Retrieval-Augmented Generation for Telemedicine: A Privacy-Preserving AI Assistant for Healthcare

The growing demand for telemedicine has intensified challenges such as doctor shortages and data privacy concerns. Generating telemedicine responses requires accuracy, clarity, and coherence, making it a time-consuming task for medical professionals. To address this, we developed a modular Retrieval-Augmented Generation system that leverages a Large Language Model to reduce doctors’ workload. The system operates on-premise, eliminating the need for an internet connection and ensuring data privacy. It integrates a database of 10 million verified PubMed articles, encoded as sparse and dense embeddings, enabling a hybrid vector search to retrieve relevant information based on patient input. Retrieved articles undergo reranking to further refine results, minimizing LLM hallucinations. The system generates telemedicine responses, reducing doctors’ average processing time per patient form from 13.46 to 2.82 minutes – an efficiency gain of nearly 80%. This solution demonstrates significant potential in enhancing telemedicine efficiency and alleviating medical professionals’ workload. Future optimizations, such as integrating additional data sources, offer further opportunities for improvement.