Multi-site brain disease identification based on tensor decomposition and personalized federated learning

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Brain diseases pose significant health challenges, necessitating early diagnosis through robust predictive models.
• Medical data privacy concerns and data silos limit the development of high-quality diagnostic models.
• Existing multi-site studies often fail to incorporate site-specific private features, hindering comprehensive analysis.

Purpose of the Study:

• To propose a Tensor Decomposition and Personalized Federated Learning (TDPFL) framework for multi-site brain disease recognition.
• To protect sensitive demographic information (age, gender, education) during multi-site data analysis.
• To improve the accuracy and efficiency of early brain disease diagnosis using integrated data.

Main Methods:

• Developed a dual feature aggregation module on the central server for efficient inter-site knowledge sharing.
• Implemented a personalized branch on client sides to safeguard private demographic data.
• Utilized a tensor decomposition module for extracting features from brain scan data (rs-fMRI).
• Incorporated a dynamic prototype aggregation module to capture evolving brain features over time.

Main Results:

• The TDPFL framework demonstrated superior performance compared to baseline methods.
• Achieved a 4% improvement in average classification accuracy on two public rs-fMRI datasets across six sites.
• Successfully identified site-specific brain disease-related biomarkers, aiding early diagnosis.

Conclusions:

• TDPFL offers an effective solution for privacy-preserving multi-site brain disease recognition.
• The framework enhances the ability to capture dynamic brain changes, leading to improved diagnostic accuracy.
• Identified biomarkers provide valuable insights for the early detection and understanding of brain diseases.