Detecting Protected Health Information in Heterogeneous Clinical Notes

Area of Science:

• Health Informatics
• Natural Language Processing
• Medical Data Privacy

Background:

• Secondary use of healthcare data requires privacy-preserving methods.
• Automatic identification of protected health information (PHI) in clinical text is crucial.
• Machine learning models offer an alternative to rule-based systems for PHI detection.

Purpose of the Study:

• To investigate the impact of domain differences on PHI detection models.
• To assess the performance of a PHI detection model across various clinical note types.
• To determine if domain adaptation is necessary for heterogeneous clinical notes.

Main Methods:

• Trained a predictive model on an existing Swedish clinical notes PHI corpus.
• Applied the trained model to diverse clinical notes from different specialties, headings, and professions.
• Analyzed model performance variations across these different domains.

Main Results:

• Significant performance differences were observed when applying the model to different clinical note domains.
• The prevalence of PHI varies across clinical specialties, note headings, and authoring professions.
• A model trained on one domain does not generalize effectively to others without adaptation.

Conclusions:

• Domain adaptation is essential for robust and accurate PHI detection in varied clinical text.
• Future research should focus on developing domain-adaptive strategies for PHI identification.
• Effective PHI detection supports secure secondary data use in healthcare.