A home-finding task in virtual reality to test entorhinal path integration performance in patients with mild cognitive impairment

Path integration is a core mechanism of spatial navigation, relying on the continuous updating of self-position based on multisensory cues. The entorhinal cortex—particularly its grid cell network—is central to this process, and is among the first regions affected by Alzheimer’s disease (AD). Path integration deficits may therefore represent a sensitive early behavioral marker of AD-related neurodegeneration. This study examined disease- and age-related alterations in entorhinal cortex–dependent path integration using an adapted home-finding task in immersive virtual reality (VR). Additionally, we investigated the effect of triangle-size scaling on path integration error magnitudes and on the ability to detect group differences. Fifteen patients with MCI, 33 age-matched healthy controls, and 39 young healthy controls completed a VR-based triangle completion task. The virtual environment was created with different triangle sizes (ratio 1:2) without distal and proximal cues to isolate path integration from landmark-based navigation. Participants also underwent neuropsychological assessment and, for patients, AD-specific cerebrospinal fluid biomarker analysis. MCI patients showed significantly greater angular errors, and a tendency toward greater absolute distance errors than older controls. Older controls exhibited larger errors than young controls, correlating with poorer performance on measures of executive function and cognitive flexibility. Absolute distance and proportional linear error increased with greater triangle size, whereas angular error remained unaffected. Group differences in path integration errors did not significantly vary as a function of triangle size. The VR-based path integration task reliably distinguishes between MCI patients and healthy controls, as well as between young and older adults. MCI was primarily associated with impaired rotational accuracy, whereas age-related impairments were evident across all path integration measures. An increase of triangle size amplified distance, but not angular errors, and scaling does not influence the interpretation of group differences. This VR paradigm has potential as a sensitive complementary behavioral marker for the early detection of MCI and AD, and warrants validation in larger, biomarker-confirmed cohorts.

Spatial navigation is a complex cognitive process that enables goal-directed movement. Internal signals such as sensory input, memory, and motor planning are relevant to establish and maintain a trajectory through the environment. These internal signals are combined with external cues such as landmarks and geometric layouts in order to form internal representations in form of a cognitive map (Ekstrom et al., 2014; Garling et al., 1984). Neuroanatomically, medial temporal brain regions configure a complex navigation network for processing spatial information (Ekstrom et al., 2017).

Path integration is a fundamental component of spatial navigation enabling to create an accurate internal representation of space using internal self-motion cues (Mi et al., 1980; Etienne and Jeffery, 2004). The internal self-motion cues allow to continuously update an individual’s position, thereby generating estimates of location and computing a homing vector (Etienne and Jeffery, 2004). The homing vector refers to the shortest vector leading back to the starting point from the current location, calculated based on linear and angular computations of traversed paths (Allen et al., 2004; Segen et al., 2025; Segen et al., 2022). The linear and angular computations result from estimations of traveled distances and rotations mediated by idiothetic signals such as proprioceptive, motor efference copy, vestibular and optic flow (Etienne and Jeffery, 2004; Gibson, 1950; Loomis et al., 1993).

Neurobiologically, the entorhinal cortex, as part of the hippocampal formation, is essential for path integration (McNaughton et al., 2006; Gil et al., 2018; Bierbrauer et al., 2020). Entorhinal grid cells provide a metric presentation of the environment and highly interact with hippocampal place cells resulting in the formation of place fields, which critically contribute to build the cognitive map (Fukawa et al., 2020). Head direction cells code for direction information integrating self-motion signals from vestibular streams (Sargolini et al., 2006). Sensory signals from proprioception and the optic flow complement the entorhinal input by recalibration of path integration, thereby supporting continuous updating of self-position, motion, and the cognitive map (McNaughton et al., 2006; Madhav et al., 2024; Wolbers et al., 2007).

As such a major hub for converging neuronal inputs for path integration, the entorhinal cortex is among the earliest brain regions to exhibit neurodegenerative changes in the context of AD. Hyperphosphorylated tau accumulates in the entorhinal cortex and subsequentially spreads to other medial temporal regions and the neocortex (Braak and Braak, 1991; Berron et al., 2021). In addition to amyloid-associated tau accumulation, amyloid-beta peptide deposition promote neurodegeneration including neuronal loss and gliosis leading to subsequential changes in excitability and synaptic function (Selkoe and Hardy, 2016; Scheltens et al., 2021). These neurodegenerative processes affect specialized spatially-tuned neurons such as hippocampal place cells and entorhinal grid cells (Ekstrom et al., 2003; Nadasdy et al., 2017; Koike et al., 2023) leading to deficits in spatial navigation (Philippen et al., 2023; Coughlan et al., 2018). In the context of path integration, entorhinal grid cell dysfunction has been shown in early AD patients as well as in individuals at genetic risk (Kunz et al., 2015; Howett et al., 2019). In mouse models of AD, animals exhibit amyloid and tau-associated dysfunctional grid cell networks, reduced integration of self-motion cues, that subsequently leads to spatial memory deficits (Koike et al., 2023; Fu et al., 2017; Ying et al., 2022).

Consistent with the progression of neuropathological alterations, deficits in path integration have been observed in individuals with AD, mild cognitive impairment (MCI), and even in at-risk populations presenting with subjective cognitive decline (Segen et al., 2022; Howett et al., 2019; Newton et al., 2024). In this context, virtual reality (VR) triangle completion tasks are particularly well suited for detecting such cognitive deficits (Howett et al., 2019; Newton et al., 2024; Merhav and Wolbers, 2019; Wiener et al., 2011). In these configural path integration paradigms, participants traverse a triangular path and must then choose a homing vector back to the starting location. In the absence of allocentric landmarks, this return cannot be achieved by landmark-based navigation strategies; instead, its computation relies exclusively on the continuous integration of idiothetic self-motion signals accumulated across the entire outbound path—a process that critically depends on entorhinal grid cell function and its interaction with hippocampal place cells (McNaughton et al., 2006; Gil et al., 2018; Bierbrauer et al., 2020). Crucially, unlike continuous path integration tasks in which the homing vector is updated online, triangle completion tasks require an offline spatial representation of the full trajectory and additionally engage working memory resources to maintain this representation until the homing response is executed (Wiener et al., 2011). Howett and colleagues (Howett et al., 2019) employed a virtual triangle-completion task demonstrating that distance estimation is impaired in MCI patients. Distance accuracy further differentiate between biomarker-positive and biomarker-negative MCI patients, indicating a link between entorhinal dysfunction and disease pathology (Howett et al., 2019). Hanyu et al. likewise reported impaired path integration performance in a MCI cohort with deficits that were more pronounced in individuals with positive biomarker status and more rapid clinical progression (Hanyu et al., 2024).

Path integration deficits have also been documented in individuals with an elevated genetic and subjective risk for AD. Among apolipoprotein E 4 (APOE4) carriers—who carry the highest genetic risk for sporadic AD—selective impairments in path integration have been observed behaviorally (Shima et al., 2025). A recent study likewise reported path integration impairment in individuals with subjective cognitive decline, a stage associated with increased risk for future progression to MCI and AD (Segen et al., 2025). Mechanistically, a human neuroimaging study demonstrated reduced grid-cell-like representations in APOE4 carriers (Bierbrauer et al., 2020), consistent with early entorhinal affection. Together, these findings indicate that path integration impairment tracks the progression from genetic and subjective risk through MCI to symptomatic AD, and that the behavioral measure reflects early dysfunction within the entorhinal–hippocampal network.

Besides the influence of AD pathology, VR home-finding tasks have also revealed age-related deficits in path integration (Segen et al., 2022). As path integration depends on the integration of multisensory input, age-related decline in these sensory streams may in turn adversely affect path integration performance (Agrawal et al., 2012; Hsieh et al., 2014; Xie et al., 2017). Increased processing noise of sensory input has been suggested as the primary contributor to diminished grid-like representation and thereby impaired path integration performance in older adults (Bates and Wolbers, 2014; Stangl et al., 2018; Stangl et al., 2020). This contrasts with the more direct and specific disruption of spatial tunes neurons in the hippocampal formation thought to underlie disease-related deficits.

Given that the entorhinal cortex is among the earliest brain structures affected in AD, and that its progressive degeneration directly undermines path integration, behavioral assessment of path integration represents a promising readout of early AD-related neurodegeneration. The clinical relevance of such a tool has grown considerably: the neuropathological cascade underlying AD unfolds years to decades before the emergence of clinical symptoms (Coughlan et al., 2018), and disease-modifying therapies targeting amyloid pathology have recently received regulatory approval—shifting the therapeutic window toward the earliest identifiable disease stages. A cost-effective, ecologically valid paradigm capable of detecting entorhinal dysfunction at this early stage is therefore of considerable clinical and diagnostic relevance.

Various cognitive assessments are available to detect spatial navigation and orientation deficits in healthy and diseased individuals, ranging from neuropsychological tests to VR-based paradigms (Tragantzopoulou and Giannouli, 2024). Especially VR-based paradigms have become increasingly relevant as ecologically valid techniques (Howett et al., 2019; Newton et al., 2024; Merhav and Wolbers, 2019; Wiener et al., 2011). However, their clinical implementation remains challenging, as such tasks often require large testing areas without physical boundaries. The influence of triangle size on the sensitivity of VR-based path integration paradigms therefore warrants systematic investigation.

To date, the influence of triangle size in human path integration tasks has been systematically investigated only once in humans (Harootonian et al., 2020). Harootonian et al. found that path integration errors and their variance increased with longer walked distances, indicative of an increased noise and random fluctuation in self-motion signals (Harootonian et al., 2020). Consistent with these findings, Segen and colleagues recently reported higher path integration errors with increasing path distances (Segen et al., 2025). However, the effects of path or triangle length have not yet been examined in cognitively impaired individuals, nor is it clear to what extent a potential triangle-dependent amplification of path integration errors improves the sensitivity to detect group differences. The availability of distal environmental cues also varies considerably across VR path integration paradigms, a factor that may confound the isolation of path integration from landmark-based navigation.

To address this, we developed a novel VR-based triangle completion task inspired by Howett et al. (Howett et al., 2019), specifically designed to be usable und comprehensible for cognitively impaired individuals. Immersive VR enables the controlled delivery of multimodal sensory inputs—including vestibular, visual and proprioceptive signals—in the absence of real-world navigational constraints. The virtual scene was constructed as an open environment lacking any proximal and distal visual cues minimizing reliance on landmarks and thereby isolating path integration processes. This design aimed to maximize the isolation of path integration from other navigational strategies, while maintaining feasibility and tolerability for participants with cognitive impairment. Additionally, we implemented two experimental conditions with identical path geometry but differing triangle sizes in a 1:2 ratio to examine whether spatial scaling affects path integration error magnitudes and the sensitivity to detect group differences.

The aim of the study was to i) investigate disease- and age-related deficits in path integration performance, and ii) examine the impact of differentially scaled triangle configurations on path integration performance. We hypothesized that disease- and age-related deficits in path integration (Segen et al., 2025; Howett et al., 2019; Newton et al., 2024; Hanyu et al., 2024; Shima et al., 2025; Stangl et al., 2018; Stangl et al., 2020) would be replicated, with disease-related effects expected to be more pronounced in MCI patients, given their early entorhinal cortex dysfunction. Mechanistically, we hypothesized that disease-related path integration deficits reflect neurodegenerative effects on entorhinal grid cell function, whereas age-related effects may represent more indirect influences on grid cell function, for example, due to increased sensory noise.

Based on the assumption of error accumulation over distance (Segen et al., 2025; Harootonian et al., 2020; Cheung et al., 2012), we further predicted that path integration errors would be greater in the large triangle condition and that group differences would be more readily detected.

Participants

We enrolled 92 participants in this study. A group of 15 patients with MCI (75.60 ± 6.74 years, range 66–87 years, 10 female) was recruited from the Memory and Dementia Clinic of the Department of Neurology at the University Hospital Kiel, Germany. All patients were diagnosed according to the current diagnostic criteria defined in the latest version of diagnostic classification (Jack et al., 2018; Jack et al., 2024). To further detect and stratify patients with amnestic type of MCI, the RAVLT delayed recall score was used to indicate hippocampus dependent encoding performance of declarative verbal memory consolidation (Rey, 1941). Exclusion criteria included the history of i) other neurologic or psychiatric diseases including depression, epilepsy or seizures, ii) current intake of any psychiatric medication, or iii) change of regular-medication within the last 3 months. Demographic and clinical characteristics are presented in Table 1 with no significant difference in age between patients with MCI and old controls (p = 0.097). In eight MCI patients AD-typical cerebrospinal fluid (CSF) biomarker were sampled during clinical diagnostic work-up, and included amyloid 1–42 (661 ± 482.72 pg/mL), total tau (479 ± 324.4 pg/mL) and phosphorylated tau (p-tau, 80.98 ± 60.30 pg/mL). Six patients showed characteristic changes of the CSF biomarker status such as reduced amyloid level and elevated p-tau levels.

As control cohorts, we enrolled 39 young healthy controls (24.02 ± 2.90 years, range 20–29 years, 22 female) and 33 old healthy controls (71.96 ± 6.12 years, range 62–83 years, 15 female). In addition to the exclusion criteria for the patient cohort, healthy participants were excluded if they met any criteria for the diagnosis of MCI or dementia defined by a history of memory decline within the past 6 months as well as an objectively measured cognitive score below age-normative thresholds in the neuropsychological assessment (see chapter 3.3.). Three participants initially assigned to the older control group met neuropsychological criteria for MCI, as indicated by a RAVLT delayed recall score below age-normative thresholds. Their data were thus reclassified post hoc into the MCI group. Controls were recruited from Kiel University, sports clubs and senior citizen groups. The sample size was not determined by a formal a priori power calculation, but was guided by comparable published studies employing VR-based path integration paradigms in clinical MCI cohorts (Howett et al., 2019; Hanyu et al., 2024).

The study was approved by the Ethics Committee of Kiel University. All participants gave their written informed consent prior to the experiment. All procedures were in accordance with the Declaration of Helsinki.

Path-integration test in virtual reality

Design and implementation

We developed an adapted immersive version of an established path integration task (Howett et al., 2019; Merhav and Wolbers, 2019) running on the virtual reality headset Oculus Quest 2 (Oculus, Irvine, CA, United States) using the Unity game engine (version 2019.4). The small and lightweight VR device was chosen to allow the users to move independently with slight spatial limitation.

The path integration task is a home-finding-task in immersive VR to quantify entorhinal dependent path integration performance. Participants entered an open virtual space without any visual borders. Neither distal nor proximal landmarks were presented, and the virtual scene featured a uniform texture without salient visual cues, thereby minimizing the possibility of landmark-based navigation (Figures 1c–f). A green grass-like surface and a light blue sky were chosen for a neutral environment. Movements in the virtual world corresponded 1:1 to the real world in order to minimize VR-related side effects.

The VR environment was implemented in two conditions with varying triangle sizes (ratio 1:2), while the overall environmental features and path configuration remained identical. The total field size were 4 × 8 and 8 × 16 m and the test was performed in a sports hall/gymnasium (Figures 1b–d). The triangle size condition in which the participants started was randomized. In order to exclude any additional visual cues for the participants, we deactivated the visibility of the guardian of the VR device which constitutes the virtual boarders of the determined field size. To avoid collisions with physical surroundings, investigators provided verbal warning to the participants in rare instance when participants approached the physical limits of the testing area. Such interventions occurred in 0.5% of all trials across all participants. Participants were equipped with a virtual reality headset that eliminated perception of the physical surroundings, and wore noise-cancelling headphones for the delivery of auditory instructions and to avoid external auditive cues. A handheld controller was used to indicate the target position in the paradigm.

Experimental set-up

Training phase

A training trial was conducted at the beginning of the experiment to familiarize participants with the virtual environment and the underlying principles of the home-finding task. During the training trial, participants received instructions via a standardized audio protocol delivered through the VR headset. Participants were instructed to locate the starting position within the VR environment indicated by a white circle on the ground (Figures 1a,e). To facilitate orientation, additional white arrows and a ground-level line were provided. Upon reaching the starting point, participants were guided to consecutively navigate to three targets indicated by virtual globes labeled ‘A’, ‘B’ and ‘C’. Each globe and its corresponding label disappeared upon the participant’s approach. After reaching the final target location ‘C’, participants were asked to return to the location of the initial target position ‘A’ (homecoming) and to confirm their arrival by pressing the A-button on the handheld controller (log, Figure 1a).

Prior to the training trial, participants spend a short period freely moving in the virtual scene in order to adapt for the immersive VR setting.

Test phase

After the training trial, all participants completed a test series consisting of eight different paths with differing architecture following the same procedure as described above (Figure 2, paths I-VIII). During the test phase, no audio-feedback or instruction was provided.

In case of any disruption or discomfort of the participant during the experiment, the investigator paused the test and restarted the unfinished path after recovery.

After completing the VR path integration task, participants gave feedback about sensory phenomena experienced during the experiment.

Data acquisition

Tracking data were recorded on the VR device at a frequency of 15–20 Hz during each trial in a three dimension coordination system including the participant’s position (x, z coordinates) and rotation (yaw). Data recording for each trial started when participants were instructed to move to the starting position and concluded by pressing the controller button to confirm their estimated arrival at the home location (‘log’, see Figure 1a). To consider the different architecture of the tests paths including varying triangle lengths across the eight paths, data were averaged for each triangle size condition, respectively.

Error computation

Analogue to Howett and colleagues (Howett et al., 2019), path integration performance was measured by computing i) the absolute distance error, indicating the distance from the logged home position to the real home location (Euclidean distance from log to A, see Figure 1a), ii) the proportional linear error quantifying the accuracy of distance estimation without taking angular deviation into account, calculated as the absolute value of the difference between the estimated (logged) home vector (Position C to log, Figure 1a) and the ideal home vector (Position C to A) using the formula x = |estimated home vector–ideal home vector|, and iii) the angular error quantified as the absolute angular deviation between the ideal home vector and the logged home vector, representing rotational accuracy (Figure 1a).

Subjective assessment of VR environment

All participants completed a questionnaire about experienced side effects during the virtual reality task that included experience of discomfort including motion sickness and dizziness and fear of collision. The 13-item questionnaire was constructed with a 6-point Likert scale ranging from 1 to 6, where 1 = strongly disagree, 2 = disagree, 3 = somewhat disagree, 4 = somewhat agree, 5 = agree, and 6 = strongly agree. For analysis, responses of 1–3 were categorized as disagreement, and responses of 4–6 were categorized as agreement. Further, participants were asked to rate their experience concerning comprehensibility and feasibility of the task.

Neuropsychological assessment

All participants completed a neurocognitive assessment to measure secondary outcome variables to characterize the cognitive profile of the MCI group, and to exclude cognitive impairment in the control groups. The utilized assessments were chosen to detect hippocampal-dependent memory deficits, typically associated with MACI according the NIA-AA and CERAD criteria (Jack et al., 2018; Morris et al., 1988). The assessment included a subset of the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) (Morris et al., 1988; Schmid et al., 2014; Wolfsgruber et al., 2014): the Rey auditory learning test (RAVLT (Rey, 1941). The Trail-Making-Test (TMT) A and B was included to assess visual attention, cognitive processing speed and cognitive flexibility (Reitan, 1979), followed by the German Regensburger Wortflüssigkeitstest (RWT) to evaluate speech fluency (Aschenbrenner et al., 2000). Working memory was evaluated by using the Digit Span as part of the Wechsler Adult Intelligence Scale (Wechsler, 1997). Selective attention and cognitive flexibility was defined using the Stroop Test (Stroop, 1935). We used the Rey-Osterrieth Complex Figure (ROCF) to quantify visuoconstructive performance and working memory function as well as nonverbal visuo-spatial memory (Kasai et al., 2006).

Statistical analysis

Recorded data from the Oculus quest were processed and analyzed post hoc on a standard desktop computer using Microsoft Excel for data aggregation and R statistics (version 4.4.2 (R Core Team, 2021)) for statistical analysis. All VR path integration error metrics were averaged across paths per triangle size, yielding three error measures per triangle size. Normality of data was assessed prior to modeling using Shapiro-Wilk test.

Hypothesis testing was computed using linear mixed-effects models using the R package lme4. Although the raw error data were non-normally distributed, visual inspection revealed that the model residuals were approximately normally distributed and supported the use of these models. These models allowed evaluation of effects of group and triangle size and their interaction effects (grouptriangle size) on each error measure respectively, while accounting for dependencies due to repeated measurements or nested data structures. To further quantify evidence for the presence versus absence of grouptriangle size interaction, we additionally conducted a Bayesian model comparison between a model including and a model excluding the interaction term. Bayes factors (BF 01) were computed using a BIC-based approximation to quantify evidence in favor of the model without the interaction. For significant main effects, we calculated post hoc pairwise contrasts using estimated marginal means with p-values adjusted using the Holm procedure. Estimates are reported as the difference between groups in the order stated; negative values therefore indicate higher errors in the second groups. Confidence intervals were obtained using simultaneous inference procedures as implemented in the R package emmeans. Because multiplicity-adjusted confidence intervals and adjusted p-values are not necessarily based on identical inferential procedures, small discrepancies between statistical significance and confidence interval coverage may occur (Lenth and Piaskowski, 2026). Effect sizes for pairwise contrasts were derived from estimated marginal means using the residual variance of the linear mixed-effects model for standardization. For omnibus tests, partial eta-squared (η 2 p) was calculated as a measure of effect size using the R package effectsize.

To assess model robustness of the primary model against potential heteroscedasticity, the model was refitted as a linear mixed-effects model with group-specific residual variances (varIdent structure). This sensitivity analysis examined whether the pattern of fixed effects was unchanged under relaxed variance assumptions.

To assess whether neuropsychological performance—especially memory consolidation, executive and visuospatial functions—moderated group differences in path integration performance, we extended the main linear mixed-effects model by including neuropsychological performance as a covariate, with group and triangle size as moderators. For investigating association of path integration with AD pathology, the main model was extended by including CSF biomarker.

Data are specified as mean ± standard deviation (SD) if not otherwise stated. A two-sided significance level of p < 0.05 was considered statistically significant.

Path integration performance

Behavioral result

Young and old controls completed all paths in both triangle size conditions. One patient in the MCI cohort only completed the small triangle paths. A linear mixed-effects model was conducted to examine the effects of group (young vs. old vs. MCI), triangle size (small vs. large), and their interaction with participant as a random effect on each error measure, respectively.

The group*triangle size interaction on absolute distance error was not significant (F (2,84) = 0.54, p = 0.582, η 2 p = 0.01). A Bayesian model comparison further indicated that the data favored the model without the interaction term (BF 01 = 100), providing strong evidence that the effect of triangle size did not differ between groups. A significant main effect of group was observed (F (2,84) = 18.92, p < 0.001, η 2 p = 0.31). Holm-adjusted pairwise comparisons based on estimated marginal means indicated that old controls made slightly lower absolute distance errors than MCI patients (mean difference = −0.66, 95% CI [-1.39, 0.08], d = −0.61, p = 0.03, Figure 3a). While the Holm-adjusted p-value suggested a statistically significant group difference, the simultaneously adjusted confidence interval included zero. Given this inconsistency between inferential criteria, the result should be interpreted as providing only limited evidence for a difference between groups. Young controls exhibited lower error magnitude than old controls (mean difference = −0.99, 95% CI [-1.55, −0.44], d = −0.99, p < 0.001, Figure 3a). In addition, the linear mixed model revealed a significant main effect of triangle size on absolute distance error (F (1,84) = 151.68, p < 0.001, η 2 p = 0.64). Holm-adjusted pairwise comparisons of estimated marginal means indicated that small triangles resulted in lower absolute distance errors (mean difference = −2.05, 95% CI [-2.38, −1.72], d = −2.03, p < 0.001, Figure 3d).

Regarding the proportional linear error, the group* triangle size interaction was not statistically significant (F (2,84) = 1.39, p = 0.25, η 2 p = 0.03). A Bayesian model comparison indicated that the data favored the model without the interaction term (BF 01 = 42), indicating that the effect of triangle size did not meaningfully affect group discrimination. The main effect of group was significant (F (2,84) = 7.20, p = 0.001, η 2 p = 0.15), with post hoc comparisons revealing no significant difference between old controls and MCI patients (mean difference = −0.16, 95% CI [-0.83, 0.51], d = −0.21, p = 0.55, Figure 3b). Young controls exhibited lower proportional linear errors than old controls (mean difference = −0.66, 95% CI [-1.17, −0.15], d = −0.86, p = 0.006, Figure 3b). Furthermore, a significant main effect of triangle size was observed (F (1,84) = 137.56, p < 0.001, η 2 p = 0.62). Consequently, pairwise comparisons revealed lower proportional linear errors in the small triangle condition compared to the large triangle condition (mean difference = −1.49, 95% CI [-1.74, −1.24], d = −1.94, p < 0.001, Figure 3e).

Concerning the angular error, no significant interaction effect was observed (F (2,84) = 0.96, p = 0.386, η 2 p = 0.02). Again, the Bayesian model comparison revealed that the model without the interaction was strongly favored by the data (BF 01 = 67), providing evidence that the triangle size did not affect group discrimination. A significant main effect for group was shown (F (2,84) = 16.82, p < 0.001, η 2 p = 0.29). Post hoc contrasts indicate that old controls made significant lower angular errors than MCI patients (mean difference = −6.58, 95% CI [-12.5, −0.64], d = −0.85, p = 0.008, Figure 3c). Young controls exhibited lower error magnitude than old controls (mean difference = −6.59, 95% CI [-11.1, −2.07], d = −0.86, p = 0.001, Figure 3c). No significant main effect of triangle size was observed on angular error (F (1,84) = 0.17, p = 0.685, η 2 p = 0.002, p =0.69, Figure 3f). Detailed statistical results are provided in Supplementary Table S1).

To visualize an exemplary performance on a MCI patient across all paths of the home-finding task, we computed radar charts illustrating the deviation of the absolute distance error from the error mean of the old controls group, exemplary for one patient (Figure 4). An error greater than a standard deviation from the group mean of the old control group was defined as a relevantly worse result.

Model robustness analysis

To assess the robustness of the primary findings against potential heteroscedasticity, we conducted a sensitivity analysis in which the main linear mixed-effects models were refitted allowing for group-specific residual variances. Figure 5 presents the raw data distributions by group for each triangle size condition. The pattern of fixed effects remained unchanged for absolute distance, proportional linear, and angular error, with no meaningful differences in the significance or direction of effects compared to the primary models (see Supplementary Table S2). This pattern indicate that the results were robust to violations of the homogeneity of variance assumption.

Neuropsychological assessment

Neuropsychological data were presented in Table 1, illustrating young controls reaching significantly higher score in all test in contrast to old controls. With regard to the MCI patients, they notably show lower scores in RAVLT and copy and delayed recall tests of the ROCF, TMT A and B, word fluency and digit span, reflecting typical impairment of episodic memory in MCI.

Neuropsychological or biological correlates of path integration performance

To evaluate influence of neuropsychological test performance on path integration, we conducted a linear mixed-effects model including neuropsychological test performance as covariates and group and triangle size as moderators for each error respectively.

The models revealed significant interactions between TMT-B and group for all error measures (absolute distance error: F (2,81) = 6.33, p = 0.003; proportional linear error: F (2,81) = 13.57, p < 0.001; angular error: F (2,81) = 8.28, p < 0.001), indicating that the association between TMT-B performance and errors differed between groups.

Follow-up comparisons of TMT-B slopes (see Supplementary Table S3) showed that the strength of this association differed between old controls and MCI patients (absolute distance error: β_diff = 0.02, SE = 0.005, t = 3.49, p = 0.002; proportional linear error: β_diff = 0.02, SE = 0.004, t = 5.2, p < 0.001; angular error: β_diff = 0.16, SE = 0.04, t = 4.06, p = 0.0003), with a stronger association in the old control group (Figures 6a,c,e).

In addition, significant interactions between TMT-B and triangle size were observed for proportional and angular errors (proportional linear error: F (1,85) = 9.16, p = 0.003; angular error: F (1,85) = 3.96, p = 0.049). Follow-up analyses indicated that the association between TMT-B and proportional linear error was stronger in the large triangle condition (β_diff = 0.006, SE = 0.002, t = −3.03, p = 0.003), whereas the association with angular error was stronger in the small triangle condition (β_diff = 0.04, SE = 0.02, t = 1.99, p = 0.049; Figures 6a,c,e).

Another model including ROCF delay showed significant interactions between ROCF delay and group for all error measures (absolute distance error: F (2,81) = 5.1, p = 0.008; proportional linear error: F (2,81) = 3.76, p < 0.028; angular error: F (2,81) = 5.25, p < 0.007), indicating that the association between ROCF delay performance and errors differed across groups. Concerning absolute distance and angular error, follow-up comparisons of ROCF delay (see Supplementary Table S4) showed that the strength of this association differed between old controls and MCI patients (absolute distance error: β_diff = −0.11, SE = 0.04, t = −2.5, p = 0.04; angular error: β_diff = −0.98, SE = 0.36, t = −2.75, p = 0.02), with a stronger association in the MCI group. Furthermore the association differed strongly between young and old controls (absolute distance error: β_diff = 0.11, SE = 0.05, t = 2.52, p = 0.04; angular error: β_diff = 0.85, SE = 0.37, t = 2.31, p = 0.0047), indicating a stronger association in the young control group (Figures 6b,d,f). For proportional linear error, follow-up comparisons of ROCF delay slopes however did not reach statistical significance after correction for multiple comparisons (all p’s > 0.05). The interaction effect between ROCF delay and triangle size was not significant for all error measures (all p’s > 0.15). A model including RAVTL delayed recall reflecting episodic verbal memory consolidation did not reveal any significant interaction effects indicating no associations with path integration performance (all p’s > 0.06).

In addition to the neuropsychological assessment, we performed linear mixed-effects models including CSF biomarkers such as ß-Amyloid 1–42, tau and phospho-tau and triangle size as moderator for each error respectively. Results revealed no significant effect of all CSF biomarker on path integration errors (all p’s > 0.57).

Subjective assessment of VR technology

The virtual reality paradigm was well tolerated across all groups with no participant discontinuing the task due to side effects of the virtual setting. No patient or participant in the older control group reported relevant discomfort or fear of collision during the VR task, with no ratings of 4–6 on the Likert scale. In the young cohort, only one participant (2.6%) reported discomfort and fear of collision (Likert scale 4). In contrast, 38% of the patients, 33% of the old controls and 48.6% of the young controls reported a prior history of motion sickness, for example, during journeys. Overall, participants described the path integration task as clearly explained, understandable and feasible, only one participant in the young control group evaluated difficulties in understanding the task procedure.

In the present study, we implemented an adapted virtual reality–based home-finding task to investigate disease-, age-, and scaling-related alterations in path integration performance as a behavioral correlate of entorhinal cortex function. First, individuals with MCI showed significantly greater angular errors and a tendency toward higher absolute distance errors compared with cognitively healthy old controls, suggesting disease-related impairments in path integration performance. Second, consistent with our hypothesis of age-related differences, the paradigm successfully distinguished between healthy young and old adults across all error measures, indicating sensitivity to normative aging effects on spatial navigation. Third, as expected, distance accuracy decreased with increased triangle size, whereas rotational accuracy remained unaffected. Although we hypothesized that larger triangle configurations would amplify performance differences, no statistically significant scaling effect on group discrimination was observed.

Disease-related impairment of path integration

In our virtual path integration task, patients with MCI exhibited significantly greater angular errors, and showed a tendency toward higher absolute distance errors compared with old controls, although the latter was not consistently supported across inferential criteria. Consequently, the present findings provide evidence for a disease-related impairment in angular estimation, but only limited support for an impairment in absolute distance estimation in patients with MCI. Our results parallel previous reports of path integration deficits in AD and MCI patients (Segen et al., 2022), which has been observed in VR settings (Segen et al., 2022; Howett et al., 2019; Hanyu et al., 2024) and real-world conditions (Mokrisova et al., 2016). Our results further indicate that angular error emerged as a more sensitive marker for impaired path integration in MCI, which is consistent with previous findings that propose angular error as a sensitive indicator for early AD-related path integration deficits (Bierbrauer et al., 2020; Newton et al., 2024; Colmant et al., 2023). Other studies have likewise reported distance errors as a marker of disease-specific path integration impairments (Howett et al., 2019), which we also observed, but with limited statistical support. The weaker effect for absolute distance error may be attributable to limited statistical power in the present sample, and to the interdependence of error measures in configural path integration tasks, where inaccuracies in angular estimation may propagate to distance estimates (Segen et al., 2022; Wiener et al., 2011). This interpretation is supported by the absence of group differences in proportional linear error, which reduces the contribution of angular deviations to distance estimation. Together, this suggests a greater role of rotational inaccuracies in the MCI group.

As path integration deficits have been observed not only in the clinical stages of AD, but also in preclinical phases and in individuals at increased genetic risk—such as those with subjective cognitive decline (SCD) and APOE ε4 carriers (Segen et al., 2025; Bierbrauer et al., 2020; Shima et al., 2025)—this pattern may suggest a potential association with early pathophysiological processes. Accordingly, Shima et al. reported associations of impaired path integration performance with higher levels of plasma phospho-tau 181 and decreased thickness of the entorhinal cortex (Shima et al., 2025). This result is further supported by findings in AD mouse models showing deficits in path integration as a result of tau accumulation in the entorhinal cortex (Koike et al., 2023). In our study, we did not observe associations of path integration performance and CSF biomarkers in the MCI cohort. This is most likely due to the limited sample size of available biomarkers and the heterogeneous CSF profile of the MCI group, which needs to be considered as a study limitation. Considering the heterogeneity of MCI phenotypes, we adopted a clinically oriented stratification approach: MCI patients in our cohort showed significantly lower scores on the RAVLT and ROCF, indicating impairments in the encoding and consolidation of declarative verbal and visuoconstructive memory. These deficits may plausibly be related to tau-associated functional alterations within hippocampal–entorhinal networks and thus possibly represent a behavioral correlate of early tau burden. In line with this interpretation, ROCF performance was associated with absolute distance and angular errors in the MCI group, suggesting that path integration performance is associated with visuoconstructive competence, which here may also be the reflection of behavioral manifestations of early tau pathology.

Moreover, MCI patients in our cohort showed lower TMT-B performance than older controls. Although the TMT-B is traditionally interpreted as a measure of executive function and cognitive flexibility, its sensitivity in MCI and early AD may as well reflect disease-related processes that extend well beyond frontal-executive networks. Notably, Hildebrandt and colleagues recently demonstrated that TMT-B performance correlates with total tau (t-tau) levels specifically in amyloid-positive AD patients, but not in amyloid-negative individuals (Hildebrandt et al., 2026). This finding suggests that, in the context of confirmed AD pathology, TMT-B may serve as a clinically accessible indicator of tau-related neurodegeneration and may help to identify patients at risk of faster disease progression. The specificity of this association to the amyloid-positive group implies that the sensitivity of TMT-B in early AD reflects not merely executive inefficiency within frontal circuits, but also the progressive spread of tau pathology to networks beyond the entorhinal cortex. Converging evidence from brain perfusion studies further links poorer TMT-B performance to reduced cerebral blood flow in posterior parietal regions in patients with mild AD and amnestic MCI (Terada et al., 2013; Shindo et al., 2013) — areas that overlap anatomically with the navigational network supporting path integration. Within this framework, the lower TMT-B performance and the group-specific association between TMT-B performance and path integration errors observed in our MCI cohort may not only reflect executive dysfunction, but may also indicate tau-associated neurodegeneration within parietal nodes of the entorhinal–hippocampal navigational network. Nevertheless, a contribution of working memory processes to path integration performance in the configural path integration task used here cannot be ruled out. In contrast to configural path integration, continuous path integration tasks are considered more sensitive to error accumulation and do not require an offline spatial representation (Segen et al., 2025; Wiener et al., 2011; Stangl et al., 2018; Stangl et al., 2020). Despite this, we used the triangle completion task as a robust and reliable task for investigating path integration performance in cognitively impaired individuals. By designing the VR home-finding task without distal landmarks or boundaries, participants were required to rely preferentially on egocentric reference frames using a configural path integration strategy. As evaluated above, this might target entorhinal cortex–mediated path integration mechanisms. Future studies incorporating functional imaging and extended neuropsychological test batteries assessing spatial working memory processes, e.g., the Corsi Block Tapping Test or spatial working memory that directly capture active spatial working memory demands (Zimmer, 2008), may help to disentangle the relative contribution of entorhinal-hippocampal circuit dysfunction, executive control and spatial working memory to path integration performance in patients with MCI.

In summary, the observed impairment of path integration performance among patients with MCI may reflect entorhinal cortex dysfunction and corroborate previous findings of entorhinal cortex-related navigational deficits, although interpretations should be made with caution due to the small sample size and therefore limited statistical power. Because of that, potential sex differences of path integration in our MCI cohort could not be elucidated. However, given the propagation and accumulation of tau affecting the entorhinal cortex in early disease AD stages, an impairment of path integration performance thus possibly reflects a stronger and specific structure-function relationship in AD and might show a greater sensitivity over global neuropsychological testing scores.

Age-related changes in path integration

With regard to age-related path integration, our VR paradigm revealed greater error magnitudes—across absolute distance, proportional linear, and angular errors—in old healthy controls compared to young healthy controls, which indicates an age-related decline of entorhinal cortex-dependent path integration performance. In contrast to our initial hypothesis, these effects were more pronounced than the disease-related deficits. Overall, these results corroborate previous findings about path integration deficits in aging adults, including a decrease in distance and angular accuracy (Segen et al., 2022).

Previous work has suggested that path integration errors may, at least in part, be related to background noise in sensory input, which is present in both young and older adults and may increase with age (Stangl et al., 2018; Stangl et al., 2020; Harootonian et al., 2020; Lappe et al., 2007). This in turn seems to contribute to diminished grid-like representations and thereby impair path integration performance in older adults (Bates and Wolbers, 2014; Stangl et al., 2018; Stangl et al., 2020). In addition to the increased background noise, age-related degradation of sensory input may contribute to deficits in path integration. As path integration relies on the integration of optic flow, motor efference copy, vestibular, and proprioceptive signals to continuously update self-position (Segen et al., 2022; Gibson, 1950; Anastasiou et al., 2023), reduced quality of these inputs may impair performance. Multisensory integration appears particularly important for successful path integration in older adults, making sensory degradation a likely contributor to age-related path integration impairment (Adamo et al., 2012; Moffat, 2009a; Lester et al., 2017). The implementation of our paradigm in VR allowed the integration of multiple sensory modalities resembling real-world conditions for path integration. Consequently, potential differences in the preferential use of specific sensory inputs and their implications on path integration performance were not examined but may be addressed in future studies. By excluding visual landmark cues, the reliance on landmark-based navigation strategies is avoided and thereby enables a more ecologically valid assessment of path integration performance.

In addition, evidence suggests that aging is associated with impairments in the downstream neural processing of path integration, particularly within the entorhinal cortex (Lester et al., 2017; Moffat, 2009b). Age-related dysfunction is also observed in the hippocampus, a structure which is critically involved in spatial navigation and closely interconnected with the entorhinal cortex (Moser et al., 2008). Age-related decline in hippocampal function has been linked to impairments in place learning and the generation and maintenance of internal cognitive maps (Stangl et al., 2020). This is particularly relevant in the context of the present study, as the configural path integration task required participants to maintain an internal spatial representation in the form of a cognitive map (Wiener et al., 2011). Thus, the present age-related deficits in path integration may reflect a deficient hippocampal-entorhinal connectivity. On the other hand, in the configural path integration, the maintenance of an internal spatial representation in form of a cognitive map does not rely solely on entorhinal–hippocampal networks, but also requires executive functions such as working memory and cognitive flexibility. The potential relevance of these processes is supported by the observed association between TMT-B performance and path integration errors, which was particularly pronounced in the old control cohort. In contrast to young controls, old controls display a lower performance on TMT-B, suggestive of age-related deficits in executive functions (Lacreuse et al., 2020; Kirova et al., 2015). These findings suggest that age-related decline in executive functions may at least partly contribute to the maintenance and updating of internal spatial representations, thereby influencing the accuracy of path integration in older adults. Importantly, as discussed in the context of disease-related impairments, poorer TMT-B performance has been linked to reduced cerebral blood flow in posterior parietal regions in patients with AD and MCI (Terada et al., 2013; Shindo et al., 2013). This raises the possibility that the association between TMT-B and path integration errors in older adults may not solely reflect frontal-executive decline, but may also indicate age-related impairment within parietal nodes of the navigational network. Furthermore, given that TMT-B performance has recently been shown to correlate specifically with total tau levels in amyloid-positive patients (Hildebrandt et al., 2026), subclinical AD pathology contributing to the observed TMT-B associations in the older control group cannot be entirely excluded—particularly in the absence of biomarker data for this cohort.

This lack of biomarker confirmation should therefore be considered an important limitation of this study. Accordingly, the absence of biomarker confirmation precludes any direct assessment of underlying age-related tauopathy or early AD-related neurodegeneration (Scheltens et al., 2021). As a result, it cannot be determined whether such processes may have contributed to the observed path integration deficits in the old control group, where stronger effects were observed compared to the MCI group. In addition to the CSF biomarkers used in the present cohort, blood-based biomarkers reflecting amyloid and tau pathology are now available, in particular plasma p-tau 217 (Teunissen et al., 2025). Associations between path integration performance and plasma p-tau levels have been reported previously, indicating that VR-based path integration may serve as a non-invasive surrogate for early AD neuropathology (Shima et al., 2025). Given the cost- and time-effective nature of the VR paradigm, incorporating plasma biomarkers in both patients and old controls in future studies would help to disentangle age-related from disease-related contributions to path integration deficits. An additional future direction concerns the potential for multisensory augmentation of spatial learning. Multisensory approaches, such as brief music listening, may transiently enhance visuospatial processing (Giannouli et al., 2024) and could be explored to support spatial learning in VR-based path integration paradigms.

Taken together, our results indicate age-related impairments of path integration reflected by distance and angular inaccuracy. Mechanistically, in contrast to disease-related deficits, age-related path integration impairments likely reflect a combination of more indirect influences on entorhinal–hippocampal network function. Contributing factors may include increased afferent processing noise and age-related degradation of multisensory inputs, disruption of hippocampal–entorhinal network connectivity, and the contribution of executive and posterior-parietal processes to the maintenance of internal spatial representations during configural path integration.

Effects of triangle size scaling on path integration

We further investigated scaling effects of triangle size on path integration, hypothesizing that errors would accumulate with increasing triangle size. In line with this, our results show higher magnitudes of absolute distance and proportional linear errors on large triangles. In contrast, no difference in angular errors was observed across triangle sizes, indicating that rotational accuracy remains unaffected by triangle scaling. These results corroborate previous findings showing that path integration errors increase as a function of path length (Segen et al., 2025; Harootonian et al., 2020). Specifically, Harootonian et al. likewise observed changes only in distance errors following manipulation of triangle size. The authors attributed that to unchanged angles, as only triangle length varied. As a potential explanation, they observed an increase in error variance as a function of triangle size, potentially reflecting increasing noise of sensory input when walking longer distances (Harootonian et al., 2020). Consistent with this, our results show stable angular errors across triangle sizes. In our paradigm, the triangle size was varied solely by scaling path length in a 2:1 ratio while turning angles were not directly manipulated. Although there is evidence for an interrelation of walked distances on angular error (Segen et al., 2022; Wiener et al., 2011), the manipulation in our triangle-completion task may not have been sufficient to detect distance-dependent changes in angular accuracy. In contrast to our findings, Segen and colleagues reported that both distance and angular accuracy decrease with longer distance traveled (Segen et al., 2025). This contrast may be attributed to the different path integration strategies as Segen and colleagues used a continuous path integration task in contrast to the configural path integration strategy required for our triangle completion task.

In contrast to our hypothesis, the linear mixed model revealed no significant interaction between group and triangle size. Importantly, the Bayes factor provided strong evidence in favor of the null hypothesis, indicating that the influence of triangle size on path integration performance was comparable across groups and therefore did not enhance group discrimination. A potential explanation is that both healthy aging and AD involve alterations in hippocampal–entorhinal circuitry, although the underlying mechanisms differ substantially, as AD involves more specific neurodegenerative changes within this network (Segen et al., 2022). Therefore, path integration accuracy may be comparably impaired across triangle sizes in both groups, rendering a multiplicative amplification of group differences unlikely regardless of triangle size. This scale-invariance of impairment offers a mechanistic account for the absence of a significant group × triangle size interaction: the entorhinal–hippocampal system re-scales its spatial representations flexibly across environmental scales, but disease- and age-related degradation of these representations appears to be similarly expressed at both scales. These interpretations should be investigated in future studies incorporating higher scaling ratio of triangles sizes, larger sample size, and functional neuroimaging to disentangle the specific circuit-level contributions to path integration in aging and AD.

Overall, our results show an increased magnitude of distance-related error measures with larger triangle size, but the scaling did not amplify disease- or age-related group differences, suggesting that a small-scale testing environment with relatively small triangle sizes may be sufficient to capture disease- and age-related deficits in path integration. This has important practical implications as smaller testing spaces may facilitate the clinical implementation of the paradigm, and also improve its cost- and effort-efficiency.

Together, our home-finding task in an immersive VR environment offers several advantages: i) it is cost- and effort-efficient, ii) it provides sensory input that closely resembles real-world navigation by the integration of visual, vestibular, and proprioceptive information into the VR environment, iii) test conditions can be adapted to target specific aspects of path integration and navigational strategies (for example, by manipulating the presence of distal or proximal cues or triangle size), and iv) it can be readily adapted to the needs and capabilities of cognitively impaired individuals due to its user-friendly design. This is supported by the absence of reported discomfort among older and patient participants indicating that the VR paradigm was safe and well tolerated.

In conclusion, our findings indicate age- and disease-related alterations in path integration. In contrast to our initial hypothesis, group differences appeared more robust across all error measures in the old control group, whereas in the patient group they were primarily reflected in angular inaccuracy, and were less consistent for distance accuracy.

Although the magnitude of distance-related error measures increased with larger triangle size, the scaling did not amplify disease- or age-related group differences in path integration performance, suggesting that small-scaled environments sufficiently capture path integration deficits. The small sample size of the MCI cohort derived from a single center as well as the absence of CSF biomarker in the old controls needs to be considered as limitations. Moreover, although the higher proportion of women in the MCI group reflects the increased prevalence of AD among women (Scheltens et al., 2021), a potential confounding influence in the form of sex-dependent differences in path integration cannot be ruled out and should be addressed in future studies. Considering these limitations, the paradigm show promise as a sensitive complementary behavioral marker for the early detection of MCI and prodromal AD. Moreover, it may serve as a promising tool for longitudinal assessment throughout the course of the disease as well over the course of novel disease-modifying therapies, enabling the monitoring of progression and the conversion to more advanced clinical stages. To further explore the diagnostic utility of our VR path integration task, future studies with larger, multicenter and diverse samples, including participants across different disease stages, are needed to establish the generalizability of the findings.