Hospital Participation in Federated Learning: Evaluating Sustainability and Clinical Utility.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Oncology

Background:

• Prostate cancer (PCa) diagnosis frequently involves invasive biopsies, posing risks and potential for unnecessary procedures.
• Traditional centralized machine learning models require sharing sensitive patient data, raising privacy concerns.
• Federated learning (FL) enables collaborative model training across institutions without direct data sharing, preserving patient privacy.

Purpose of the Study:

• To evaluate the feasibility and performance of federated learning (FL) for prostate cancer (PCa) risk prediction.
• To benchmark FL against local training and free-riding strategies using real-world heterogeneous healthcare data.
• To analyze the influence of data diversity and consortium size on the predictive accuracy of FL models.

Main Methods:

• Utilized real-world, heterogeneous datasets from 19 hospitals for training and evaluation.
• Compared predictive performance across three strategies: local model training, federated learning, and free-riding on federated models.
• Assessed the impact of varying data diversity and the number of participating institutions on model generalizability and accuracy.

Main Results:

• Federated learning (FL) demonstrated improved model generalizability compared to local training.
• Local models achieved comparable performance to FL models, particularly for large hospitals with extensive data.
• A smaller consortium of high-quality data institutions showed potential for developing robust predictive models.

Conclusions:

• Federated learning is a feasible approach for privacy-preserving PCa risk prediction.
• The direct benefit of FL participation may be limited for large institutions; strategic collaboration is key.
• Recommendations are provided for the sustainable implementation of FL in real-world clinical networks.