Classifying Cyber-Risky Clinical Notes by Employing Natural Language Processing

Area of Science:

• Health Informatics
• Natural Language Processing
• Cybersecurity

Background:

• Clinical notes in electronic medical records are vital for patient care and research.
• Increasing patient access to clinical notes necessitates robust data protection methods.
• Existing de-identification techniques inadequately address sensitive information risk classification.

Purpose of the Study:

• To investigate methods for identifying and classifying security/privacy risks within clinical notes.
• To develop models that can detect sensitive information for improved data protection.
• To aid in identifying clinical notes that may require further de-identification.

Main Methods:

• Developed and tested several machine learning models using unigram and word2vec features.
• Employed different classifiers to categorize sentence-level risk.
• Utilized the i2b2 de-identification dataset for experimental evaluation.

Main Results:

• The Support Vector Machine (SVM) classifier with word2vec features achieved a maximum F1-score of 0.792.
• The models demonstrated effectiveness in categorizing sentence risk within clinical notes.
• Identified specific features and classifiers that perform well for risk assessment.

Conclusions:

• The developed models offer a promising approach to assessing and mitigating cyber risks in clinical notes.
• Classifying sensitive information risk is a fundamental step towards safeguarding patient health information.
• Future work should align risk assessment with global regulatory requirements.