Impact Analysis of De-Identification in Clinical Notes Classification.

Area of Science:

• Health Informatics
• Natural Language Processing
• Data Privacy

Background:

• Clinical notes in telemonitoring are valuable for disease management.
• Automated analysis requires structured data, often achieved through text classification.
• De-identification is crucial for privacy but may affect data utility.

Purpose of the Study:

• To investigate the impact of de-identification on text classification accuracy.
• To evaluate how different pseudonymization and classification algorithms perform on de-identified clinical text.

Main Methods:

• Applied two pseudonymization and two classification algorithms to telehealth clinical messages.
• Measured the divergence in classification results between de-identified and clear text.

Main Results:

• De-identification significantly altered classification outcomes.
• A sensitive classification algorithm showed notable performance degradation, particularly in sensitivity.
• A simpler classification method demonstrated robustness, with negligible impact when paired with a suitable pseudonymization technique.

Conclusions:

• De-identification can substantially affect text classification performance.
• Developing classification methods with de-identification considerations is recommended for telemonitoring systems.
• Balancing privacy preservation and data utility is key for effective clinical text analysis.