Active deep learning to detect demographic traits in free-form clinical notes

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Privacy

Background:

• Clinical notes contain valuable research data but require de-identification to protect patient privacy.
• Existing de-identification methods focus on direct identifiers, potentially leaving residual demographic traits (DTs).
• Residual DTs, such as marital status or housing type, can pose re-identification risks when combined.

Purpose of the Study:

• To investigate residual re-identification risks in de-identified clinical notes.
• To identify and quantify the prevalence of demographic traits (DTs) after standard de-identification.
• To develop and evaluate an effective method for detecting and potentially redacting DTs.

Main Methods:

• Manual annotation of over 140,000 words of clinical notes to identify DTs.
• Development of an annotation guide for Demographic Traits (DTs).
• Implementation of a bootstrapped active learning iterative process using a BERT-based classifier for DT identification.
• Training and validation on MIMIC-III and i2b2-2006 clinical note datasets.

Main Results:

• No directly identifying information was found in the investigated medical notes.
• A low prevalence of demographic traits (DTs) was observed.
• The active learning approach significantly improved classifier accuracy for DT detection.
• BERT-based models outperformed non-neural models in identifying DTs.

Conclusions:

• Directly identifying information is virtually absent in de-identified clinical notes.
• Demographic traits (DTs) are present but detectable with high accuracy.
• A cost-effective human-in-the-loop active learning process can effectively identify and redact DTs, enhancing patient privacy.