Understanding Learning and Predictive Factors for Performance of AR-Based Clinical Tasks: Quantitative Performance Analysis (Preprint)

BACKGROUND Augmented Reality (AR) offers a risk-free training modality for medical trainees. However, little is known about the effect that prior experiences with technology and visuo-spatial ability have on performance in AR. Understanding these relationships and the learning curve that users face is vital to designing effective AR training programs. OBJECTIVE We aimed to identify predictors of performance in novice AR users. We hypothesized that individuals with higher visuo-spatial ability and greater technology experience will perform AR-based tasks quicker and with fewer errors. METHODS Participants were composed of undergraduate, graduate, and medical students between June and December 2024. Participants with prior AR experience were excluded. A mixed-methods framework was employed including a survey on previous experience, a test of visuo-spatial ability Shepard and Metzler mental rotation task (MRT), a series of AR tasks, and the NASA-TLX cognitive load assessment. Participants performed a standardized seven task AR protocol with Microsoft HoloLens 2 that mimicked clinical use cases. Participants were instructed to resize and rotate a hologram of a human skull model (task 1), outline the orbits on the hologram (task 2), look through coronal, sagittal, and axial views of the hologram (task 3), place virtual trajectory markers on the hologram (task 4), place virtual landmarks on the hologram (task 5) and physical landmarks on the 3D-printed human skull model (task 6), and perform a trajectory alignment (task 7). Performance was analyzed using metrics such as time to completion, number of slips, and quality of tracing. RESULTS Twenty-one participants completed the task and were included in the analysis. Performance on the MRT did not predict baseline performance time (p=0.54) nor error rates in AR tasks (p=0.62). Participants with extensive video game experience (≥ 5 hours/week) performed fewer slips compared to participants with minimal experience (p=0.046), despite no time improvement (p=0.24). Video game experience was also not a predictor of baseline performance in AR (p=0.12). Participants demonstrated a clear learning effect in the baseline performance task (task 1). Participants had significantly faster task completion between attempts 1 and 2 (p CONCLUSIONS Here, we show that visuo-spatial ability does not predict performance of clinically relevant AR tasks. Interestingly, we observed a clear learning effect, demonstrated by a reduction in task completion times along with improved accuracy. These findings may support the potential for the integration of AR into clinical workflows that require fast adoption and precise task execution.