Benchmarking Modern Named Entity Recognition Techniques for Free-text Health Record Deidentification

Area of Science:

• Health Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic Health Records (EHRs) are the standard for medical data in the US.
• Federal regulations restrict sharing of EHR data containing Protected Health Information (PHI).
• De-identification is essential for enabling public access to EHR data for research.

Purpose of the Study:

• To evaluate deep learning-based Named Entity Recognition (NER) methods for EHR de-identification.
• To determine the optimal NER approach for removing PHI from clinical text.

Main Methods:

• Training and testing various deep learning NER models on the i2b2 dataset.
• Utilizing a Bi-LSTM-CRF architecture as a primary model.
• Assessing model performance qualitatively on local hospital EHR data.

Main Results:

• The Bi-LSTM-CRF model demonstrated superior performance as an encoder/decoder combination for de-identification.
• Character embeddings enhanced precision but reduced recall in PHI detection.
• Transformer models, when used alone as context encoders, showed suboptimal performance.

Conclusions:

• Bi-LSTM-CRF is a highly effective method for EHR de-identification.
• Further research into structuring medical text could improve semantic and syntactic information extraction for de-identification.