Learning to identify Protected Health Information by integrating knowledge- and data-driven algorithms: A case study...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• De-identification of clinical narratives is crucial for enabling healthcare data research.
• Existing tools require expansion and tailoring to effectively process diverse clinical text.
• Psychiatric evaluation notes present unique challenges for automated de-identification.

Purpose of the Study:

• To evaluate and improve de-identification methods for clinical narratives.
• To adapt existing tools for a shared task focused on psychiatric notes.
• To assess the performance of machine learning approaches in identifying various types of Protected Health Information (PHI).

Main Methods:

• Utilized and expanded two existing de-identification tools.
• Applied machine learning models on both large and small feature spaces.
• Implemented additional strategies such as two-pass tagging and multi-class models.
• Evaluated methods on a dataset of psychiatric evaluation notes for up to 25 PHI types.

Main Results:

• Achieved overall F1-scores of approximately 90% and above for identifying Health Information Portability and Accountability Act (HIPAA) defined PHIs.
• Demonstrated the effectiveness of machine learning with enhanced feature spaces and multi-class strategies.
• Identified specific PHI classes, such as Profession and Organization, as persistently challenging due to expression variability.

Conclusions:

• The integrated methods significantly improve the de-identification of clinical text, facilitating data sharing for research.
• Machine learning approaches, particularly with tailored strategies, are highly effective for PHI detection.
• Further research is needed to address the variability in identifying certain information types like professions and organizations.