While there exist numerous validated neuropsychological tests and batteries to measure cognitive and behavioral capacities, the vast majority of these are time intensive and difficult to administer and score outside of the clinic. Moreover, many existing assessments may have limited ecological validity in some contexts (e.g., military operations). Therefore, we have been developing a novel approach to administering neuropsychological assessment using a virtual reality (VR) “game” that will collect simultaneously acquired multidimensional data that is synthesized by machine learning algorithms to identify neurocognitive strengths and weaknesses in a fraction of the time of typical assessment approaches. For our initial pilot project, we developed a preliminary VR task that involved a brief game-like military “shoot/no-shoot” task that collected data on hits, false alarms, discriminability, and response times under a context-dependent rule set. This prototype task will eventually be expanded to include a significantly more complex set of tasks with greater cognitive demands, sensor feeds, and response variables that could be modified to fit many other contexts. The objective of this project was to construct a rudimentary pilot version and demonstrate whether it could predict outcomes on standard neuropsychological assessments.

To demonstrate proof-of-concept, we collected data from 20 healthy participants from the general population (11 male; age=24.8, SD=7.8) with high average intelligence (IQ = 112, SD=10.7). All participants completed the Wechsler Abbreviated Scale of Intelligence-II (WASI-II), and several neuropsychological tests including the ImPACT, the Attention and Executive Function modules of the Neuropsychological Assessment Battery (NAB), and the VR task. Initially, we used a prior dataset from 359 participants (n=191 mild traumatic brain injury; n=120healthy control; n=48 sleep deprived) to serve as a training sample for machine learning models. Based on these outcomes, we applied machine learning, as well as standard multiple regression approaches to predict neuropsychological outcomes in the 20 test participants.

In this limited study, the machine learning approach did not converge on a meaningful prediction due to the instability of the small sample. However, standard multiple linear regression using stepwise entry/deletion of the VR task variables significantly predicted neuropsychological performance. The VR task predicted WASI-II vocabulary (R=.457, p=.043), NAB Attention Index (R=.787, p=.001), and NAB Executive Function Index (R=.715, p=.002). Interestingly, these performances were generally as good or better than the predictions resulting from the ImPACT, a commercially available neuropsychological test battery, which correlated with WASI-II vocabulary (R=.557, p=.011), NAB Attention Index (R=.574, p=.008), and NAB Executive Function Index (R=.619, p=.004).

Our pilot VR task was able to predict performances on standard neuropsychological assessment measures at a level comparable to that of a commercially available computerized assessment battery, providing preliminary evidence of concurrent validity. Ongoing work is expanding this rudimentary task into one involving greater complexity and nuance. As multivariate data integration models are incorporated into the tasks and extraction features, future work will collect data on much larger samples of individuals to develop and refine the machine learning models. With additional work this approach may provide an important advance in neuropsychological assessment methods.