Evaluating the effects of adjunct VR-haptic simulator practice in preclinical restorative training and predicting student performance using machine learning: a randomized controlled trial

Although undergraduate preclinical training is commonly carried out on phantom mannequins, virtual reality (VR) devices have started to be used as supplementary tools. This study aimed to investigate the impact of adjunct VR-haptic simulator practice on undergraduate dental education in restorative practice and to explore whether machine-learning models can predict faculty-assigned letter grades. A total of 100 undergraduate students participated in the study, with 50 receiving 8 hours of phantom training only, while the other 50 received 8 hours of phantom training plus an additional 4 hours of VR-haptic simulator training. All students prepared a Class I cavity, which was graded independently by two experienced lecturers. (100-point scale). Inter-rater agreement was quantified using the intraclass correlation coefficient (ICC; two-way random-effects, absolute agreement, average measures; ICC(2,2) = 0.85, (95% CI: 0.72-0.86)). The effect of dental simulator use on cavity performance was analyzed. Twenty-six features were extracted to predict students’ letter grades using four different machine learning methods (Naive Bayes, decision tree, random forest (RF), and support vector machines). In addition, the result of the best-performing method was explained using Shapley additive explanations (SHAP) to describe feature-prediction associations (not causality). The best scores were obtained with the RF (Accuracy=0.8273, Macro-F1-score=0.8205, Macro average precision=0.8304, Macro average recall=0.8309). The SHAP explainability results indicate that the most deterministic features are the occlusal outline, simulator usage, and point angle. These findings suggest that simulator usage significantly impacts student performance, contributing to improved grades. In preclinical restorative training, adjunct VR-haptic simulator exposure was found to be associated with faculty-assigned performance grades on a Class I cavity preparation task. Machine-learning models can be employed to predict the letter grades assigned by faculty members.