CUBA: A Conversational AI for Medical Education

CUBA is a Large Language Model based Virtual Patient Simulator designed as a training system for medical students to improve their communication skills and practice on their diagnostic reasoning skills through interactive conversations with the virtual patients. The system is build on two strong base models namely ClinicalBERT and Phi-3-mini. ClinicalBERT, fined tuned on medical questions, which is used to check the relevance of questions asked by the interns. It reached an accuracy of 97.8% and F1 score of 98.7%. Only relevant queries are forwarded to the Phi-3-mini model, which is fine-tuned using LoRA (training loss 0.29) to generate realistic patient responses. The System integrates voice-based interaction and multilingual support using the Google Translate API. After each conversation, interns are required to predict the patient primary diagnosis, and the system then provides feedback regarding the correctness of prediction and gives a explanatory guidance. Overall, CUBA offers a safe and interactive environment for improving AETCOM, clinical questioning, and diagnostic reasoning skills among medical trainees.