or potential physical uses, thus being closer to situated experience with the environment (Ferreres et al., 2007). Given that concept retrieval demands are greater for abstract than for concrete items, these patterns suggest that PLWH might have a tendency to reduce cognitive effort during semantic memory navigation.

This claim aligns with results from our word property analyses. Compared with PLWoH, PLWH produced shorter (r = .32), more frequent words (r = .39) which presented more typical phonological structures (r = .22). Similarly, a verbal fluency study on PLWH (Rofes et al., 2021) showed that task performance is modulated by word properties like frequency, familiarity, and imageability. This pattern also mirrors findings from other cognitively impaired populations. For instance, people living with Alzheimer’s disease, compared with healthy controls, produce words that are more frequent, less specific, and with more phonological neighbors (Ferrante et al., 2024). Likewise, when describing properties of specific concepts, people living with Parkinson’s disease produce more concrete and more imageable words than do healthy controls (Toro-Hernández et al., 2024). Importantly, compared with low frequency, short, and fewer phonological neighbors, words with high frequency, fewer phonemes, and more phonological neighbors are known to entail reduced cognitive effort, as seen in measures of response time (Grainger, 1990; Monsell et al., 1989), reading time (Kuperman et al., 2024), and brain activation patterns (Carreiras et al., 2009; Schuster et al., 2016). Thus, in line with previous findings, our results suggest that vocabulary navigation in HIV would be typified by a tendency to favor the most accessible spaces of semantic memory.

Yet, note that accuracy is at ceiling and did not significantly differ between PLWH and PLWoH. In a similar vein, previous studies reported spared outcomes in PLWH with mild cognitive deficits when compared to PLWoH on tasks requiring access to semantic memory (White et al., 1997). Thus, these psycholinguistic patterns probably reflect a particular cognitive strategy rather than an impairment per se, potentially influencing daily communicative choices. The need to reduce cognitive effort during spontaneous conversation might entail a (plausibly unconscious) preference for low-demand (e.g., hyponymic or use-related) associations and easily retrievable (e.g., highly frequent) words. These strategies would allow PLWH to bypass potential deficits in abstract semantic processing while facilitating continual lexical selection –as observed in individuals with other language difficulties, who often favor more imageable, frequent words to convey meaning (Bastiaanse, Wieling & Wolthuis, 2015; Gahl & Menn, 2016).

Finally, results of the correlations analyses in PLWH showed no significant association between word property patterns and either executive capacity or viral load. Previous works have revealed mixed findings in this regard. For instance, Poutiainen, Elovaara, Raininko, Hokkanen, Valle, Lähdevirta, & Livanainen (1993) showed no correlation between language skills, other cognitive skills, and cerebral atrophy. Likewise, a study on verbal memory and learning capacity found no significant correlations with outcomes from the Stroop color-word task, the Color Trails Test, or the Wisconsin Card Sorting Test (Witten et al., 2015). Conversely, Gongvatana, Woods, Taylor, Vigil & Grant (2007) showed that, in patients with HAND, verbal memory and learning capacity depended on executive outcomes, with more dysexecutive patients showing differences on word recall and semantic grouping. In our case, the lack of correlations could be driven by our inclusion criteria, as most participants were at early stages of neurocognitive decline and had normal scores in at least two out of four executive tests, reducing cognitive variance needed for correlations to emerge. Be that as it may, our findings extend previous results by suggesting that the use of psycholinguistically more accessible words characterizes PLWH irrespective of their cognitive skills. While these links could not yet be established, it is matter of future research to explore whether a more comprehensive battery of executive functions tests and sample could provide better insight on this matter.

More generally, our findings should be interpreted against previous research aiming to disentangle the effects of HAND and age-related cognitive decline. Prior studies on the interaction between natural aging and HIV have yielded mixed results, with some suggesting that HIV accelerates neurological aging (Cysique & Brew, 2014; Pfefferbaum et al., 2014; Sacktor et al., 2010) and others providing contrary outcomes (Cole et al., 2018). Although our study does not offer conclusive evidence on this matter, we note that our groups were age-matched, ruling out the role of ‘age’ as a key driver of the observed effects.

Lastly, this work bears clinical implications. Cognitive assessments in HIV have typically focused on attention, executive functions, and processing speed, reducing language testing to verbal working memory and fluency tasks. Our study extends an incipient literature showing that HIV may affect semantic memory processes. Crucially, while most semantic memory studies have focused on response accuracy and speed, our work suggests that cognitive markers of HIV may be found in the type of responses provided. Such analyses have already been claimed to streamline cognitive assessments in other populations, such as persons with Alzheimer’s and Parkinson’s disease (Ferrante et al., 2024; Toro-Hernández et al., 2024). By the same token, we propose that semantic memory assessments may contribute to the characterization, detection, and phenotyping of persons with HIV, broadening the scope of relevant cognitive tests in clinical settings.

5. Limitations and avenues for further research

Our study is not without limitations. First, though our sample was similar to or larger than those of previous studies, replications with larger and more diverse samples would be important to test the robustness of our findings. This could be done, for instance, by examining ethno-racial differences, the mediating role of socioeconomic background in participants’ cognitive performance, and the interaction of HAND with other comorbidities directly or indirectly related to HIV. Second, we lacked a direct measure of abstraction capacity, which would provide a direct testing ground for our interpretation of conceptual association results. Finally, more fine-grained insights could be gained through longitudinal and cross-sectional studies with larger samples of participants at different stages of HAND (ideally including nadir CD4 data), potentially revealing whether and how semantic memory patterns change over time across cognitive phenotypes.

6. Conclusion

Cognitive alterations in HIV seem to extend to fine-grained aspects of semantic memory. Compared to PLWoH, PLWH seem to produce fewer hypernymic and more troponymic conceptual associations, suggesting reduced abstraction capacity. Also, they exhibit a preference for short, frequent and phonologically common words, revealing a preference for highly accessible items during semantic memory search. Such patterns were not associated with patients’ cognitive outcomes, speaking to their potential generalizability across neuropsychological profiles in the HIV population. Future studies along these lines may further our understanding of the multifarious impact of HIV on semantic memory and cognitive abilities at large.