A comprehensive tool for creating and evaluating privacy-preserving biomedical prediction models.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Privacy

Background:

• Data-driven medical research utilizes machine learning for clinical decision support.
• Sensitive personal information in medical data poses privacy risks.
• Existing methods lack practical tools for creating and evaluating privacy-preserving machine learning models.

Purpose of the Study:

• To bridge the gap in practical tools for privacy-preserving machine learning in biomedicine.
• To develop and evaluate methods for creating accurate, privacy-preserving prediction models from clinical data.
• To provide an open-source solution for researchers in the field.

Main Methods:

• Extended the ARX anonymization tool with machine learning techniques.
• Integrated various privacy protection methods (k-anonymity, differential privacy, game-theoretic approach).
• Developed a versatile framework with graphical user interfaces for model creation, evaluation, and refinement.

Main Results:

• Created accurate, privacy-preserving prediction models without adding noise to data.
• Demonstrated applicability through case studies in breast cancer diagnosis, urinary system inflammation, and contraceptive method prediction.
• Supported binomial and multinomial target variables and diverse prediction models.

Conclusions:

• The developed tool enables the creation of accurate prediction models that preserve individual privacy.
• The open-source implementation offers a versatile solution for various threat scenarios in biomedical research.
• The methods are intuitive, explainable to non-experts, and preserve data truthfulness.