Methods and Impact for Using Federated Learning to Collaborate on Clinical Research

Area of Science:

• Medical AI
• Neurosurgery
• Machine Learning

Background:

• Accurate machine learning (ML) requires diverse data, which is challenging in healthcare due to data sensitivity and silos.
• Federated learning (FL) offers a decentralized approach, distributing algorithms to data and updating a global model centrally to avoid data aggregation.

Purpose of the Study:

• Establish a multicenter collaboration to assess the feasibility of using FL for training ML models.
• Develop an FL model for intracranial hemorrhage (ICH) detection without inter-site data sharing.

Main Methods:

• Five US neurosurgery departments formed a federated network.
• Trained a convolutional neural network (CNN) to detect ICH on computed tomography (CT) scans using FL.
• Benchmarked the FL model against a centrally trained model and local models.

Main Results:

• Successfully initiated a federated network for ICH prediction model training.
• The FL model achieved an AUC of 0.9487 for all ICH subtypes, compared to 0.9753 for the benchmark model.
• The FL model demonstrated consistent top-tier performance across local datasets, indicating improved generalizability.

Conclusions:

• Demonstrated the feasibility of implementing a federated network for multi-institutional clinical collaboration.
• Showcased the utility of FL for machine learning research in neurosurgery.
• Opened a new paradigm for collaborative neurosurgical research.