Enhancing Automated De-identification of Pathology Text Notes Using Pre-trained Language Models

With the recent advancements in artificial intelligence (AI), particularly the surge in large language models (LLMs) led by industry giants such as OpenAI, Microsoft, and Google, a new era in clinical medicine and digital health has been ushered in. These advancements, transformative as they are, introduce significant privacy concerns, particularly in the healthcare sector, which heavily relies on Electronic Health Records notes (EHRs) text notes. Our research directly addresses these privacy challenges by enhancing the T5-Efficient-BASE-DL2 model for the de-identification of patient data. We fine-tuned the T5 model with specific hyperparameter adjustments and implemented exact match evaluations to significantly reduce false positives and improve the accuracy of entity recognition, particularly for sensitive information processing. This was achieved using regular expressions and the Pycountry library. Our enhancements resulted in a notable improvement in the model’s performance, with a 14% increase in the macro-average F1-score for Task 1, rising from 0.7787 to 0.9187, and a 30% improvement for Task 2, from 0.4871 to 0.79. These advancements not only enhance the security and efficiency of healthcare information management but also uphold the integrity of individual privacy and facilitate the safe sharing of data for research purposes without compromising confidentiality. Our research lays the groundwork for future developments in improving rule-based methodologies and expanding datasets, aiming to further integrate machine learning with linguistic rules for more effective data privacy solutions in healthcare.