Abstract
This column attempts to describe the characteristics of current cyberpsychology research in Europe. In particular, CyberEurope aims at describing the leading research groups and projects running on the other side of the Ocean.
Mild Cognitive Impairment: A Condition between the Present and the Future
T
The syndrome is usually divided into two main subtypes: amnestic and non-amnestic. Amnestic MCI (a-MCI) primarily affects a person's memory, while other cognitive functions, such as language or visuospatial abilities, tend to be preserved. It should be noted, however, that this subtype of MCI can also lead to the deterioration of cognitive domains not exclusively related to memory. In the latter case, it is defined as a-MCI-multiple domains, whereas in the former case, it is defined as a-MCI-single domain. In contrast, non-amnestic MCI (na-MCI) affects attention/executive functions, visuospatial abilities, and language but not memory. Again, a distinction is made between na-MCI single domains if it affects only one cognitive domain other than memory, and na-MCI if it affects multiple domains. In general, people with MCI can present with a very heterogeneous symptomatology. So, the severity of impairment or the type of cognitive domains affected can vary widely among individuals with this disorder.
Old and New Tools for Assessing MCI: Limitations and Current Perspectives
MCI and other forms of cognitive decline are commonly diagnosed using techniques such as labeling specific antigens in the cerebrospinal fluid, functional magnetic resonance imaging, and clinical assessment with cognitive tests (e.g., the Montreal Cognitive Assessment). 3 These tools appear to be effective but have limitations such as invasiveness, high cost, poor ecological validity, and low degree of individualization. 3
Specifically, paper-and-pencil neuropsychological assessments are effective, but they take a long time to administer, require the presence of an experienced and trained professional, are prone to human error, and are also associated with practice effects. 4 In addition, another weakness of cognitive tests is that when used with heterogeneous populations, they have difficulty adapting effectively to different levels of education, ethnicity, and language. 5
To overcome these problems, virtual reality (VR) has been used in recent years to detect MCI and other forms of cognitive decline, with comparable or better results than commonly used paper-and-pencil tools. This technology, regardless of the level of immersion proposed in previous studies (i.e., non-immersive, semi-immersive, and full-immersive VR), creates a sense of agency and comfort in the elderly while increasing the degree of ecological validity. The VR assessments typically use scenarios that correspond to everyday activities (e.g., virtual supermarket tasks, spatial orientation tasks, etc.). As the elderly user interacts with these scenarios, the clinician can assess different cognitive domains (e.g., memory, spatial memory, executive functions) that are typically affected by MCI.
In this case, VR makes it possible to consider and extract new types of behavioral data useful for early detection of cognitive decline, such as average performance time, distance traveled in the VR environment (VRE), and movement patterns performed in the scenario. In addition, VR tools can be used quickly (5–20 minutes for a complete assessment), allowing older adults to use VR and maintain motivation. Furthermore, there is evidence that these tools are well accepted by the elderly population, as they do not often experience complications such as cybersickness. 6
However, VR assessment suffers from several limitations, the most important of which is the lack of automated tools for extracting and evaluating the data collected in the virtual environment, resulting in less efficient classification and diagnosis of cognitive decline. In this context, artificial intelligence (AI) and especially machine learning (ML) have been widely applied in this medical field. ML is a branch of AI that consists of developing computer programs that, instead of being programmed to perform specific tasks, are able to learn from different types of data and patterns to make predictions, identify patterns, and solve problems. 7 ML has been extensively used in the field of medical diagnosis, also moving toward the detection of conditions such as MCI and AD, achieving good results in predicting and assisting medical diagnoses. 8
The AIVRA-MCI Project: Artificial Intelligence and Virtual Reality for the Assessment of MCI
Following on from these considerations, it is clear how VR and ML can be considered as valid tools not only for the identification of MCI but also for the recognition of specific patterns in the collected data, thus facilitating and speeding up the classification and diagnosis of cognitive decline.
The AIVRA-MCI project is a collaboration between the Catholic University of Milan (Milan, Italy) and the University of Pisa (Pisa, Italy). The aim is to promote the early detection of MCI by developing a new VRE capable of guiding elderly people suffering from MCI in the performance of various cognitive tasks. During these tasks, behavioral data related to the elderly's movement will be collected (e.g., path tracing, time of performance, speed, degree of headset rotation, etc.), together with information from the cognitive tasks that the participants have performed in the VRE. The collected data will then be used to train different ML models, with the goal of identifying specific patterns in the results and predicting MCI.
The combination of VR and ML for the detection of MCI could lead to the development of an innovative tool with greater ecological validity, lower cost, and reduced administration time and degree of invasiveness than commonly used tools for the diagnosis of this condition.
Step 1: Creation, Familiarization, and Testing of a New VRE for Older Adults with MCI
The first step of this project will be divided into three subphases:
Subphase 1: In this subphase, we will create a VRE that is as similar as possible to a typical elderly home in order to increase the ecological validity of the study. We will create the scenario using Unreal and Unity software and set it up to record and extract different types of behavioral data (i.e., different movements that the elderly will make during the tasks), such as eye tracking/heat map, route tracking, average speed, position of joysticks in VR space, total distance traveled, and so on. Then, after the VRE is completed and tested, we will recruit a sample of elderly with a-MCI and na-MCI and a control group of healthy elderly without cognitive decline.
Subphase 2: In this subphase, all three groups (i.e., amnestic, non-amnestic, and healthy participants) will attend a session to familiarize themselves with the VR equipment. During the first session, the experimenters will explain to the participants how the VR setup works—how to use the joysticks and how to wear the head-mounted display—thus giving them the opportunity to move around and interact with a VRE similar to the one that will be used in the next phase of the test (i.e., a house composed of several rooms). The interactions with the VRE that the elderly will perform in this and the following subphase will be simple in order to facilitate the learning of the use of VR. During this initial meeting, each senior will be closely supervised by the experimenter, both to prevent them from injuring themselves while using the VR equipment and to answer any doubts or questions they may have in an effort to make the subsequent testing phase as smooth as possible.
Subphase 3: After the familiarization phase with the VR equipment, the three groups will attend a second session to participate in the testing phase. During this session, they will be immersed in a virtual house consisting of a living room, kitchen, balcony, bedroom, and entrance hall.
The testing phase will then be divided into:
Subphase 1: This will be an exploration and interaction phase. The participants will freely explore the different areas of the house, interacting with the objects present there. In this way, they will visualize the location of objects and associate them with a specific place in the house.
Subphase 2: This will be a memorization and orientation phase. Next, the participants will be given a list of objects to find in the house, such as a toothbrush, a glass cup, blankets, and so on. The list will show both the picture and the written name of the object to facilitate memorization. After a few minutes, the displayed list will no longer be available to the participants, and the elderly will start moving around the VRE to search the house for all the objects that were on the list.
These tasks will assess some cognitive domains normally affected by MCI, such as memory (spatial and episodic), executive functions (attention, planning), and spatial orientation. During administration, behavioral data and cognitive task scores (i.e., items selected correctly, items selected incorrectly, items missing) will be collected.
Step 2: Building and Evaluating an ML Capable of Predicting MCI
Once the testing and data collection phase is complete, the raw data (behavioral and cognitive) will be subjected to several processing steps in order to train the ML properly.
First, the raw data will undergo a preprocessing phase to clean the dataset from any incorrect, duplicate, or missing data. After that, we will normalize all the features (i.e., z-score, min–max normalization) to put them all on the same scale and make the data more homogeneous.
Once we have prepared the dataset, we will continue to select the most appropriate ML models for this type of data, such as support vector machine (SVM), random forest (RF), or neural networks (NN). Then, we will divide the dataset into two parts: a training dataset to train the different ML models, and a test dataset to evaluate the performance of the chosen models.
Through the training dataset, we will train the models to learn to recognize specific patterns in the data that correspond to the condition we are interested in (i.e., a-MCI, na-MCI, and healthy people). We will then use the test dataset to evaluate the predictive capabilities of the trained models. In this part, we will use some metrics to evaluate the ML performance, such as precision, which evaluates how many of the positive predictions made by the model are actually correct; recall, which evaluates the proportion of positive instances correctly predicted by the model compared to the total number of actual positive instances; and F1 score, which allows an overall measure of model performance to be obtained through a weighted average of precision and recall.
In addition to these steps, we will also consider the occurrence of overfitting. This phenomenon corresponds to a state in which the ML overfits the data on which it was trained and then develops difficulties in correctly generalizing the predictions to new data. For this reason, we will use cross-validation strategies, which may consist, for example, of splitting the available data into different parts that are used to train and evaluate the model iteratively.
Finally, where possible, we will also use explicable AI models, such as SHAPE, to understand the weight of each feature and to understand better the decision-making process implemented by the models. This will help us to understand which features are most important for predicting MCI.
Conclusion
In conclusion, this project aims to expand the limited field of study that combines the use of VR tools and ML models for the detection of conditions such as MCI. Thanks to this, early prevention will potentially slow down the deterioration process that can lead to the development of worse clinical conditions such as AD.
The novelty of this study lies in the combination of VR and ML, in the creation of a new type of VRE, and in the collection and analysis of a large amount of behavioral data that, to our knowledge, has not been used in previous studies.
As pointed out by Chiara Stramba Badiale, a neuropsychologist working on the assessment and rehabilitation of elderly people with cognitive decline, “The AIVRA-MCI project introduces a novel approach by combining VR-based assessments and ML techniques to improve the accuracy, efficiency and accessibility of MCI detection and diagnosis. Indeed, VR provides a sense of control, comfort, and ecological validity by simulating everyday activities and providing positive feedback. ML models have the ability to identify patterns in scores, allowing for faster classification and diagnosis of cognitive decline. The combination of VR and ML could provide a low-cost alternative to traditional diagnostic tools.
Overall, future directions aim to refine, validate, and implement the innovative VR-ML approach for MCI detection, potentially changing the way cognitive decline is identified and rehabilitated in clinical settings.”