PPFL: A personalized progressive federated learning method for leveraging different healthcare institution-specific...

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training on distributed healthcare data while preserving privacy.
• Conventional FL struggles with heterogeneous data, as institutions collect distinct features.
• Leveraging institutional data diversity is crucial for improving predictive model performance.

Purpose of the Study:

• To introduce a Personalized Progressive Federated Learning (PPFL) approach.
• To address the challenge of leveraging client-specific features in FL for healthcare.
• To evaluate PPFL's effectiveness in improving predictive accuracy using real-world datasets.

Main Methods:

• Developed a PPFL algorithm designed for heterogeneously distributed clients.
• Incorporated client-specific features into the federated learning framework.
• Evaluated PPFL against conventional models (local training, FedAvg) for in-hospital mortality prediction.
• Utilized real-world healthcare datasets, including mortality and cancer data.

Main Results:

• PPFL achieved superior performance in in-hospital mortality prediction, with an accuracy of 0.941 and AUROC of 0.948.
• PPFL's performance surpassed that of local models and the FedAvg algorithm.
• The model demonstrated comparable effectiveness on cancer datasets.
• Identified key client-specific features contributing to mortality prediction.

Conclusions:

• PPFL effectively leverages unique, institution-specific features within a federated learning framework.
• The personalized approach enhances model performance in heterogeneous healthcare data settings.
• PPFL offers a promising solution for privacy-preserving, collaborative AI in diverse clinical applications.
• This method expands the feature space for personalized vertical feature learning.