An Automated Virtual Reality Training System for Teacher-Student Interaction: A Randomized Controlled Trial

Abstract

Background Shortages in qualified supervision and other resources prevent education personnel from rehearsing effective practices. Interactive simulations, although increasingly used in education, frequently require instructor management. Automated simulations rarely engage trainees in skills related to practice (eg, speech). Objective We evaluated the capability of delivering behavioral skills training through an automated virtual reality (VR) simulation using artificial intelligence to improve the implementation of a nondirective mathematical questioning strategy. Methods We recruited and randomly assigned 30 college-aged participants to equivalent treatment (ie, lecture, modeling, and VR; 15/30, 50%) and control groups (ie, lecture and modeling only; 15/30, 50%). The participants were blind to treatment conditions. Sessions and assessments were conducted face to face and involved the use of VR for assessment regardless of the condition. Lessons concerned the use of a nondirective mathematical questioning strategy in instances where a simulated student provided correct or incorrect answers to word problems. The measures included observed and automated assessments of participant performance and subjective assessments of participant confidence. The participants completed the pretest, posttest, and maintenance probes each week over the course of 3 weeks. Results A mixed ANOVA revealed significant main effects of time (F2,27=124.154; P<.001; ηp2=0.816) and treatment (F1,28=19.281; P<.001; ηp2=0.408) as well as an interaction effect (F2,28=8.429; P<.001; ηp2=0.231) for the average percentage of steps in the questioning procedure. Posttest scores for the intervention group (mean 88%, SD 22.62%) exceeded those of the control group (mean 63.33%, SD 22.64%), with t28=3.653, P<.001, and Cohen d=1.334. Maintenance scores indicated a positive effect of the intervention (mean 83.33%, SD 24.40%) relative to the control (mean 54.67%, SD 15.98%), t28=3.807, P<.001, Cohen d=1.39. A Mann-Whitney U test indicated that the treatment groups’ self-ratings of confidence (mean 2.41, SD 0.51) were higher than those of the control group (mean 2.04, SD 0.52), U=64, P=.04, r=0.137. Conclusions The results demonstrate the potential of artificial intelligence-augmented VR to deliver effective, evidence-based training with limited instructor management. Additional work is needed to demonstrate the cascading effect of training on authentic practice and to encompass a wider range of skills.