The battery we administered comprised memory, language, attention, executive function, and visuospatial domains. The full set of neuropsychological tests and descriptions is reported in Multimedia Appendix 1.

We used a VR landmark- and boundary-based spatial memory recall similar to that described by Tuena et al [38]. Participants were seated in a comfortable position at a distance of 50 cm from the computer screen. In this task, a 3dRudder and a VR-ready computer (15.6-inch MSI GF63 Thin 11UC Intel Core i7-11800H, NVIDIA GeForce RTX 3050 4 GB, 16 GB RAM, and 512 GB SSD) were used to interact with the virtual environment. The 3dRudder is a circular platform designed for use while seated in a comfortable and safe position. To operate it, the user places the feet on top of the device, which is mounted on a semispherical base. This design enables intuitive movement control by tilting the feet in the desired direction or by pushing the toes or heels to move forward or backward. The device is equipped with inertial and pressure sensors that detect the user’s movements and convert them into virtual actions. The use of a motion foot pad instead of a classic joypad is intended to recruit lower-limb motor commands instead of hand and finger motor efferent copies, which is closer to real-world navigation [39].

The structure of the spatial memory task was as follows. During the encoding phase, participants were instructed to collect 4 items at a time and encode their locations within a circular arena (with a diameter of 50 virtual meters). Critically, during encoding, both the arena boundary (wall) and a fixed intra-arena landmark (obelisk; coordinates [−10.5, 10.5]) were simultaneously visible, allowing participants to encode object locations using both cues.

Items were presented randomly, and each object was collected 4 times in a random order. The 4 objects were the same for all the participants and were presented at the same fixed coordinates (cat [−5, −14], chair [−4, 20], bike [18, 10], and carrot [5, 17]). To visualize the object, participants had to navigate to its exact location. Upon reaching it, the object disappeared, and they proceeded to find the next one. The starting position (coordinates [0, 0]) of this phase was always the center of the circular arena, facing an arbitrary north.

During the immediate recall phase, participants had to remember and navigate to the exact location where the item (cued at the bottom of the screen) had previously been collected and then press a button on the keyboard to respond. The next object was then presented. Either the wall or the obelisk was randomly removed, forcing participants to rely primarily on one type of spatial cue. In the allocentric (wall) condition, the obelisk was removed, but the visible wall remained, encouraging boundary-based spatial memory. In the egocentric (obelisk) condition, the wall became invisible, but the obelisk remained visible, encouraging landmark-based spatial memory. Fixed distal cues (ie, mountain range and clouds) remained visible in both recall conditions as orientation cues. Each object was tested twice under both egocentric and allocentric conditions, resulting in a total of 16 trials (8 trials per recall condition).

An invisible wall was implemented as a constraint during the obelisk condition to ensure comparability between the 2 recall cue conditions. Figure 1 provides an overview of the task.

The recorded virtual variables were the encoding time to collect and encode the object locations; the Cartesian coordinates (x, y) of the recalled location; the Cartesian coordinates (x, y) of the paths made by participants during the encoding phase; and the error for each trial at recall, measured as the Euclidean distance (expressed in virtual Unity software meters) between the recalled position and the actual location. Details of the setup are shown in Figure S1 in Multimedia Appendix 1.

First, we analyzed path length in the HC and aMCI groups at encoding. To quantify participants’ movement within the arena, we computed the Euclidean distance between consecutive position samples (path Cartesian coordinates) for each participant. For each participant, we then computed the mean step distance (average distance traveled per time step) and the total distance (sum of all stepwise distances). Group differences in mean step distance between participants with HC and aMCI were assessed using an independent samples t test (with Welch correction for unequal variances). The same procedure was adopted for paths at recall. Encoding and recall path coordinates were sampled at 1 Hz (2/80 files could not be saved due to technical issues).

We also computed the Euclidean distance of each path point at encoding from the fixed landmark and the center of the arena and averaged this distance for each participant to carry out an independent samples t test (Welch correction for unequal variances) between the 2 groups.

Second, to analyze the spatial distribution of encoding and response patterns in the arena, we developed a systematic approach that allowed us to examine potential biases toward specific reference points in both egocentric and allocentric conditions, similar to that described by Lee et al [10]. Recall coordinates corresponded to each participant’s response.

Regarding the allocentric condition data point densities (encoding paths and recall data points), we divided the circular arena (radius=25 units) into inner (radius≤12.5 units) and outer annular regions to examine biases toward the center or boundary. We calculated point densities in each region by dividing point counts by their respective areas (inner area=π x 12.5² and annular area=π x [25² – 12.5²]). To determine whether participants showed significant deviations from the expected distribution based on relative areas (25% inner and 75% outer), we conducted binomial tests comparing observed proportions to the expected proportions. Additionally, we calculated density ratios (inner to outer) to quantify preferences toward central or peripheral regions.

For the egocentric condition densities (encoding paths and recall data points), we compared the quadrant containing the egocentric (obelisk) landmark (coordinates [−10.5, 10.5]) with the diametrically opposite quadrant. Specifically, we assessed whether participants placed objects more frequently in the upper-left quadrant (containing the obelisk) compared to the lower-right quadrant. We performed binomial tests to determine whether the proportion of points in the landmark quadrant significantly differed from 0.5 (the expected proportion under the null hypothesis of equal distribution between the 2 quadrants). We also calculated density ratios between these quadrants to quantify the strength of any spatial recall cue–based bias. In addition, we also calculated the ratio of density ratios between groups to quantify the relative strength of spatial biases across clinical populations.

Finally, before proceeding to compute data point densities for recall, we ensured that responses were unbiased to fixed object locations. We computed an intra-arena landmark attractor index, normalized by the distance of each object from the landmark. This index quantifies how much a response is shifted toward the intra-arena landmark ([d_target – d_responseLM]/d_target), where d_target is the Euclidean distance from the object’s original location to the landmark and d_responseLM is the distance from the participant’s response to the landmark. Positive values indicate a bias toward the landmark, whereas negative values indicate a bias toward the object’s true location. We also computed an arena center attractor index, normalized by the distance of each object from the arena center. This index quantifies how much a response is shifted toward the center of the arena ([d_target – d_responseC]/d_target), where d_target is the distance from the object’s original location to the arena center, and d_responseC is the distance from the response to the center. Negative values indicate a bias toward the center of the arena, whereas positive values indicate a bias toward the object’s true location. These 2 attractor indices were analyzed separately with a linear mixed-effects model, with participants and objects as random effects and group (HC vs aMCI) as a fixed predictor.

To study group and spatial recall cue differences, we performed a linear mixed-effects ANOVA for the error metric.

We found a significant main effect of the group variable (F_1,77=8.37; P=.005; and f=0.33). Regardless of the spatial recall cue, patients with aMCI demonstrated a higher mean error of 23.9 (SD 11.3) virtual meters than HC, who showed a mean error of 21.3 (SD 11.7). We did not find a main effect of spatial recall cue (P=.37) or the interaction between spatial recall cue and group (P=.27). The significant effect of the group variable (P=.03) was also retained after adjusting for the covariates [26], namely age, years of education, gender (P=.03), retrieval time for each object, global cognition, age by time, and age by gender (P=.02). Figure 2 shows the results of these analyses with (A) a significant main effect of the group variable on spatial memory error, and (B) nonsignificant spatial memory performance by spatial recall cue in the groups (mean and 95% CIs), with a detailed summary of these analyses reported in Multimedia Appendix 1.

Crucially, an ANOVA was conducted to explore any differences in recall time. We did not find a main effect of group (P=.07), spatial recall cue (P=.92), nor the interaction between group and spatial recall cue (P=.65). Results were unchanged when the model was adjusted for age, years of education, gender, and global cognition.

We also performed a polynomial linear mixed-effects model to investigate the performance across testing trials.

Concerning the error, the results of the group main effect above were replicated (P=.005). Crucially, we did not find any improvement or deterioration across testing trials (P=.40). The main effect of spatial recall cue (P=.37), the interaction between group and spatial recall cue (P=.27), the interaction between trials and group (P=.43), between trials and spatial recall cue (P=.98), and among trials, spatial recall cue, and group (P=.37) were not significant.

To analyze the effects of group and spatial recall cues on binding errors, we performed a Poisson regression. Neither the main effect of group (P=.24) nor spatial recall cue type (P=.85) reached statistical significance. The interaction between group and spatial recall cue type was also nonsignificant (P=.51). Results remained unchanged after adjusting for the covariates suggested by Castegnaro et al [26], namely age, education, gender, MMSE, and gender by age. Refer to Figure S2 for the binding errors in the groups. Figure S2 in Multimedia Appendix 1 shows the binding errors for the 4 objects.

To analyze the effects of group (aMCI vs HC) and recall cue switching condition (ego-to-allo vs allo-to-ego vs no switch) on error, we performed a linear mixed-effects ANOVA.

We found a significant main effect of group (F_1,78=4.71; P=.03; and f=0.25). Regardless of the switching condition, HC had higher retrieval accuracy (mean 21.3, SD 6.89) compared to the aMCI group (mean 23.5, SD 6.89). The switching condition (P=.97) and the group-by-switching condition (P=.18) were not significant. We also performed a model adjusting for age, education, gender, retrieval time, MMSE, age by gender, and age by time; however, in this case, the main effect of group did not survive (P=.11). No other significant effects were found for the adjusted model. See Figure S3 in Multimedia Appendix 1 shows the unadjusted frame-switching cost results.

In addition, we performed a logistic regression analysis to explore the predictive power of the recall cue switching condition in classifying aMCI vs HC. We found that only the no-switch error (β=.12; P=.01) was positively associated with aMCI diagnosis (OR 1.13, 95% CI 1.03-1.25). The other switch predictors were not significant (ego-to-allo: P=.78 and allo-to-ego: P=.57). The association for no-switch error remained significant after adding relevant covariates (P=.04), such as age, education, gender, MMSE (β=−.4; P=.01), and age by gender. This suggests that frame switching is impaired in both conditions, as the no-switch condition was a good predictor of aMCI, whereas the other 2 conditions were not.

We investigated the predictive power of group classification of egocentric and allocentric performances using a logistic regression model.

Regarding the error, our model revealed that egocentric error was a significant predictor of group status (β=.98; P=.01), with higher error associated with aMCI diagnosis. In contrast, allocentric error did not significantly predict group membership (β=.33; P=.39). The intercept of the logistic regression was significant (β=−5.96; P=.01); when egocentric and allocentric errors were 0, there was a greater chance of being classified as HC.

The egocentric error effect also remained significant (β=1.11; P=.03) after controlling for relevant covariates, namely age, education, gender (β_male=−1.65; P=.02), retrieval time, and MMSE (β=−.43; P=.01) [26]. This pattern of results suggests a selective impairment in egocentric spatial navigation in patients with aMCI compared with HCs, while allocentric spatial representations appear relatively preserved. Table 2 reports the full model results.

A receiver operating characteristic curve analysis was conducted to evaluate the diagnostic performance of the full model for the classification of aMCI. The AUC was 0.79, indicating moderate discriminatory ability (Figure 3). The optimal threshold was determined using the Youden index, which maximizes the sum of sensitivity and specificity. The analysis yielded an optimal probability cutoff value of 0.43. At this threshold, the test demonstrated a sensitivity of 0.88 and a specificity of 0.63, with a global accuracy of 0.75.

We also analyzed the cognitive measure of the model separately outside the logistic regression model. A receiver operating characteristic curve analysis was conducted to evaluate the diagnostic performance of the egocentric error measure. The AUC was 0.70, indicating moderate discriminatory ability. The optimal threshold was determined using the Youden index, which maximizes the sum of sensitivity and specificity. The analysis yielded an optimal cutoff value of 4.35. At this threshold, the test demonstrated a sensitivity of 0.80 and a specificity of 0.58. This means that egocentric error is particularly effective for ruling out aMCI, given the high sensitivity and, consequently, the low number of false negatives (ie, when egocentric error is below the cutoff, the participant is likely to be HC). Conversely, values above the cutoff may result in false-positive classifications because of the lower specificity.

Finally, MMSE showed an AUC of 0.68, indicating poor discriminatory ability. The analysis yielded an optimal cutoff value of 28.15. At this threshold, the test demonstrated a sensitivity of 0.68 and a specificity of 0.65. Consequently, the egocentric error is more accurate than MMSE for classifying healthy older adults.

Despite not being a significant predictor, we also computed a receiver operating characteristic curve analysis for the allocentric error. The AUC was 0.61, and the best cutoff was 4.87, which yielded a sensitivity of 0.5 and specificity of 0.75. This is specific to the egocentric error measure; in this case, the higher specificity is a better predictor of aMCI given the lower number of false positives, and a positive score above the cutoff may indicate aMCI.

We computed the patterns of encoding Cartesian (path) coordinates to explore navigation behavior in this phase. This data type allows us to explore the paths within the virtual environment during the encoding phase of the object locations. Refer to Figure 4 for the regions used in these analyses and the path density distributions in the groups at encoding. In the figure, (A) shows heat maps representing the density of encoding path points, with dark blue indicating higher concentration. Black lines represent paths taken by participants during the encoding phase, and black cross markers show object positions. The black dot represents the obelisk landmark (–10.5, 10.5), with dashed lines delineating the analyzed upper-left and lower-right quadrants; the solid circle represents the arena boundary (radius=25 units), with a dashed inner circle (radius=12.5 units) separating central and annular regions. (B) Shows heat maps of response point densities at recall, with dark blue indicating higher concentration. Black cross markers show encoding object positions. In the egocentric condition, the black dot represents the obelisk landmark (–10.5, 10.5), and in the allocentric condition, the solid circle represents the arena boundary (radius=25 units), with a dashed inner circle (radius=12.5 units) separating central and annular regions. Both cues were presented during encoding. Allocentric refers to a boundary cue presented, but no local landmark, and egocentric refers to local landmark presented, but no boundary. Legends show density levels.

First, we calculated whether the groups differed in terms of path length. A t test showed that groups did not differ in path length (P=.08). Along with encoding time (refer to Table 1), this result shows that time and path do not differ quantitatively in aMCI from normal cognitive aging.

Second, a Welch 2-sample t test was conducted to compare the mean distance from the intra-arena landmark between participants with aMCI and HC. There was a significant difference in mean distance between the 2 groups (t₆₃=2.35; P=.02). Participants with aMCI (mean 19.57, SD 1.34) maintained a greater average distance from the landmark during encoding navigation compared to HC participants (mean 18.65, SD 2.05). Similarly, a Welch 2-sample t test was conducted to compare the mean distance from the center of the arena between participants with aMCI and HC. There was a significant difference in mean distance between the 2 groups (t₇₆=2.35; P=.02). Participants with aMCI (mean 15.45, SD 1.96) maintained a greater average distance from the center compared to HC participants (mean 14.45, SD 1.77). These analyses showed a tendency to navigate closer to the boundary of the arena and away from the landmark, which is replicated in the data point density analyses.

HC demonstrated significantly higher point density in the inner circle compared to patients with aMCI (36.4% vs 29.16%, binomial test P<.001), suggesting a centripetal pattern during the encoding phase for HC. Conversely, patients with aMCI showed significantly higher concentration in the outer ring (70.83% vs 63.6%, binomial test P<.001), suggesting a centrifugal bias toward the boundary during the encoding phase.

Quadrant analysis revealed that HC had significantly higher point density in the top-left quadrant (region that contained the egocentric cue [obelisk]) compared to participants with aMCI (27.75% vs 24.37%; P<.001), suggesting a bias toward the egocentric landmark for HC. Conversely, participants with aMCI showed a significantly higher concentration in the bottom-right quadrant (14.98% vs 13.64%; P<.001), suggesting a centrifugal pattern away from the egocentric cue.

Overall, the results indicate a centrifugal pattern of encoding navigation for aMCI away from the egocentric cue and toward the boundary (away from the center of the arena) and a more clustered pattern for HC toward the center of the arena and the egocentric cue.

We analyzed the response patterns to investigate behavioral recall differences between the groups.

First, we calculated whether the groups differed in terms of path length. Results showed that aMCI had significantly (t₇₄=3.64; P<.001) longer paths (mean 2.83, SD 0.89) compared to HC (mean 2.16, SD 0.74).

Second, normalized attractor index analyses for egocentric and allocentric performance showed that the HC group had a lower bias toward the landmark (β=−.21; SE=0.07; P=.007) and toward the center of the arena (β=−.14; SE=0.06; P=.02) compared to the aMCI group. This ensured that the following analyses were relatively unbiased by fixed object locations. Indeed, this result was replicated in the data point densities analyses reported in the next paragraph: aMCI recalled locations in the allocentric condition toward the center of the arena and toward the landmark during the egocentric condition.

Regarding data point densities, we started by investigating the point pattern for allocentric performance. We analyzed the spatial distribution of response points in the 2 groups by examining the density patterns relative to the structure of the circular arena. For this purpose, we divided the circular arena into an inner region (within the inner circle of radius 12.5 units, far from the wall) and an outer annular region (between the inner and outer circles, 12.5< radius ≤25 units, closer to the wall). The inner area constituted 25% of the total area, while the outer area represented 75%.

Refer to Figure 4 for the regions used in these analyses, the groups’ density point distributions, and recall cue conditions.

The HC group placed 17.81% of their recall points in the inner area (57/320 points) and 82.19% in the outer area (263/320 points). This distribution significantly differed from the expected proportions based on the inner and the outer portion of the arena (binomial test; P=.002).

In contrast, the aMCI group placed 23.75% (76/320) of their recall points in the inner area and 76.25% (244/320) in the outer area. This distribution did not significantly differ from the expected proportions for the inner and outer sections of the arena (binomial test; P=.65). To account for the different sizes of the inner and outer areas, we calculated point densities (points per square unit). In the HC group, the density was 0.11 points/unit² in the inner area and 0.18 points/unit² in the outer area, yielding an inner-to-outer density ratio of 0.65.

For the aMCI group, the density (points per unit area) was 0.15 points/unit² in the inner area and 0.17 points/unit² in the outer area, resulting in an inner-to-outer density ratio of 0.93. The difference in density ratios between groups was −0.28, indicating that patients with aMCI showed a more balanced distribution of points between inner and outer areas than HC participants, who demonstrated a stronger preference for the outer area. This suggests a potential alteration in allocentric spatial memory strategies in aMCI, characterized by reduced usage of the circular boundary as a spatial reference point compared with healthy older adults.

We then inspected the egocentric performance. We examined the spatial distribution of response points across opposing arena quadrants in patients with aMCI and HC. We focused on comparing the upper left quadrant, which contained the egocentric cue (obelisk at coordinates [−10.5, 10.5]), with the diametrically opposite lower right quadrant.

In the HC group, 32.19% (103/320) of recall points were positioned in the upper left quadrant, while 17.19% (55/320) were located in the lower right quadrant. This asymmetric distribution between the 2 quadrants was statistically significant (binomial test; P<.001).

Similarly, the aMCI group placed 32.50% (104/320) of recall points in the upper left quadrant and 20.00% (64/320) in the lower right quadrant. This distribution also showed a significant preference for the upper left quadrant (binomial test; P=.003).

To account for the spatial area, point densities were calculated for each quadrant. In the HC group, the point density was 0.21 points/unit² in the upper left quadrant and 0.11 points/unit² in the lower right quadrant, yielding an upper-left–to–lower-right density ratio of 1.87.

For the aMCI group, the point density was 0.21 points/unit² in the upper left quadrant and 0.13 points/unit² in the lower right quadrant, resulting in a density ratio of 1.62. The difference in density ratios between groups was 0.25, indicating that HC participants showed a slightly stronger preference for the upper left quadrant compared to patients with aMCI. The absolute differences in point densities between groups were minimal in both the upper left (−0.01 points/unit²) and lower right (−0.01 points/unit²) quadrants, with patients with aMCI showing slightly higher densities in both regions.