Optimizing healthcare delivery: leveraging large language models to do pre-consultation interactions

Abstract

Medical resources are scarce, especially in densely populated areas where there is a high demand for hospital diagnoses. Often, there are long queues for appointments, which can be frustrating for patients. If patients could express their symptoms effectively and receive preliminary diagnoses promptly, it would free up significant medical resources and enhance the efficiency of patients gaining the right knowledge to manage their conditions. This article leverages the open-source MetaAI language model LLaMA2 as the base model. Firstly using TF-IDF Vectorization and Cosine Similarity Filtering to finish text de-duplicates job. The model utilizes Huggingface's PEFT Library to perform LoRA fine-tuning technique and Quantization Techniques to fine-tune LLaMA2 on NVIDIA 3090 device, enabling it to make initial symptom assessments based on patient descriptions. Finally, an attempt will be made to have this Language Model (LM) participate in the USMLE examination for evaluation.