Development of time to event prediction models using federated learning.

Area of Science:

• Computational biology
• Biostatistics
• Machine learning in healthcare

Background:

• Training predictive models often requires large, diverse datasets, necessitating multi-site data aggregation.
• Centralized data pooling raises privacy concerns and logistical challenges.
• Federated learning (FL) offers a decentralized alternative for model training.

Purpose of the Study:

• To develop and evaluate FL algorithms for training time-to-event prediction models using distributed datasets.
• To enable individual-level survival curve prediction without exposing sensitive patient data.

Main Methods:

• Proposed two FL-based methodologies for time-to-event prediction.
• Utilized kernel smoothing for baseline hazard in Cox models.
• Applied general parametric likelihood theory for right-censored data.
• Validated methods via simulations and a real-world Hodgkin lymphoma dataset.

Main Results:

• FL models demonstrated performance comparable to non-distributed models across four simulations.
• Minor deviations observed in predicted survival probabilities compared to true models.
• Real-world data analysis showed similar performance between FL and centralized approaches for Hodgkin lymphoma.

Conclusions:

• The proposed FL methods effectively train time-to-event models on distributed data.
• Individual-level data and event times are not shared, preserving patient privacy.
• Achieved predictive performance equivalent to centralized methods.