Virtual Patient Chatbot for Medical Training using Small Language Models

Abstract

Medical education is characterized by students learning sufficient diagnostic and communication skills. Nevertheless, they are normally constrained from acquiring access to real patients for ethical or expensive training reasons. This paper presents a virtual patient chatbot that is capable of simulating doctor–patient consultations using Small Language Models (SLMs). The chatbot focuses on Ear, Nose, and Throat (ENT) diseases, with clinical thinking and diagnostic questioning abilities to be exercised by medical students in a safe, controlled, and low-budget environment. For low-end device performance optimization, lightweight models such as TinyLLaMA-1.1B and Gemma-3-1B were experimented with and optimized. A Retrieval-Augmented Generation (RAG) framework was used to improve factual accuracy. Experimental results show that the fine-tuned Gemma-3-1B model achieved the best trade-off between precision and recall. The system demonstrates that SLMs are capable of providing scalable, cost-effective, and realistic medical training devices in resource-poor environments.

Ähnliche Arbeiten