Privacy-Preserving Federated Survival Support Vector Machines for Cross-Institutional Time-To-Event Analysis:...

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Centralized patient data collection faces privacy challenges, limiting large-scale clinical studies.
• Federated learning offers a privacy-preserving solution for distributed medical data analysis.
• Large sample sizes are crucial for time-to-event studies but often unavailable at single institutions.

Purpose of the Study:

• To develop and validate a privacy-preserving federated survival support vector machine (SVM).
• To enable cross-institutional time-to-event analyses for researchers.
• To provide an accessible tool for federated survival analysis.

Main Methods:

• Extended the survival SVM algorithm for federated environments.
• Implemented the federated survival SVM as a FeatureCloud app.
• Evaluated the algorithm on synthetic and real-world microbiome datasets, comparing it to a central model.

Main Results:

• The federated survival SVM yielded highly similar results to the centralized model (max weight difference of 0.001).
• Federated learning improved prediction accuracy by incorporating more data, even with site-specific batch effects.
• The approach demonstrated robustness across benchmark datasets.

Conclusions:

• The federated survival SVM enhances federated time-to-event analysis with a robust machine learning method.
• The FeatureCloud app is the first publicly available federated survival SVM, freely accessible to researchers.
• This tool facilitates direct use within the FeatureCloud platform for collaborative research.