Target informed client recruitment for efficient federated learning in healthcare.

Area of Science:

• Machine Learning in Healthcare
• Artificial Intelligence for Medical Decision Support

Background:

• Healthcare data aggregation for machine learning poses significant privacy risks and administrative burdens.
• Federated learning (FL) enables distributed model training without data aggregation, but faces challenges in client selection and efficiency.
• Existing FL methods require improvements for practical, privacy-preserving healthcare applications.

Purpose of the Study:

• To develop an enhanced client recruitment strategy for federated learning in healthcare.
• To improve the efficiency and privacy-preserving capabilities of federated learning models.
• To address challenges in client selection for federated learning by incorporating local hardware and data characteristics.

Main Methods:

• Extended a prior client recruitment approach by integrating local hardware knowledge.
• Recruited federated learning clients based on representativeness, considering local target distribution divergence, sample size, and hardware efficiency.
• Evaluated the approach on medical regression and classification tasks.

Main Results:

• The proposed recruitment approach achieved performance on par with or superior to central and federated methods.
• Reduced training time by a factor of 3-4, requiring a fraction of the data.
• Demonstrated that excluded clients can benefit from the federated model via local fine-tuning.

Conclusions:

• A novel client recruitment strategy was defined using local target distribution, sample size, and hardware efficiency.
• This approach significantly reduces federated learning training time without compromising predictive performance.
• The method enhances privacy-preserving characteristics compared to standard federated learning and central approaches.