Automatic Detection of Personally Identifiable Information in Medical Data Using Large Language Model

Abstract

In this study, we proposed a method for the automatic detection of personally identifiable information from various formats of medical data using a locally operated large language model (LLM). We performed optimized text extraction for each data format, then input the extracted text into Llama3 for detection. In addition, we also performed fine-tuning using LoRA and compared performance with the base model. The base model achieved a high detection rate on text data, but the detection rate decreased on image data due to misrecognition by Easy-OCR. While the output format improved after fine-tuning, the detection rate for patient IDs significantly decreased.