Data encoding for healthcare data democratization and information leakage prevention

Area of Science:

• Computer Science
• Healthcare Informatics
• Quantum Computing

Background:

• Deep learning in healthcare is limited by data privacy concerns and information leakage from models.
• Current methods struggle to balance data utility for training with robust privacy protection.

Purpose of the Study:

• To propose and evaluate an irreversible data encoding framework for secure and democratized healthcare data.
• To ensure encoded data retains semantic properties for effective deep learning model training.
• To minimize information leakage from trained deep learning models using encoded data.

Main Methods:

• Exploited random projections and random quantum encoding for irreversible data transformation.
• Developed an encoding framework to create an imperceptible data space while preserving semantics.
• Applied the framework to dense and longitudinal/time-series healthcare data.

Main Results:

• Models trained on encoded time-series data adhered to the information bottleneck principle.
• Demonstrated reduced information leakage from deep learning models trained on encoded data.
• Validated the effectiveness of the proposed encoding framework for privacy-preserving deep learning.

Conclusions:

• Irreversible data encoding offers a viable solution for data democratization in healthcare.
• The proposed random projection and quantum encoding methods effectively balance privacy and utility.
• This approach enhances the security and acceptance of deep learning solutions in clinical settings.