A federated learning framework for ethical dynamic treatment allocation across heterogeneous hospitals

Area of Science:

• Machine Learning
• Healthcare Informatics
• Clinical Decision Support

Background:

• Hospitals often serve diverse patient populations, making universal treatment protocols suboptimal.
• Existing treatment allocation methods may not adequately address inter-hospital variations or patient privacy concerns.

Purpose of the Study:

• To develop an adaptive federated learning framework for optimizing individual hospital treatments.
• To enhance treatment allocation efficiency while ensuring privacy, fairness, and safety.

Main Methods:

• A federated treatment recommendation strategy formulated as a Multi-Armed Bandit (MAB) problem.
• A lead hospital coordinates adaptive learning and transfers Upper Confidence Bounds (UCB) and Personalized Upper Bounds.
• Method validated in a simulated clinical trial using real COVID-19 data.

Main Results:

• Collaborative learning reduces data samples needed per institution while preserving patient privacy.
• Fairness is ensured by mitigating biases from differing patient populations.
• Improved safety by minimizing sub-optimal treatment administration.
• Outperforms state-of-the-art, requiring 36%-75% fewer data and administering optimal treatments to 0.95%-48.6% more patients.

Conclusions:

• An adaptive federated learning strategy effectively recommends optimal treatments for individual hospitals.
• The framework addresses privacy, fairness, and safety in heterogeneous healthcare settings.