Enabling qualitative research data sharing using a natural language processing pipeline for deidentification: moving...

Area of Science:

• Health Informatics
• Computational Linguistics
• Qualitative Research Methods

Background:

• Sharing qualitative health research data is vital for translating findings into practice but is hindered by deidentification challenges and re-identification risks.
• Existing deidentification systems often fail to capture non-HIPAA Safe Harbor (HSH) identifiers prevalent in unstructured qualitative data.

Purpose of the Study:

• To establish and evaluate a framework for deidentifying qualitative health research data using automated computational techniques.
• To develop and validate a natural language processing (NLP) pipeline capable of identifying and removing both HSH and non-HSH identifiers.

Main Methods:

• Developed an NLP pipeline employing named-entity recognition, pattern matching, dictionaries, and regular expressions.
• Analyzed and qualitatively reviewed diverse qualitative health research datasets to inform pipeline development.
• Validated the pipeline using a gold-standard dataset of 280,000 words from 70 files.

Main Results:

• The NLP deidentification pipeline achieved a consistent F1-score of approximately 0.90 across two datasets totaling 1.2 million words.
• Demonstrated that the majority of identifiers in qualitative data are non-HSH and not addressed by current systems.
• Successfully identified and removed both HSH and non-HSH identifiers from qualitative research texts.

Conclusions:

• NLP methods provide an effective solution for deidentifying qualitative health research data, encompassing both standard and non-standard identifiers.
• Automated deidentification tools are essential for researchers, particularly in light of new data-sharing mandates from organizations like the National Institutes of Health (NIH).
• This framework facilitates more secure and widespread sharing of qualitative health research data, accelerating knowledge translation.