De-Identification of Electronic Medical Records Using Large Language Models: A Case Study

De-identification of electronic health records (EHRs) is essential to safeguard patient privacy, particularly when such data are utilized for large-scale population studies. In this study, we assess the performance of locally deployed large language models (LLMs) both general-purpose and domain-specific in de-identifying maternal EHRs. To evaluate general-purpose LLMs, we employ a combination of prompt engineering techniques and post-processing strategies aimed at improving the accuracy and robustness of the anonymization process. These models are then compared against a domain-specific LLM fine-tuned for de-identification tasks. Our results show that general-purpose LLMs exhibit significant limitations in performing named entity recognition (NER), underscoring the advantages of task-specific tuning for sensitive medical applications.