Decentralised, collaborative, and privacy-preserving machine learning for multi-hospital data.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Privacy

Background:

• Machine learning (ML) models in healthcare require large, diverse datasets for accuracy and generalizability.
• Sharing medical data across institutions is hindered by privacy regulations and logistical challenges.
• Collaborative ML training is crucial but difficult without compromising data privacy.

Purpose of the Study:

• To propose a decentralized, collaborative, and privacy-preserving ML framework for multi-hospital data analysis.
• To enable multi-party ML model training without direct data sharing or centralization.
• To safeguard patient privacy during collaborative model development.

Main Methods:

• Introduced Decentralized, Collaborative, and Privacy-preserving ML for Multi-Hospital Data (DeCaPH).
• Enabled collaborative training without data transfer, ensuring no data centralization.
• Implemented privacy safeguards to limit information leakage during training.
• Facilitated training without reliance on a central server.

Main Results:

• DeCaPH demonstrated generalizability on patient mortality prediction, cell-type classification, and pathology identification using real-world medical data.
• Models trained with DeCaPH showed <3.2% performance drop compared to non-privacy-preserving methods.
• Average vulnerability to privacy attacks decreased by up to 16%.
• DeCaPH models outperformed models trained solely on private data (up to 70% improvement) and previous privacy-preserving methods (up to 18.2% improvement).

Conclusions:

• DeCaPH improves the utility-privacy trade-off, enabling high-performance models while preserving data privacy.
• DeCaPH enhances model generalizability by facilitating collaboration across individual datasets.
• The framework effectively supports privacy-preserving collaborative machine learning in multi-institutional healthcare settings.