Statement of Research Significance
Research Question(s) or Topic(s): How does a source memory task contribute to memory processes across different age groups? Main Findings: Older adult groups benefit more from the source memory task than their younger counterparts when performing delayed recall tasks. Introducing a source memory task may help some age groups elaborate further on material to be learned and remembered. Study Contributions: Some age groups may be more adequately assessed using a source memory task that explores both the content and context of memory. The novelty of the study is the introduction of a source memory task that requires the processing of contextually relevant information. This contextual processing does not seek to interfere with or create confusion in the respondent, but rather to delve deeper into the coding of the material to be remembered, which appears to improve subsequent recall, especially in certain age groups.
Introduction
Memory refers to the processes by which an individual encodes, stores, and retrieves information (Strauss et al., Reference Strauss, Sherman and Spreen2006). According to Lezak et al. (Reference Lezak, Howieson, Loring and Fischer2012), many common neurological and psychiatric conditions show a decline in the efficiency of memory processes, and, consequently, assessment of memory is often a central issue in a neuropsychological examination.
With regard to this, a proliferation of assessment tools exists for the assessment of specific aspects of verbal memory (see, for example, Lezak et al., Reference Lezak, Howieson, Loring and Fischer2012; Sherman & Hrabok, Reference Sherman and Hrabok2023) and many of these tools have employed the use of interference as an integral part of standardized verbal memory assessment (Brophy et al., Reference Brophy, Jackson and Crowe2009). Some examples are the widely used Rey Auditory Verbal Learning Test (Rey, Reference Rey1942) or the California Verbal Learning test (Delis et al., Reference Delis, Kramer, Kaplan and Ober2000). In these tools, there is a set of learning trials of a List A (or Monday shopping list) that is presented five times to the testee, followed by a List B (or Tuesday shopping list) that is presented in a single trial. When the testee has to repeat items from List B, according to the proactive interference theory (Underwood, Reference Underwood1957), the information presented before (List A, presented five times) may prevent or make the acquisition of new materials difficult, thus making it difficult to remember items from List B. Moreover, some items of List A may mistakenly be remembered as part of List B (i.e., intrusions). After the presentation of List B or Tuesday list, the testee is usually requested to remember items from List A, which opens the door to retroactive interference, thus causing that information presented after the target material (i.e., List B presented once after List A presented five times) negatively affects recall (McGeoch, Reference McGeoch1932).
This way of testing verbal memory has been the main paradigm for years, both in research and clinical practice, and only recently have some studies appeared to question or at least test the effects of interference paradigms in memory performance. For example, Brophy et al. (Reference Brophy, Jackson and Crowe2009) investigated the effect of interference on delayed recall scores of the WMS-III and other commonly used memory measures. They found that the introduction of interference items during the delay affected negatively delayed recall performance on almost all sub-tests. Libon et al. (Reference Libon, Bondi, Price, Lamar, Eppig, Wambach, Nieves, Delano-Wood, Giovannetti, Lippa, Kabasakalian, Cosentino, Swenson and Penney2011) found that, as compared to other groups, amnestic MCI patients appeared to be very susceptible to the deleterious effect of interference test conditions, along with greater penetration of list B words into subsequent list A recall. Separately Rahimi-Golkhandan et al. (Reference Rahimi-Golkhandan, Maruff, Darby and Wilson2012) suggested that knowing about differences in susceptibility to interference across tasks may provide some important diagnostic and cognitive information for researchers and clinicians, and the choice of verbal learning tests should be guided by the knowledge of interference effects and the susceptibility of patient groups to this effect. Overall, while the use of interference could be a differential factor for specific memory profiles, it can also constitute a confounding factor for memory and recall if its effects are not properly examined.
In this context of aiming to overcome problems traditionally associated with memory assessment, virtual reality (VR) offers a potentially interesting alternative for the assessment of many cognitive processes and the inclusion of other memory processes potentially relevant for clinical diagnosis, such as source memory. VR offers several advantages in the context of neuropsychological assessment and cognitive research, such as increased ecological validity (allowing observation of behaviors in complex, dynamic settings that better reflect everyday life) and the ability to minimize examiner-related variability (Diaz-Orueta et al., Reference Diaz-Orueta, Blanco-Campal, Lamar, Libon and Burke2020; Parsons, Reference Parsons2015; Pieri et al., Reference Pieri, Tosi and Romano2023; Rizzo et al., Reference Rizzo, Schultheis, Kerns and Mateer2004). However, VR is not without its limitations. Ceccato et al. (Reference Ceccato, Ricci, Mazza, Bartolini, Di Crosta, La Malva, Biondi, Colasanti, Mammarella, Palumbo, Roma and Di Domenico2024) showed that recognition accuracy and confidence were significantly higher in real-life modality environments than in similar VR modalities, and that additional research is needed to make VR environments sufficiently comparable to real-life contexts, paying special attention to the impact of stimuli typicality and emotional valence in VR contexts. In addition, variability in user familiarity with technology and challenges in standardizing VR protocols across studies may affect the reliability and generalizability of results.
Despite these drawbacks, ongoing technological advancements and increasing empirical support suggest that VR holds significant potential as a complementary tool in neuropsychological evaluation. Recently, Climent et al. (Reference Climent, Rebon-Ortiz, Saura-Carrasco and Diaz-Orueta2024) published the normative data for a new developed VR-based neuropsychological test, the Suite test, on a population between 12 and 85 years old. Suite is a VR-based test designed to evaluate visual memory and aims to help clinicians support the diagnosis of memory-related conditions or disorders. One of the features of Suite Test is the inclusion of a source-memory task.
Source memory can be defined as memory for details that accompany the central component of an event, or also as memory for temporal details –as well as other contextual details- of the encoded information (Palombo et al., Reference Palombo, Te, Checknita and Madan2021). In their study, Minor and Herzmann (Reference Minor and Herzmann2019) defined source memory in an operational way as the ability to recall both an image and its block, as opposed to item memory, which was represented by memory for the image irrespective of the memory for its block.
When assessing source memory, it can be done by asking the individual where, from where, how, or when the information was learned (like in Ceccato et al., Reference Ceccato, La Malva, Di Crosta, Palumbo, Gatti, Momi, Logrieco, Fasolo, Mammarella, Borella and Di Domenico2022, where they requested their participants to remember when each time was seen among three possible timepoints). Separately, Squire et al. (Reference Squire, Wixted and Clark2007) presented a task that required individuals to be exposed to items such as a list of words, in two separate states (at the top vs. the bottom of a computer screen), and then declare whether the information originated from the top or the bottom of the screen. Symeonidou & Kuhlmann (Reference Symeonidou and Kuhlmann2021) used a task with audio software to distinguish ‘sources’, alongside visual faces and corresponding names. The participants were not only asked to differentiate whether an item came from source A, source B, or was new; but they were also asked to describe their retrieval strategies. This is different from item memory, as item memory involves memory retrieval linked to the semantic features of information (Guo et al., Reference Guo, Shubeck and Hu2021), although these two forms of memory have been shown to have an almost symbiotic relationship. The retrieval of both source memory and item memory increases when the relationship between these two components is strongly linked (Guo et al., Reference Guo, Shubeck and Hu2021), and this was the rationale behind the development of the source memory task within the Suite Test.
In summary, the exploration of source memory reveals a nuanced and interconnected cognitive operation essential for recalling not just what we learn but also how and where we learn it. The symbiotic relationship between source and item memory, intricately orchestrated by the prefrontal cortex and the hippocampus (Guo et al., Reference Guo, Shubeck and Hu2021) suggest that a memory test comprising a set of tasks that measure both item and source memory could be a potential addition to tests relying on interference-based tasks or paradigms, and a way to further understand the development of memory processes across the lifespan.
For all these reasons, the goal of the current study is to focus on the source memory task of the Suite test, a recently developed VR-based neuropsychological assessment tool for memory. This research aims to contribute to building a body of knowledge about this type of memory and its relevance, even in non-pathological processes. In order to achieve that, the present study aims to evaluate how source memory, integrated into a virtual reality (VR)-based neuropsychological task (the Suite Test), contributes to overall memory performance across different age groups. In line with clinical reporting conventions, we will use immediate recall and delayed recall to label test scores. Conceptually, we will refer to memory processes (e.g. source/contextual memory) when discussing mechanisms, in order to avoid conflating score labels with theoretical claims about memory systems or stores.
Method
Participants
This study employs the normative sample of the SUITE test (N = 676) as delineated by Climent et al. (Reference Climent, Rebon-Ortiz, Saura-Carrasco and Diaz-Orueta2024). The participant cohort included individuals aged 12 to 85 years, demonstrating a balanced gender distribution, with 49.7% of female participants (mean age = 32.65 years old) and 50.3% of male (mean age = 30.46 years old). Exclusion criteria encompassed individuals with visual, auditory, or motor impairments that could adversely impact their ability to interact with the test. Furthermore, individuals with a history of acquired brain injury, psychiatric disorders, or neurodegenerative diseases were excluded to maintain the validity of the findings. Table 1 below shows the age distribution of the normative groups.
Table 1. Normative sample age distribution
The original sample included 676 subjects, but a total of 29 participants were excluded from the final sample. In 14 cases, reasons for exclusion included the participant removing the VR headset during the test, misunderstanding the instructions (which led to invalid performance), or hardware malfunctions during the test administration. Additionally, another 15 subjects were excluded due to poor performance on a forced-choice recognition task (i.e., worse performance than what would be expected if the subjects answered by chance -i.e. less than half of correct answers in that task, as will be detailed later in the procedure), which was attributed to factors like faking their performance, fatigue, or tremors. As a result, the final sample comprised 647 subjects, representing a 4.29% exclusion rate.
Ethics approval statement
The study obtained ethical approval from the Human Research Ethics Committee of the University of the Basque Country in northern Spain, and all participants provided informed consent before the beginning of the study. The study was conducted following the Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans.
Study procedure
To administer the SUITE test, the participant sat comfortably in a chair and was provided with a VR headset and a controller operated with the dominant hand, like a joystick with buttons that allowed the participant to click and interact with various features of the virtual environment. The entire setup process, which included the participant putting on the VR headset, the evaluator’s computer starting the SUITE VR desktop control application and connecting to the same network as the participant’s VR headset, the evaluator entering basic participant information into the SUITE computer application (name or code, age, gender, and whether the participant was left-handed or right-handed), and starting the test administration, took a maximum of 30 minutes.
Instruments
Suite is a comprehensive neuropsychological assessment tool designed to evaluate memory functions, including immediate memory, source memory, short and long delayed free recall, recognition, and memory strategies (visual, verbal, primacy, recency), in individuals aged 12 years and older, through an immersive VR environment facilitated by an Oculus Quest headset. The virtual environment is set in a furniture store where participants are tasked with organizing various pieces of furniture based on specific criteria for packing and shipping. Using the head-tracking technology of the VR headset, participants can explore the environment with a 360-degree view. As described in Climent et al. (Reference Climent, Rebon-Ortiz, Saura-Carrasco and Diaz-Orueta2024), in the virtual scenario, there is a television screen located in a central position. On that screen, stimuli (i.e., pieces of furniture) are presented with an image of each of them and their name. A male voiceover in European Spanish names the furniture that must be packed, and the respondent has to locate each item in the virtual room, point and click on it. Additionally, the voice informs participants about different customer groups—referred to as “families”— each requesting between four and six pieces of furniture that must be remembered. The customer groups include (1) a family of four (a man, a woman, and two children), (2) a single man, (3) two men, (4) a single woman, and (5) a family of five (a man, a woman, and three children). Figure 1 below shows a sketch of the original Suite environment. To preserve the validity of the assessment, further details regarding the interface and specific stimuli or customer groups are not disclosed to prevent overexposure.
Figure 1. Sketch of the Suite virtual reality scenario.
SUITE collected data related to response accuracy, defined as measures of how accurately participants completed tasks within the VR, such as correctly identifying or grouping furniture items (correct answers and errors), response times, and strategies used.
Below, the different tasks administered in Suite are displayed:
-
Task 1: Immediate Recall (Order List): The user must select and recall furniture ordered by different families (5 groups with orders of 4-6 pieces each). Three rounds of the same order are presented to assess the learning curve. The selection strategy is analyzed (i.e., does the user select the furniture in sequential order or by category?). The recorded data are: (1) selected furniture and their order; (2) response times in milliseconds, and (3) memory strategies (primacy, recency, verbal, or visual strategies).
-
Task 2: Source Memory: A total of eight items of previously ordered furniture are presented, and the user must remember which family requested each of them. Images of the family groups appear on the screen to select the answer. The recorded data are: (1) accuracy in matching each piece of furniture ordered with the right family; (2) response time, and (3) use of a virtual red button available on the screen if the user cannot remember the answer. Source memory, as measured here, refers to the contextual information around target items that is not explicitly requested to the individual to remember. In Suite test, that contextual information would refer to “families” who make the orders, and target items would be the specific set of pieces of furniture requested in each order that need to be learned and remembered by the testee.
-
Tasks 3: Free Delayed Recall (short delay): Each family is shown again, and the user must select the furniture that corresponds to them, without receiving any clues from previous tasks. The recorded data are: (1) accuracy in retrieving orders; (2) response time and pauses (i.e., latencies) between selections, and (3) a path map within the environment (i.e., a graphic representation of the scanning sequence performed by the user to select the target stimuli).
-
Task 4: Prospective memory item: The user must remember to turn off the store lights when a bell rings. It is registered whether the user remembers to act or not, and the response time.
-
Task 5. Free delayed recall (long delay): Same task as task 3, presented after a 20-minute delay.
-
Task 6. Recognition (Yes/No): Images of furniture are presented, and the user must indicate whether they were part of the order. Targets were mixed with distractors present in the shop (seen but not ordered) and items never seen before. More specifically, a total of 18 items are presented, six of them being target stimuli, nine of them also being in the shop (but not being part of any order), and the other three are furniture items never presented before. The recorded data are (1) accuracy in recognizing the correct piece of furniture, (2) response time, and (3) false positives and negatives.
-
Task 7: Forced choice recognition task: Each trial presented a studied target stimulus paired with a novel item from the same semantic category, equating familiarity and category cues while imposing a binary, comparative decision. More specifically, a total of 7 pairs of items are displayed (i.e., the user has to make seven decisions), seven of which are target stimuli and seven that have never been shown to the user before. Subsequently, the user must choose each one that was present. Under random responding, expected performance would obtain between 3 and 4 correct answers (out of 7). We used ≤2/7 as a heuristic screen (together with ancillary indicators such as very short reaction times or inconsistent responding) to identify non-credible performance, and flagged cases were excluded. This task served as a performance validity screen and was not included in psychometric summaries of memory. The recorded data are: (1) selection accuracy, (2) response time, and (3) a discriminability index.
Previously performed reliability analyses (Climent et al., Reference Climent, Rebon-Ortiz, Saura-Carrasco and Diaz-Orueta2024) showed excellent internal consistency for immediate, short, and long delay recall tasks (Cronbach’s alpha of 0.93 for tasks 1, 3, and 5); moderate for source memory (Cronbach´s alpha of 0.64 for task 2) and good for recognition (Cronbach´s alpha of 0.79 for task 6). Reliability analyses for Task 4 showed very poor results (with less than 10% of individuals from the whole normative sample remembering to switch off the light, which led to the decision to exclude this task from the normative data), and items from Task 7 (i.e., forced choice recognition, merely used as a task to estimate performance validity and/or malingering/simulation) were not considered. Construct validity (PCA) is also detailed in the Suite normative study (Climent et al., Reference Climent, Rebon-Ortiz, Saura-Carrasco and Diaz-Orueta2024).
The study also considers two other variables: the age and gender of the examinees, as Suite reveals differences in the behavior of the subjects according to age group (12 – 12, 13 – 26, 27 – 44, 45 – 58, 59 – 85) and, in some groups, even according to gender (13 – 26 and 45 – 58).
Data analysis
The following diagram in Figure 2 shows an overview of the data analysis process performed.
Figure 2. Graphical overview of data analysis performed. 1. Random init: Neuron weight vectors are initialized randomly, allowing the map to adapt to data without initial bias. 2. BMU (best matching unit): The neuron whose weights are most similar to the current input (usually by Euclidean distance); this “winner” is updated most strongly. 3. Neighborhood: Not only the BMU but also neighboring neurons are updated, with the degree of adjustment decreasing with distance from the BMU and over time. This preserves the topological structure of the map.
The statistical analyses and data management were carried out with R version 4.4.1 through several key libraries (R Core Team, 2021). The RVAideMemoire package facilitated the execution of non-parametric statistical tests (HERVE, Reference Herve2023), while tidymodels streamlined the modeling process for data consistency and efficiency (Kuhn & Wickham, Reference Kuhn and Wickham2020). The kohonen package was used to implement the Kohonen network algorithm (Wehrens & Buydens, Reference Wehrens and Buydens2007; Wehrens & Kruisselbrink, Reference Wehrens and Kruisselbrink2018). Additionally, the boot package enabled bootstrapping procedures (Canty & Ripley, Reference Canty and Ripley2024), and randomForest package was utilized for random forest analyses (Liaw & Wiener, Reference Liaw and Wiener2002). The cluster package assisted in creating a silhouette index (Maechler et al., Reference Maechler, Rousseeuw, Struyf, Hubert and Hornik2023), while performance package was employed to evaluate and diagnose statistical models (Lüdecke et al., Reference Lüdecke, Ben-Shachar, Patil, Waggoner and Makowski2021). For data visualization, ggplot2 package was used to generate graphs (Wickham, Reference Wickham2016), complemented by the magick package for graph manipulation (Ooms, Reference Ooms2024). Data management and manipulation were conducted using dplyr package (Wickham et al., Reference Wickham, François, Henry, Müller and Vaughan2023), and the tidylog package provided insights into data transformations throughout the analysis (Elbers, Reference Elbers2024).
Self-Organizing Maps (SOM) were utilized to analyze memory behavior patterns during both general and scale group testing (Brereton, Reference Brereton2012; Oliver et al., Reference Oliver, Vallés-Ṕerez, Baños, Cebolla, Botella and Soria-Olivas2018). SOM is an effective tool to visualize patterns in complex data, enabling the identification of similar data clusters and their separation. The neuron architecture consists of two layers: the input layer that contains one neuron per input variable, and the competition layer, typically structured as a two-dimensional, low-dimensional topological grid. Each neuron in the input layer is connected to every unit in the competition layer, with each unit assigned to an N-dimensional weight vector. SOM training started with the random initialization of weight vectors assigned to each neuron. During the training phase, input data was presented as characteristic vectors that were compared to the weight vectors of all neurons within the map. The neuron that exhibited the closest match to a given input vector was designated as the “winner” or “best matching unit” (BMU). Subsequently, the BMU and its neighboring neurons updated their weight vectors to more closely align with the input vector, employing neighborhood functions and progressively decreasing learning rates. This iterative process was repeated across multiple epochs, incorporating the entire dataset. Upon completion of the training, the neurons were organized such that those with similar weight vectors were positioned adjacent to one another, thereby facilitating the visualization of patterns and the identification of clusters within the data.
Authors such as Tervonen et al. (Reference Tervonen, Puttonen, Sillanpää, Hopsu, Homorodi, Keränen, Pajukanta, Tolonen, Lämsä and Mäntyjärvi2020) and Younger et al. (Reference Younger, Schaerlaeken, Anguera and Gazzaley2024) performed a subsequent clustering of the nodes using k-means to identify similar nodes. Due to the limited number of variables (four types of memory processes, named here immediate recall, source memory, short-term delayed recall, and long-term delayed recall), the resulting profiles were easily interpretable, making further grouping unnecessary. To assess the quality of the generated clusters, the Silhouette index was calculated. This index measures the similarity of each object to other clusters, providing a quantitative evaluation of clustering effectiveness (Rousseeuw, Reference Rousseeuw1987). In this context, the Silhouette index was applied to the clusters identified by the SOM algorithm.
Conducting several SOM analyses for the groups linked to the test (12 – 12, 13 – 26-M, 13 – 26-F, 27 – 44, 45 – 58-M, 45 – 58-F, 59 – 85) enables the monitoring of pattern evolution over time. To improve comparability, the nodes were restructured based on the strength of the defining variables. This approach ensured that the interpretation of patterns remained clear and was not influenced by the specific nodes where they appear.
After analyzing the memory profiles of each group, the source memory was subjected to an incremental validity assessment (IVA). It was used to determine whether new tests or measures provided significant additional information on a psychological variable beyond what was obtained with existing measurements. IVA is a statistical bootstrap technique that allows statistical distribution estimates (e.g., average, median, and regression coefficient) by repeating the original data with replacement (Davison & Hinkley, Reference Davison and Hinkley1997). The statistic used is the determination coefficients (R 2), which measure the proportion of the predictable variance of independent variables in dependent variables. The difference between R2 between the two models was calculated for more than 10,000 replications: the first model did not include source memory as an independent variable, while the second model did. Both models used long-term memory as a dependent variable because long-term memory was considered a successful fingerprint if the subject performed it correctly (Cotton & Ricker, Reference Cotton and Ricker2022).
To ascertain the most suitable technique for the models, it was essential to conduct a study, at the very least in graphical form, on the distribution of the residuals to detect any unusual patterns. Furthermore, it was essential to evaluate the presence of multicollinearity among the predictor variables, ascertaining the linear relationship between the predictor variables and the response variable, verifying whether the residuals had constant variance, and confirming whether the residuals followed a normal distribution (Osborne & Waters, Reference Osborne and Waters2019).
The choice of techniques depended on the results of the study. Statistical or regression models such as simple linear regression, general linear model, generalized least squares, additive model (AM), least absolute deviation (LAD), general AM, general linear mixed model (GLMM), general additive mixed model, are based on statistical and mathematical principles to model relationships between variables, or machine learning models, such as support vector machine, decision trees or random forests, which are more flexible and are used to capture complex patterns in data (Hastie et al., Reference Hastie, Tibshirani and Friedman2017).
Results
Once the sample was obtained, the SOM algorithm was applied to the whole sample and each age group, resulting in the graphical representation of the SOM (see Figure 3 in the text and Supplemental Figure 1 in the Appendix).
Figure 3. Profiles from the complete sample provided by SOM. Figure explanation: Each circle represents a node identified by the SOM, functioning as the centroid of a cluster. The “+” symbols indicate individual data points assigned based on similarity; the gray intensity may reflect different feature values. The proximity between symbols indicates greater similarity among observations. Percentages and numbers below each node show the proportion and number of data points assigned to each group relative to the total sample.
In this graph, the patterns of memory behavior in the subjects can be visually verified. For example, Figure 1 shows that Node 1 is completed by 33.25% of the subjects whose four memories have a similar weight. In contrast, Node 2 is completed by 21.48% of the subjects, where the source memory has the highest weight, followed by the immediate recall, and the short-term and long-term delayed recall have a similar weight.
In node 3, completed by 27.2% of the subjects, immediate recall has the highest weight. Here, the relationship between short-term and long-term delayed recall is stronger than in Node 2, and the presence of source memory is lower than that of the other memories.
In node 4, which includes 18.08% of the subjects, all the memories are represented, but their strength in the distribution of the patterns in the other nodes is so low that their graphing is not observable. In order to improve the interpretation, these patterns have been tabulated (see Tables 2 and 3).
Table 2. Importance of variables in each node by group
Note: A (immediate recall), B (short-term delayed recall), C (long-term delayed recall), D (source memory).
M = male; F = female.
Table 3. Memory patterns in the groups.
Note: A (immediate recall), B (short-term delayed recall), C (long-term delayed recall), D (source memory); M (male), F (female).
Data in Tables 2 and 3 illustrate how different memory types engage with various age groups, with younger individuals tending to exhibit stronger immediate and short-term delayed recall performance, while older adults may rely more on long-term recall and source memory. Before obtaining Table 2, it was confirmed that the identified nodes represented distinct groups. This was achieved by plotting the silhouette index for each group on the appropriate scale (see Figure 4 in the text and Supplemental Figure 2 in the Appendix). In all cases, a silhouette index value greater than zero is observed, which indicates that the groups are well delimited. However, some values are not well classified in terms of distance, suggesting the need to increase the number of clusters. Nevertheless, these subjects not classified within a group are insignificant compared to those belonging to the well-defined groups (see Supplemental Table 1 in the Appendix).
Figure 4. Silhouette indexes on the general sample. Figure explanation: Silhouette plot evaluating clustering quality. Each horizontal bar represents the silhouette width of an observation, grouped by cluster. Values near 1 indicate strong cohesion and clear separation between clusters; low or negative values suggest possible misclassifications or boundary points. Cluster sizes and averages are shown on the right.
The graph illustrates the silhouette analysis for data clustering with four groups. Group 1 is the largest with 215 members (n1 = 215), followed by Group 3 (n3 = 176), Group 2 (n2 = 139), and Group 4 (n4 = 117). The silhouette width measures the similarity of a data point to its own group compared to other groups. A higher value indicates better clustering. Therefore, Group 1 has the highest average silhouette width (0.44), suggesting it is well-separated from the other groups. Conversely, Group 2 has the lowest average silhouette width (0.16), indicating there may be some overlap with other groups. The overall clustering quality is moderate, suggesting that there may be room for improvement in the selection of the number of groups used, as previously explained.
After completing the study of memory profiles for each age group, the focus turned to conducting the IVA. Figure 5 in the text and Supplemental Figure 3 in the Appendix show that in none of the cases is there linearity, homogeneity of variance in the residuals, or normality. There is also no multicollinearity, as confirmed by the visual inspection of the model assumptions.
Figure 5. Memory patterns in the groups. Figure explanation: Diagnostic plots for assessing regression model assumptions and quality. Each panel displays a different aspect: fit between observed and predicted data, linearity, homogeneity of variance, presence of influential observations, collinearity among predictors, and normality of residuals. These plots help identify potential deviations from model assumptions and validate model adequacy.
This analysis indicates that the model to be implemented in IVA should be a machine learning model. Random Forests algorithm was selected for its high accuracy and robustness to overfit, due to the combination of multiple trees. Moreover, it offers stability, is efficient in terms of training and prediction time, and provides a measure of the importance of the variables, which facilitates the interpretation of the model.
The non-parametric bootstrap analysis results presented in Table 4 reveal distinct patterns in memory performance across different age and gender groups. Notably, younger participants (ages 13 – 26) exhibit higher initial estimates along with greater imprecision and bias, indicating less stable memory performance. In contrast, the oldest age group (59 – 85) shows a negative original value and the lowest standard error, suggesting greater precision in their memory estimates. Furthermore, significant differences between genders are particularly evident in the younger age group.
Table 4. Ordinary non-parametric bootstrap results by group.
Note: M (male), F (female).
Discussion
The present study aimed to show how performance in a VR-based source memory task contributes to performance in immediate and delayed memory trials of the same test (the VR-based Suite test) across the different age groups included in the normative sample (Climent et al., Reference Climent, Rebon-Ortiz, Saura-Carrasco and Diaz-Orueta2024). The data presented here show the relevance of source memory for memory performance in individuals between 12 and 85 years old. The main variables of the tests are linked to immediate, short-term and long-term delayed recall memory processes, but instead of using interference tasks that may compromise performance (Brophy et al., Reference Brophy, Jackson and Crowe2009; Libon et al., Reference Libon, Bondi, Price, Lamar, Eppig, Wambach, Nieves, Delano-Wood, Giovannetti, Lippa, Kabasakalian, Cosentino, Swenson and Penney2011), the effects of a source memory task inserted between the immediate memory trials and the short-term delayed recall trial were investigated. While interference tasks are typically designed to assess susceptibility to forgetting and the ability to resist the intrusion of competing information, thereby draining working memory resources and measuring long-term consolidation, this new source memory task operates under a different neuropsychological principle. In Suite, participants are instructed to learn the association between each item and its corresponding ‘family’ (the contextual source). No manipulation or transformation of the context is requested. Thus, encoding is explicitly associative/contextual —via item–source binding— which may deepen the memory trace and facilitate retrieval. Neuropsychologically, this implies a greater emphasis on frontal lobe functions related to source monitoring and memory elaboration, and on hippocampal connections for the formation of item-context associations, in contrast to a primary focus on the also frontal but substantially distinct inhibition, or resistance to interference, that is key in traditional paradigms (Guo et al., Reference Guo, Shubeck and Hu2021).
Results highlight how source memory contributes to memory performance when considered together with other memory types across age groups. Younger individuals tend to rely more on immediate and short-term delayed recall, while older adults exhibit a stronger reliance on source memory and long-term delayed recall, suggesting a shift from rapid recall in youth to more reflective processing in older age (Čepukaitytė et al., Reference Čepukaitytė, Thom, Kallmayer, Nobre and Zokaei2023). This evolution in memory aligns with theories of cognitive aging, where older adults use deeper processing strategies to elaborate and remember information (Craik et al., Reference Craik, Byrd and Swanson1987; Light & Anderson, Reference Light and Anderson1985; Salthouse et al., Reference Salthouse, Fristoe and Rhee1996). Additionally, older adults’ increased reliance on long-term memory supports previous research indicating more effective retrieval strategies over time (Park & Reuter-Lorenz, Reference Park and Reuter-Lorenz2009). Gender differences were particularly noticeable in younger participants, emphasizing the complex interplay between age, gender, and memory. Incorporating source memory tasks into the assessment of memory performance provides further insight into long-term memory functioning across the lifespan, though it appears less relevant for 12-year-olds and females aged 45 – 58, where negative values were observed. These findings contribute to a deeper understanding of age-related memory changes and emphasize the crucial role of source memory in assessing memory across different demographics. As source memory seems to play a more prominent role in memory processes associated with older populations, a defective ability to rely on source memory may lead to a decrease in memory performance and thus be an important indicator of a decline in memory processes due to different etiologies. Subsequently, incorporating assessment tasks to further evaluate the individual’s ability to take advantage of source memory to boost their overall memory performance may help identify early signs of memory decline across different conditions.
The current study has obvious limitations. First, it is focused on a normative population from Spain. To enhance its generalizability, future research should involve cross-cultural validations, examining diverse populations in both community and clinical settings. Second, as literature has shown that selected clinical populations may not benefit from a memory assessment paradigm that includes interference tasks (Brophy et al., Reference Brophy, Jackson and Crowe2009; Libon et al., Reference Libon, Bondi, Price, Lamar, Eppig, Wambach, Nieves, Delano-Wood, Giovannetti, Lippa, Kabasakalian, Cosentino, Swenson and Penney2011), and since replacing it with source memory may potentially strengthen memory performance in certain age groups, a further confirmation of this hypothesis would require a massive administration of the test to larger samples in different clinical settings. Third, the disparity in the number of participants in the composition of age groups and, especially, the small number of subjects in the 12-year-old age group may demand a careful interpretation of results for this age group. Finally, further studies with specific clinical populations would help to determine the sensitivity and specificity of the test.
Despite the limitations, these preliminary results on the impact of introducing source memory tasks instead of interference tasks may improve the way neuropsychological assessment of memory is performed, particularly in clinical populations where interference paradigms may exacerbate cognitive challenges. Unlike traditional interference-based approaches, which primarily highlight memory deficits, source memory tasks provide a richer understanding of both the content and context of memory retrieval, making them more versatile for diverse populations.
The novelty of the study is the introduction of a source memory task that requires the processing of contextually relevant information. This contextual processing does not seek to interfere with or create confusion in the respondent, but rather to delve deeper into the coding of the material to be remembered, which appears to improve subsequent recall, especially in certain age groups.
This distinction highlights the practical implications of the Suite Test: by integrating source memory tasks, clinicians and researchers can assess cognitive processes with greater ecological validity and adaptability. Moreover, these findings pave the way for designing interventions that leverage source memory to enhance encoding and retrieval strategies, particularly in populations affected by neurodegenerative conditions or age-related cognitive changes.
The findings of this study underscore the relevance of integrating source memory into neuropsychological assessments as an addition to interference paradigms. Interference tasks primarily assess how prior or subsequent information affects recall, often emphasizing memory limitations, particularly in clinical contexts. In contrast, source memory explores both the content and context of memory, providing a more comprehensive perspective.
In conclusion, this study emphasizes the potential of a new VR based memory test that includes a source memory task as a valuable tool in neuropsychological assessment, providing a pathway to explore diverse memory patterns across the lifespan, with the possibility to identify these patterns across clinical and neurotypical populations in upcoming research. Future studies need to emphasize its concurrent validity in comparison with other well-established tests of visual memory, as well as focus on the study of clinical samples that may add valuable information about diagnostic accuracy.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S1355617725101483.
Author contribution
GC, UD, and FR worked in the conceptualization and planning of the paper. JC and UD wrote the first draft of the introduction, while GC, IACG, and FR worked in the methodology and results section. UD, IACG, and FR worked in the Discussion, and GC updated the introduction and discussion, as well as final refinements to the paper. All authors reviewed the final version of the paper.
Funding statement
This work was supported by the European Commission under Horizon 2020 Programme (Grant number 733901, from Project VRMIND – Virtual Reality Based Evaluation of Mental Disorders).
Competing interests
FR and IACG work in the R+D Department of Giunti-Nesplora, publisher of the test.
Open access
