Sensitive Data Detection with High-Throughput Machine Learning Models in Electrical Health Records.

Area of Science:

• Health Informatics
• Data Science
• Machine Learning

Background:

• The Health Insurance Portability and Accountability Act (HIPAA) protects sensitive health information but lacks efficient de-identification tools.
• Heterogeneous data structures across healthcare entities challenge rule-based PHI detection.
• Secure data sharing is crucial for improving health outcomes and advancing research.

Purpose of the Study:

• To develop a machine learning approach for automatically identifying protected health information (PHI) in structured Electronic Health Record (EHR) data.
• To address the limitations of rule-based systems in detecting variable PHI fields across different datasets.
• To facilitate the de-identification process for enhanced data sharing and research collaboration.

Main Methods:

• Engineered over 30 features from the metadata of structured EHR data.
• Utilized a novel observation of differing metadata distributions between PHI and non-PHI fields.
• Developed and trained machine learning classification models on diverse EHR databases.

Main Results:

• Achieved 99% accuracy in detecting PHI-related fields on unseen datasets.
• Demonstrated the effectiveness of metadata-based feature engineering for PHI identification.
• Validated the algorithm's performance across various large EHR databases from different sources.

Conclusions:

• Machine learning offers a robust solution for automated PHI identification in structured EHR data.
• The developed method significantly enhances the de-identification process, enabling secure data sharing.
• This approach has broad implications for industries handling sensitive data, improving data security and research capabilities.