Exploring the Privacy-Preserving Properties of Word Embeddings: Algorithmic Validation Study.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Data Privacy

Background:

• Word embeddings are crucial for neural networks in representing language.
• Publicly available clinical word embeddings were previously nonexistent.
• This study is the first to investigate the privacy risks associated with releasing clinical word embedding models.

Purpose of the Study:

• To demonstrate that traditional word embeddings from deidentified clinical text can expose sensitive patient data.
• To explore the privacy implications of different word embedding methods applied to clinical corpora.

Main Methods:

• Utilized word embeddings generated from 400,000 doctor-written consultation notes.
• Experimented with three common word embedding techniques.
• Assessed the privacy-preserving capabilities of each method.

Main Results:

• Reconstructed up to 68.5% of patient names from embeddings trained on deidentified data.
• Linked sensitive information to specific patients within the corpus.
• Found that vector distances between patient names and billing codes reveal diagnostic information.

Conclusions:

• Current methods of deidentifying clinical text for word embeddings pose significant patient privacy risks.
• Personal health information (PHI) removal alone is an imperfect anonymization strategy.
• Anonymization by PHI replacement may offer a more robust privacy-preserving alternative for clinical word embeddings.