Privacy-preserving patient clustering for personalized federated learning

Area of Science:

• Machine Learning
• Medical Informatics
• Data Privacy

Background:

• Federated Learning (FL) trains models without central data sharing but struggles with non-identically independently distributed (non-IID) data, common in healthcare.
• Personalized FL and Clustered FL variants aim to address non-IID data by tailoring models to site-specific distributions.
• Existing Clustered FL methods face privacy risks or performance degradation due to data aggregation during patient clustering.

Purpose of the Study:

• To introduce Privacy-preserving Community-Based Federated machine Learning (PCBFL), a novel framework for privacy-preserving patient clustering in FL.
• To enable secure clustering using patient-level data without compromising privacy through Secure Multiparty Computation.
• To evaluate PCBFL's effectiveness in improving federated model performance for medical applications.

Main Methods:

• Developed PCBFL, a Clustered FL variant utilizing Secure Multiparty Computation for calculating patient-level similarity scores across institutions.
• Implemented PCBFL to train a federated mortality prediction model using data from 20 sites in the eICU dataset.
• Compared PCBFL's performance against traditional and existing Clustered FL frameworks.

Main Results:

• PCBFL successfully formed clinically relevant patient cohorts (low, medium, high-risk).
• The framework demonstrated significant performance improvements over baseline methods.
• Achieved an average Area Under the Curve (AUC) improvement of 4.3% and an Average Precision-Recall Curve (AUPRC) improvement of 7.8%.

Conclusions:

• PCBFL offers a robust solution for privacy-preserving patient clustering in federated learning settings.
• The proposed method effectively addresses non-IID data challenges in healthcare, enhancing model accuracy.
• PCBFL facilitates the creation of meaningful patient subgroups, leading to superior federated model performance in medical prediction tasks.