FedComDist: Towards Effective Personalized Federated Learning for Patient Outcome Prediction Using Multi-Center...

Area of Science:

• Medical Informatics
• Machine Learning
• Healthcare Data Analytics

Background:

• Accurate patient outcome prediction is crucial for healthcare improvement.
• Utilizing diverse medical data for prediction faces privacy and security challenges.
• Federated learning offers a solution for collaborative model training while preserving data privacy.

Purpose of the Study:

• To propose FedComDist, a personalized federated learning approach to enhance patient outcome prediction using heterogeneous clinical features across hospitals.
• To address the suboptimal performance of conventional federated learning due to data heterogeneity.

Main Methods:

• Categorizing clinical features into common and distinct based on their presence across hospitals.
• Decoupling model parameters into common global and distinct local parameters.
• Training common parameters on common features globally and distinct parameters on distinct features locally.

Main Results:

• FedComDist demonstrated effectiveness in predicting patient mortality and Length of Stay (LoS) on the eICU dataset.
• The proposed approach outperformed various federated learning methods.
• Enhanced data privacy was achieved through parameter decoupling.

Conclusions:

• FedComDist effectively utilizes heterogeneous clinical features for improved patient outcome prediction.
• The personalized federated learning approach balances global knowledge sharing with local data specificity.
• This method offers a promising solution for privacy-preserving, accurate clinical outcome prediction in multi-center settings.