Visuospatial working memory in Huntington’s disease: behavioural and structural brain correlates

Abstract

Background

Visuospatial working memory (VSWM) deficits are among the earliest cognitive impairments evident in Huntington’s disease (HD). They often precede motor symptoms and likely reflect early disruption of cortico-striatal circuits. Both the cognitive components and brain correlates are, however, poorly delineated.

Objectives

This study aims to characterise the progression of VSWM deficits across HD stages and identify their structural brain correlates.

Methods

In this prospective international study, 92 HD mutation carriers (HDmc) (21 premanifest and 71 manifest) and 39 healthy controls underwent cognitive and clinical assessments and structural MRI at baseline; 49 manifest HD and 36 controls completed a 1-year follow-up. VSWM was assessed using the object–location task (OLT), providing continuous measures of location error, object recognition and object–location binding. MRI analyses included caudate, putamen and hippocampal volumes and cortical thickness.

Results

We found increased location error and spatial imprecision under high memory load, without binding errors in premanifest HD with an area under the curve of 0.731. Across all HDmc, these location deficits correlated with striatal atrophy and cortical thinning in parietal and middle cingulate regions. Manifest HD individuals exhibited swap errors and hippocampal atrophy. Both location and swap errors worsened over time alongside progressive striatal and hippocampal volume loss. Principal component analysis confirmed a dissociation between spatial location-related deficits (striato-cortically mediated) and binding errors (hippocampally mediated), supporting a hierarchical model of VSWM decline.

Conclusions

HD-related VSWM impairment follows a staged, anatomically dissociable trajectory. The OLT offers sensitive, mechanistic markers of cognitive decline with potential utility for early detection and clinical trial stratification.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Visuospatial working memory (VSWM) deficits occur early in Huntington’s disease (HD), but prior studies have largely relied on coarse recognition-based measures that do not dissociate spatial precision from object–location binding or clarify their neural substrates across disease stages.

WHAT THIS STUDY ADDS

• Using a continuous object–location recall task with cross-sectional and longitudinal data, this study shows that spatial precision deficits emerge in premanifest HD before object–location binding failures, with early impairments associated with parietal and cingulate cortical thinning and later binding errors linked to hippocampal atrophy in manifest HD. These findings support a hierarchical breakdown of VSWM networks, progressing from cortico-striatal dysfunction to hippocampal-dependent associative memory failure.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• By isolating distinct VSWM components and their neural correlates, this study highlights the value of load-sensitive, continuous spatial measures as early cognitive markers in HD, with potential utility for disease monitoring, stratification and outcome measures in clinical trials.

Introduction

Visuospatial working memory (VSWM) refers to the temporary retention and manipulation of visuospatial information and relies on a distributed yet partially distinct neuronal network.¹ Object identity, the recognition of visual objects, mainly engages the ventral visual stream, whereas spatial updating, that is, the dynamic tracking of object locations, is mediated by the dorsal visual stream and dorsal striatum.² The hippocampus supports object–location binding and allocentric spatial representations.³ Parietal and prefrontal cortices provide attentional and executive resources under increasing memory load.^{4 5} Although these components are functionally distinct, they are rarely assessed together in clinical settings, limiting our understanding of their interactions and hierarchical deterioration in neurodegenerative disorders.

Huntington’s disease (HD), an autosomal dominant neurodegenerative disorder caused by a CAG repeat expansion in the HTT gene,⁶ provides a compelling model with which to examine the degradation of different VSWM components over time. Visuospatial disruption may appear before clinical diagnosis (usually defined by onset of motor signs) and is often reported early by HD mutation carriers (HDmc) and their relatives as spatial disorientation or object misplacement.^{7 8} These symptoms likely reflect subtle VSWM deficits already detectable in the premanifest stage.^9–11 These emerge alongside striatal degeneration while hippocampal and cortical atrophy remains minimal,¹² providing a window to tease apart striatal (updating), hippocampal (object-location binding) and cortical (executive) contributions to VSWM across disease progression.

Current neuropsychological tools lack resolution to distinguish between different types of memory error, limiting the interpretability of VSWM performance. Specifically, identification accuracy is linked to object recognition, location errors index spatial precision, swap errors reflect failures of object–location binding and nearest-neighbour control distance provides a control measure of general spatial placement independent of object identity. Traditional tasks such as the Visual-Spatial Learning Test (VSLT), the Paired Associates Learning task (PAL) or egocentric versus allocentric paradigms detect group differences but rely on categorical scoring, preventing discrimination between error types.^{8 13} For example, premanifest individuals may show isolated location errors on the VSLT, while manifest patients display broader impairments, yet standard scoring cannot differentiate swap (binding) errors from random guesses.⁸ The PAL task often fails to detect early impairments, potentially due to preserved binding in premanifest HDmc, but its reliance on discrete spatial grids reduces its sensitivity. Mobile-phone-based and virtual-reality tasks have improved ecological validity but still depend on binary pass/fail metrics masking more subtle deficits. Only a single longitudinal study has linked object–location memory with functional neural changes, without fine-grained characterisation of error subtypes.¹⁴

To address these limitations, we implemented the object–location task (OLT) in a cross-sectional comparison of healthy controls, premanifest HDmc and individuals with manifest HD, with longitudinal follow-up in a subset of the HD and control groups. The OLT provides continuous, parametric measures that dissociate spatial imprecision (location errors) from binding errors such as swaps.¹⁵ Our aims were to: (1) characterise VSWM error profiles across HD stages; (2) relate specific error types to structural alterations in striatal, hippocampal and cortical regions and (3) evaluate the sensitivity of these measures as markers of cognitive decline. We hypothesise that distinct patterns of VSWM errors will emerge across the disease continuum, reflecting selective neural vulnerability. Premanifest HDmc, with relatively preserved hippocampi, is expected to show subtle location errors without binding deficits, whereas manifest HDmc should exhibit elevated swap errors associated with hippocampal atrophy and longitudinal memory decline.

Methods

Participants

HDmc and healthy controls were recruited from Repair-HD centres (Cardiff, Créteil and Manchester; NCT03119246) and the Bio-HD cohort (Créteil; NCT01412125). Manifest HDmc (≥38 CAG repeats) were at stages 1–2 (Shoulson staging).¹⁶ Premanifest HDmc (>40 CAG repeats) had no motor (Total Motor Score, TMS<5) or functional (Total Functional Capacity, TFC of 13) impairments. HD-ISS staging² is reported in table 1. Controls were community volunteers or gene-negative relatives. Exclusion criteria included substance abuse and neurological comorbidities (including attention-deficit/hyperactivity disorder). All participants had normal/corrected vision and provided informed consent. French ethics committees approved the protocols (Saint Louis for Repair-HD and Créteil for Bio-HD).

Table 1.

Demographics, neurological description and general assessment at baseline (means and SD)

	Controls	Premanifest	Manifest	p values*	Post-hoc p values†
					C vs PM	C vs M	M vs PM
Number	39	21	71
(Cardiff/Créteil/Manchester)	(5/32/2)	(0/21/0)	(19/41/11)
Stage HD-ISS (0–1/2/3/NA)	–	(19/2/0/0)	(0/11/58/2)
Age	50.75 (8.13)	47.24 (8.59)	53.92 (9.61)	<0.01	0.320	0.180	<0.01
Gender	18F/21M	14F/7M	26F/44M	0.050	0.210	0.440	0.030
Education	13.51 (3.19)	14.29 (2.12)	13.27 (3.05)	0.410
Laterality	0A/34R/5 L	0A/18R/3 L	2A/62R/7 L	0.700
CAG repeats	–	42.05 (2.56)	43.15 (2.77)	0.09
TFC	13.00 (0.00)	13.00 (0.00)	10.31 (2.06)	<0.001	1.000	<0.001	<0.001
cUHDRS	17.23 (1.37)	17.23 (1.18)	9.97 (3.00)	<0.001	0.990	<0.001	<0.001
CAP score	–	386.80 (80.30)	494.70 (86.60)	<0.001
Estimated onset	–	6.50 (7.88)	–
IES-global	1.66 (0.36)	1.81 (0.39)	3.58 (1.55)	<0.001	0.51	<0.0001	<0.0001
IES-visuospatial	0.87 (0.31)	0.97 (0.31)	2.11 (1.20)	<0.001	0.65	<0.0001	<0.0001
IES-executive	0.81 (0.27)	0.997 (0.36)	1.59 (0.76)	<0.001	0.03	<0.0001	<0.0001
IES-language	3.52 (0.92)	3.75 (1.00)	7.51 (4.08)	<0.001	0.80	<0.0001	<0.0001
IES-memory	1.44 (0.41)	1.54 (0.61)	3.02 (1.51)	<0.001	0.78	<0.0001	<0.0001
SDMT	52.05 (10.06)	52.81 (8.02)	29.17 (9.76)	<0.001	0.79	<0.0001	<0.0001
Letter fluency	40.88 (9.85)	41.95 (9.18)	26.19 (9.67)	<0.001	0.63	<0.0001	<0.0001
HVLMT-RI	28.56 (3.63)	26.52 (5.50)	19.58 (6.09)	<0.001	0.0049	<0.0001	<0.0001
HVLMT-RD	10.26 (1.83)	9.72 (1.90)	6.14 (3.01)	<0.001	0.20	<0.0001	<0.0001
MDRS	141.87 (2.69)	141.14 (3.52)	130.71 (7.76)	<0.001	0.58	<0.0001	<0.0001

In all cognitive tasks except for the SelfCog higher values reflect better performance.

1 missing value for education level. Age, education and estimated onset are presented in years. CAP Cumulative CAG-related disease burden. IES Inverse Efficiency Score from the SelfCog battery.

*p-values for chi-2 test and ANOVA.

†post hoc Tukey

A, ambidextrous; C, control; CAG, cytosine adenine guanine and represent the number of sequence repeats; cUHDRS, Composite Unified Huntington’s Disease Rating Scale score; DBS, Disease Burden Score; F, female; L, left; M, male; M, manifest; PM, premanifest; R, right; TFC, total functional capacity.

Assessments were conducted at Month 0 (M0, baseline), M1 and M12. As previously demonstrated in HD, using M1 rather than M0 as a baseline reduces retest effects and better captures cognitive decline over 1 year.¹⁷ M1 was therefore used as a baseline in all longitudinal analyses. The UHDRS (Unified Huntington’s Disease Rating Scale,¹⁸ TMS and TFC) was collected at M0 and M12, whereas the UHDRS cognitive battery—Letter Verbal Fluency, Symbol Digit Modalities Test (SDMT) and Stroop Word Test—was administered at all three visits. Additional cognitive measures at M0, M1 and M12 included the Mattis Dementia Rating Scale (MDRS), the Hopkins Verbal Learning Test (HVLT) and the SelfCog battery.¹⁹ The MDRS assesses global cognition; the HVLT evaluates verbal memory and SelfCog provides a digital assessment of language, visuospatial, executive and memory functions, yielding an Inverse Efficiency Score (IES; higher values indicate poorer performance). The composite UHDRS (cUHDRS)²⁰ and the cumulative CAG-related disease burden²¹ were computed.

131 participants (71 manifest, 21 premanifest and 39 controls) completed baseline assessments; 81 had MRI data (35 manifest, 15 premanifest and 31 controls). Manifest participants were older and more often male; there were no differences in TFC/cUHDRS between premanifest and controls, but manifest HD showed poorer performance on all measures (table 1). 85 participants (49 manifest and 36 controls) completed the 1-year follow-up (online supplemental table S1), with MRI from 59 (29 manifest, 30 controls, matched for demographics). Longitudinal follow-up was embedded within the RepairHD cohort design (M0, M1 and M12) and was therefore available only for manifest HD participants and healthy controls. Premanifest participants were recruited from the BioHD cohort and did not undergo follow-up assessments in the present study. Among manifest participants and controls, some did not complete follow-up due to limited availability of specialised assessors at certain centres.

Follow-up sessions were not offered to premanifest participants, and some manifest patients and controls did not complete them due to limited availability of specialised assessors at certain centres.

Object location task

The OLT was adapted from Pertzov et al.⁸ (see online supplemental data). The OLT was implemented in PsychoPy (Psychopy.org), which recorded all behavioural responses. Participants completed the task on laptops with comparable physical screen sizes (~15.3″) across sites and were seated in a standardised position to maintain consistent visual angle calculations.²²

Stimuli consisted of 60 coloured drawings of everyday objects (manufactured items, plants/flowers, fruits/vegetables and animals), each scaled to fit within an 8×8° visual-angle square. In the trial, all objects were positioned at least 3.9° from the screen edges, 6.5° from the centre, and a minimum of 9° from each other (centre-to-centre).²²

Participants completed 40 trials (eight per memory-load level), with memory load defined by the number of objects presented at encoding (from 1 to 5). Load levels were randomly assigned, with the constraint that no more than three consecutive trials shared the same load. Objects were randomly sampled with no more than three occurrences per object and no consecutive repetitions.

Participants performed a brief visuomotor calibration before and after the OLT (five dot-click trials each), in which they were asked to click on a dot presented on the screen, to control for potential motor confounds. In the OLT, on each trial (figure 1A), participants encoded an array of 1–5 objects for up to 5 s (duration scaled with memory load). After a 1 s delay, they completed: (1) object-identity recognition (target vs distractor) and (2) object-location recall by using the mouse to place the recognised object to the remembered position on the screen. Participants confirmed responses with the spacebar to reduce errors due to choreic movements and to minimise potential manual laterality effects.

Figure 1.

Design of the Object-Location Memory Task: sequence of task procedure (A) and the four visuospatial memory measures employed in the task (B).

Behavioural measures (see figure 1B)^{15 22} included: (i) identification accuracy, the percentage of correctly identified objects; (ii) location error, the angular deviation (degrees of visual angle) between response and the correct location; (iii) swap errors, the percentage of trials in which the object was placed within 3° of another object’s location and (iv) the nearest-neighbour control quantified the distance to the nearest object regardless of identity, thereby controlling for general spatial localization ability and isolating performance related to spatial position independently of object–location binding.¹⁵ Location error, swap errors, and nearest-neighbour measures were computed only for correctly identified objects, with swap and nearest-neighbour measures restricted to trials with more than one object.

MRI data acquisition and pre-processing

3D T1-weighted MRI scans were acquired on the day of clinical and neuropsychological assessments at Créteil. Acquisition parameters (online supplemental data) and preprocessing details are described in Lunven et al.^{19 23} Subcortical (caudate, putamen) and hippocampal volumes were obtained using FreeSurfer 6.0 and its longitudinal pipeline for follow-up (https://surfer.nmr.mgh.harvard.edu) and normalised to intracranial volume. Cortical thickness was analysed using the CAT12/SPM12 (https://neuro-jena.github.io/cat12-help/), with standard baseline and longitudinal pipelines.

Statistical analyses

Analyses were performed in R (V 2022.12.0+353). Age, sex, education, and centre were included as covariates in all linear and linear mixed-effects models.

Behavioral analyses (OLT performance)

At baseline, OLT performance was analysed using linear models including group (controls, premanifest and manifest HDmc), memory load and their interaction. Pairwise comparisons used estimated marginal means with Tukey correction (‘emmeans’). To examine the structure of cognitive and OLT measures, all variables were standardised, and Principal Component Analysis (PCA) was conducted (‘psych’, principal function, two components and varimax rotation).

Longitudinal changes (M1–M12) were assessed with linear mixed-effects models (‘lme4’), including fixed effects of group, time, and memory load and their interactions, as well as random intercepts and slopes at the individual level. Model-based predictions were visualised using ‘ggeffects’. ROC analyses assessed discriminative performance when premanifest HDmc differed from controls (‘pROC’ package), with area under the curve (AUC) as the summary accuracy metric.

Imaging analyses

At baseline, group differences in subcortical and hippocampal volumes were assessed using linear regression, with Spearman correlations examining relationships with OLT performance. To assess the independent contribution of structural volumes beyond disease burden, partial Spearman correlations controlling for disease burden score (age × (CAG repeat − 35.5)) were computed (“ppcor” package). Cortical thickness analyses used FWE-corrected thresholds (p<0.05) and a 20-vertex cluster threshold. Group differences were assessed using a full-factorial model, and relationships between cortical thickness and OLT measures in HDmc were tested using linear regression.

Longitudinal atrophy was analysed using mixed-effects models with group, time, and their interaction as fixed effects. Longitudinal cortical-thickness changes (M12 minus M0) were compared between controls and manifest HDmc using linear regression, and their associations with changes in OLT performance were examined using regression models.

Results

Baseline evaluation

Premanifest participants performed comparably to controls, except on SelfCog executive and HVLT immediate recall. Manifest HDmc showed widespread deficits across all cognitive measures (table 1).

Calibration

Online supplemental table S2 shows comparable location accuracy across groups, with stable performance before and after the OLT, indicating motor symptoms did not bias responses.

Task performance

Accuracy was high across groups (>90%, online supplemental table S3)(figure 2). Memory load reduced accuracy in controls (β=–0.90, p=0.047), with a stronger effect in manifest HD (interaction β=–2.09, p<0.001), but not in premanifest individuals. Manifest HD performed worse than controls (CE=5.02, p<0.001) and premanifest participants (CE=4.37, p=0.0002), while premanifest and controls showed similar performance (CE=0.65, p=0.82).

Figure 2.

Behavioural results. (A) Model-predicted identification memory accuracy, location, swap error and nearest neighbour control measures as a function of memory load (Objects: number of objects presented during memory array). (B) Principal component analysis results in HD gene expanded individuals with the two components (RC1 and RC2). HVLT, Hopkins Verbal Learning Memory; TaskIES, Inverse Efficiency Score from the SelfCog battery; OLT, Object–Location Task; RD, delayed recall; RI, immediate recall.

Memory load increased location error in controls (β=0.84, p<0.001). Both premanifest and manifest HD showed larger errors than controls, with manifest individuals exhibiting the largest deficits (manifest vs controls: CE=–3.25, p<0.001; premanifest vs controls: CE=1.42, p<0.001). The effect of memory load was amplified in HDmc (premanifest: β=0.49, p=0.04; manifest: β=0.96, p<0.001). Group differences widened with increasing load, and manifest HD was impaired even at load 1. The AUC distinguishing premanifest from controls was 0.731 (95% CI: 0.603 to 0.860, DeLong’s method), indicating significant discriminative ability.

Manifest HDmc made more swap errors than controls (CE=–3.35, p=0.0012), driven by a steeper load-dependent increase (β=1.87, p=0.02). Premanifest individuals did not differ from controls (all p>0.5).

Memory load increased nearest neighbour distance in controls (β=0.08, p<0.001), with a trend toward a difference effect in manifest HD (interaction β=−0.05, p=0.057), but no effect in premanifest (interaction β=0.02, p=0.56). Both premanifest and manifest HD showed higher minimal distances than controls (premanifest: CE=−0.58, p=0.0013; manifest: CE=−1.41, p<0.001), with manifest HD also differing significantly from premanifest (CE=−0.83, p<0.0001). Unlike controls and premanifest, memory load did not significantly affect nearest neighbour distance in manifest HD.

Baseline imaging measures

Group characteristics are shown in online supplemental table S4. HDmc had reduced caudate and putamen volumes compared with controls (caudate: CE=0.18, p<0.001; putamen: CE=0.262, p<0.001), with manifest HDmc lower than premanifest caudate: CE=0.100, p<0.001; putamen: CE=0.136, p<0.001). Hippocampal volumes were similar in premanifest and controls but reduced in manifest HDmc (controls vs manifest: CE=0.078, p<0.001; premanifest vs manifest: CE=0.06, p=0.007; figure 3).

Figure 3.

Brain volumetric results: (A) subcortical and hippocampal volumes differences between groups, ***p<0.001 and (B) association between object–location task measures and subcortical and hippocampal volumes.

In HDmc (figure 4), putamen volume correlated with identification accuracy (r=0.36, p=0.03), location error (r=–0.37, p=0.03), and nearest-neighbour distance (r=–0.53, p<0.001). Hippocampal volume correlated with nearest-neighbour distance (r=–0.42, p=0.015) and showed a trend with location error (r=–0.32, p=0.05), while caudate volume trended with nearest-neighbour distance (r=–0.31, p=0.06).

Figure 4.

Model-predicted location error and swap error as a function of memory load and time. Objects: number of objects displayed in trials. M1, Month 1; M12, Month 12; HD, Huntington’s disease.

Cortical thickness analyses (online supplemental tables S5,S6 and figure S1) revealed widespread thinning in manifest HD relative to controls, notably in occipital, temporal, premotor and frontal regions. Compared with premanifest individuals, manifest HD showed predominantly posterior thinning, including primary and secondary visual cortices, superior parietal and temporal areas. Location error correlated with cortical thickness in two right-hemisphere clusters (p-Family-Wise Error (FWE) corrected<0.001): middle cingulate and inferior parietal cortex (online supplemental table S7).

Clinical associations

In mutation carriers, cUHDRS and CAG-Age Product (CAP) score correlated with identification accuracy (respectively: rho=0.53, p<0.001; rho=−0.32, p=0.002) and location error (respectively: rho=0.59, p<0.001; rho=0.31, p=0.003). PCA identified two components explaining 60.5% of variance (figure 2B). RC1 (34.8%) reflected global cognition (verbal memory, processing speed, fluency, SelfCog). RC2 (25.7%) captured VSWM, with high loadings from OLT measures. Executive IES and SDMT also have moderate load on the RC2.

Longitudinal study M1 – M12

Manifest HD participants showed declines over 1 year in TFC, cUHDRS, MDRS, SelfCog (global and language), Stroop interference and letter fluency, after adjusting for demographic covariates (online supplemental table S8).

Accuracy remained stable over time in both controls (reference group, β=0.07, p=0.952) and manifest HD, with no significant group×time (β=0.48, p=0.750) or group × time × load interactions (β=−0.41, p=0.153).

There was no mean change over time in controls (reference group, β=−0.02, p=0.933) or group × time effect (β=−0.25, p=0.51) for location error performance. However, the three-way interaction with memory load was significant (β=0.22, p=0.010): manifest HD showed greater load-dependent increases in location error over 1 year (figure 4).

Swap errors were stable over time in controls (reference group, β=2.17, p=0.180) but increased over time in manifest HD (group × time β=−8.30, p<0.001), with a stronger rise at higher loads (three-way interaction β=3.17, p<0.001) (figure 4).

The nearest-neighbour control measure remained stable across groups, with no significant longitudinal effects (all p>0.05).

Longitudinal imaging measures

Manifest HD exhibited significant 1-year atrophy in the caudate (CE=0.02, p<0.001), putamen (CE=0.009, p=0.006) and hippocampus (CE=0.009, p=0.002). Decline in identification accuracy correlated with caudate atrophy (CE=245.60, p=0.022), whereas increased location error was associated with putamen atrophy (CE=−43.00, p=0.049). No longitudinal changes in cortical thickness were observed, nor were cortical associations with OLT performance.

Discussion

In this study, we investigated the deterioration of distinct components of VSWM in HD and their neural correlates across disease stages. Using the OLT²² cross-sectionally (21 premanifest HD, 71 manifest HD and 39 controls) and longitudinally over 1 year (49 manifest HD and 36 controls), we found that spatial memory impairments, reflected by increased location errors, are detectable in premanifest gene carriers before deficits in object recognition (identification accuracy) or object–location binding (swap errors) emerge. Early impairments correlated with cortical thinning in parietal and middle cingulate regions, suggesting that early striatal–parietal–cingulate atrophy reduces VSWM precision prior to hippocampal involvement. As HD progresses, binding errors (swap errors) increase and are evident in manifest HD, concomitant with hippocampal atrophy. Longitudinal data confirm a progressive emergence of binding deficits. Together, these findings support a hierarchical model of network breakdown in HD, wherein early cortico-striatal disruptions compromise spatial updating and precision, followed by hippocampal-dependent failures in associative memory as the disease progresses.

In the premanifest phase, we observed increased location recall errors and greater nearest-neighbour distances, without a corresponding rise in swap errors. This suggests that early deficits primarily reflect disruptions in attentional control or spatial updating rather than associative binding. Impaired attentional control may compromise the precise encoding and maintenance of spatial information, consistent with observed declines in executive function (SelfCog) and verbal episodic memory (Hopkins Verbal Learning Memory Test) in premanifest participants. These cognitive alterations likely indicate early striatal-linked disruptions in executive and spatial updating processes.^{24 25} Although the OLT minimises verbal mediation, the use of nameable everyday objects may facilitate spontaneous verbal tagging and rehearsal strategies, potentially introducing a verbal working memory component²⁶. This strategy may be inconsistently effective in individuals with subtle language impairments, further contributing to performance variability. For example, Glikmann-Johnston et al ^{4 5} employed both PAL tasks and a virtual house paradigm using concrete objects, while Brandt et al ⁷ used the Hopkins Board with nameable items. The use of such stimuli was motivated by their ecological validity and their ability to enhance task engagement in clinical populations. No explicit instructions were given to suppress verbalisation; future studies employing abstract stimuli may provide a purer measure of VSWM.

Although hippocampal volume in premanifest HD does not differ significantly from controls, we observed significant correlations in HDmc individuals between hippocampal volume and spatial memory measures, including location error and nearest-neighbour distance. This suggests early microstructural or functional hippocampal changes that precede overt atrophy. Binding failures, reflected by swap errors, emerged predominantly in manifest HD; however, we did not observe any association between swap errors and hippocampal atrophy in our cohort, although hippocampal atrophy and disrupted hippocampal–striatal connectivity have been commonly reported in manifest HD.^{7 27}

Further supporting this functional dissociation, manifest HD individuals showed load-dependent increases in swap errors that were independent of nearest-neighbour error, reinforcing the interpretation of binding deficits rather than generalised spatial degradation.²⁸ This pattern reflects a breakdown in the integration of object identity and location, likely resulting from disrupted hippocampal–striatal communication. In contrast, premanifest HD and control participants primarily exhibited reduced spatial precision (indexed by location errors) under load, implicating striatal-dependent spatial updating rather than hippocampal binding. Principal component analysis further corroborated this dissociation: measures of spatial recognition, precision and executive efficiency loaded onto a core VSWM component, whereas swap errors formed a separate component, suggesting a distinct binding process emerging at later stages.

Our findings support a hierarchical model of network breakdown in HD. Early degeneration of striatal and cortical regions impairs spatial location memory, followed by associative binding deficits as hippocampal involvement increases. This pattern aligns with early dysfunction along the dorsal visual stream, including the striatum, inferior parietal cortex and middle cingulate cortex. The inferior parietal lobe, particularly the intraparietal sulcus, is critical for maintaining and manipulating visuospatial information,²⁹ with activity scaling with memory loads.³⁰ Functional MRI studies suggest compensatory parietal recruitment in premanifest individuals nearing symptom onset, reflecting early adaptive changes.³¹ The striatum’s role in VSWM is supported by correlations between striatal activation and task performance in healthy individuals³² and by reduced activation in fronto-parietal-striatal networks in HD under high memory load.¹¹ Reduced middle cingulate thickness is linked to increased object–location errors, consistent with its role in attentional control and error monitoring.^{33 34} Early disruption of cingulate connectivity may contribute to emerging spatial memory deficits in premanifest carriers.^{14 31 35}

Longitudinal data support this trajectory. Over 1 year, manifest HD participants showed significant atrophy in the caudate, putamen and hippocampus. Caudate atrophy correlated with declines in object recognition, while putamen atrophy was associated with increased location error, underscoring their distinct contributions to cognitive function.³⁶ The caudate appears primarily involved in object identification and the putamen in spatial location accuracy. Although the hippocampus also atrophied, this was not directly linked to cognitive performance over the study period. No significant cortical thickness changes were observed between HD participants and controls, nor were cortical changes associated with OLT performance, suggesting that subcortical degeneration, rather than cortical atrophy, predominates in early VSWM decline.

Our findings replicate and extend prior work documenting early object–location memory impairments in HD gene carriers.10 12 37 38 ROC analysis showed that location errors on the OLT significantly discriminate preHD from controls (AUC = 0.731), supporting the sensitivity of continuous spatial recall measures in detecting subtle VSWM deficits early. Consistent with previous studies,^{7 10 13} increasing memory load disproportionately affected gene carriers, highlighting the advantages of load-sensitive paradigms over recognition-based tasks, which may underestimate early deficits.¹¹ Moreover, our data indicate that recognition memory is also impaired in manifest HD, challenging the notion that recall deficits are the primary concern and reflecting limitations of earlier, less sensitive paradigms.^{13 39} The OLT thus provides a nuanced assessment of VSWM, capturing both spatial precision and binding failures.

Clinically, OLT-measured VSWM deficits correlate with established severity scales (cUHDRS, CAP), underscoring their relevance for disease monitoring and caregiver burden.^{37 38} Its ability to detect premanifest deficits suggests potential as a biomarker in trials and prognosis, though larger samples and longer follow-up are needed. A limitation remains the difficulty of fully disentangling attentional from mnemonic contributions in early deficits due to their interplay.⁴⁰ To the best of our knowledge, this is the first implementation of this adapted OLT in a HD cohort. Although this task has been previously used in experimental memory research, formal psychometric properties have not yet been established. Nevertheless, our longitudinal findings provide preliminary support for its potential sensitivity to disease progression in manifest HD.

In conclusion, VSWM deficits emerge early in HD following a hierarchical pattern: striatal and parietal atrophy first impairs spatial precision, followed by hippocampal-related associative memory failures. The OLT offers a sensitive, mechanistically grounded tool for detecting and tracking these changes. To the best of our knowledge, this is the first study to comprehensively map the progressive, hierarchical disruption of distinct VSWM components in HD, linking specific error types to regional brain atrophy across the disease continuum.

Data availability statement

Data are available upon reasonable request. The anonymised datasets of the current study are available from the corresponding author on reasonable request.

Ethics statements

Patient consent for publication

Not applicable.

Ethics approval

This study involves human participants and was approved by French ethics committees approved the protocols (Saint Louis for Repair-HD: NCT03119246, ; and Créteil for Bio-HD: NCT01412125). Participants gave informed consent to participate in the study before taking part.