Developing federated time-to-event scores using heterogeneous real-world survival data

Area of Science:

• Healthcare Analytics
• Biostatistics
• Machine Learning in Medicine

Background:

• Survival analysis is crucial for clinical decision-making, predicting time-to-event outcomes.
• Current survival scoring systems require centralized data, limiting multi-institutional research due to privacy concerns.
• Federated learning offers a solution for collaborative model training without data sharing.

Purpose of the Study:

• To develop a novel, privacy-preserving federated framework for constructing survival scoring systems.
• To enable efficient and secure collaboration among multiple data owners for survival outcome prediction.
• To address the limitations of single-source data assumptions in existing survival score construction.

Main Methods:

• A federated learning framework was designed for multi-site survival outcome analysis.
• The approach was applied to heterogeneous survival data from emergency departments in Singapore and the US.
• Local survival scores were independently developed at each participating site for comparison.

Main Results:

• The federated scoring system consistently outperformed local models across all testing sites.
• Integrated area under the receiver operating characteristic curve (iAUC) showed a maximum improvement of 11.6% with the federated approach.
• Federated scores demonstrated superior time-dependent AUC(t) values with narrower confidence intervals compared to local scores.

Conclusions:

• The proposed federated survival score generation framework is effective and applicable to real-world heterogeneous data.
• This privacy-preserving method enhances prediction accuracy and efficiency for survival models.
• The framework holds promise for future collaborative healthcare research, improving risk prediction in clinical settings.