Large Language Models as a Virtual Endodontic Patient. Can They Help Us Teach Endodontic Diagnosis?

In dental education, standardized patients (SPs) [1], played by actors, are commonly used to train students in diagnostic and communication skills. While effective, these methods can be resource-intensive and limited in their availability. Chatbots powered by Generative Artificial Intelligence (GenAI), such as ChatGPT, present a potential solution. Trained through massive textual data models known as Large Language Models (LLMs), these GenAI chatbots are able to interpret and generate human-like language based on patterns learned from this training data. In an educational context, GenAI chatbots can serve as virtual patients, offering a scalable way to provide dental students with real-time interactions and feedback. The key question is whether LLMs can accurately and realistically simulate a patient's responses in a way that enhances dental education while aligning with curriculum goals and improving students' learning outcomes. To address this problem, we explored the performance of a GenAI chatbot as a virtual patient. LLMs such as GPT-4, the model behind ChatGPT, can be fine-tuned to incorporate specific information that brings desired specificity to their responses. We fine-tuned the model with an endodontic clinical scenario, including a radiograph and a "patient persona", a set of specifications regulating the style of the expected responses (Figure 1). The clinical case was not revealed to the student before the activity. The student was instructed to interact with the chatbot through text, simulating a conversation with an actual patient (Figure 2). The student was instructed to initiate the conversation, and ChatGPT responded to each student's questions, establishing a patient interaction through which the student was able to determine objective and subjective clinical findings. Two endodontic faculty provided synchronous feedback during the interaction or asynchronous feedback by later reviewing recorded transcripts directly through the chatbot interface. A pilot program was performed with third- and fourth-year dental students and post-graduate endodontic students. Students and faculty provided positive feedback regarding the natural feel of the conversations with virtual patients. Students reported that the patient interaction felt realistic and that it helped increase their engagement in the training process. Endodontic instructors valued the ability to review student interactions and assess their clinical assessment skills. This allowed instructors to provide targeted feedback based on the specific questions asked by students during their conversations with the virtual patient. ChatGPT allows for continuous interactions with students, providing 24/7 access to training cases. This approach is scalable and cost-effective, and it provides real-time feedback on student performance while offering the flexibility needed for self-directed learning. However, a challenge emerged: ChatGPT was not 100% accurate in providing consistent answers. While the system generated functional responses, answers were inconsistent between sessions, and some "hallucinations" [2] were observed. This variation in responses emphasized the need for further refinement and benchmarking of different LLMs. A comprehensive comparison of several LLMs will help identify the model that delivers the most accurate and consistent answers. Ensuring the accuracy and reliability of the responses is crucial to the effectiveness of this technology in dental education. The authors declare no conflicts of interest.