De-identification of free-text clinical notes

Abstract

Clinical notes contain rich information that is useful in medical research and investigation. However, clinical documents often contain explicit personal information that is protected by federal laws. Researchers are required to remove these personal identifiers before publicly release the notes, a process known as de-identification. In recent years, the healthcare community has initiated several competitions to expedite the development of automated de-identification systems. Notably, models built using recurrent neural networks achieved state-of-the-art performance on the de-identification task. Since the competition, new architectures based on transformers have been developed with excellent performance on general domain natural language processing tasks. Examples include BERT and RoBERTa. In this work, we evaluated de-identification using different choices of bidirectional transformer models and classifiers. Further, we developed a hybrid system that incorporates rule-based features into the bidirectional transformer model. Our results demonstrated state-of-the-art performance with an average 98.73% binary token F1 score, a 0.45% increase from current baseline models.