OpenDeID Pipeline for Unstructured Electronic Health Record Text Notes Based on Rules and Transformers:...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Privacy

Background:

• Unstructured electronic health records (EHRs) are rich data sources for clinical and biomedical research.
• Patient privacy necessitates the removal of sensitive health information (SHI) from EHRs before research use.
• While rule-based and machine learning methods exist for de-identification, few studies combine them with transformer models.

Purpose of the Study:

• To develop a hybrid de-identification pipeline for Australian EHR text notes using rules and transformers.
• To investigate the impact of pretrained word embeddings and transformer-based language models on de-identification accuracy.

Main Methods:

• Developed the OpenDeID pipeline, a hybrid approach integrating associative rules, supervised deep learning, and pretrained language models.
• Utilized the OpenDeID Corpus, an Australian multicenter EHR corpus comprising 2100 pathology reports with 38,414 SHI entities.
• Fine-tuned the Discharge Summary BioBERT model and incorporated preprocessing/postprocessing rules.

Main Results:

• Achieved a best F1-score of 0.9659 using the fine-tuned BioBERT model within the OpenDeID pipeline.
• The OpenDeID pipeline has been successfully deployed in a large tertiary teaching hospital.
• Processed over 8000 unstructured EHR text notes in real-time.

Conclusions:

• The OpenDeID pipeline offers an effective hybrid approach for de-identifying sensitive health information in unstructured EHR text notes.
• The pipeline's performance has been validated on a large, multicenter corpus.
• Future work includes external validation to further assess the pipeline's effectiveness.