Communication Efficient Federated Generalized Tensor Factorization for Collaborative Health Data Analytics.

Area of Science:

• Computational biology
• Data science in healthcare
• Machine learning for electronic health records

Background:

• Healthcare systems generate vast amounts of high-dimensional relational data.
• Tensor factorization is effective for extracting medical concepts (phenotypes) from this data.
• Federated learning offers privacy-preserving collaborative learning for sensitive health data.

Purpose of the Study:

• To address limitations of existing federated tensor factorization methods, such as high communication costs and reduced accuracy.
• To propose a communication-efficient federated generalized tensor factorization (GFT) for collaborative phenotyping.
• To enhance the flexibility and applicability of federated tensor factorization in healthcare.

Main Methods:

• Developed a novel federated generalized tensor factorization (GFT) approach.
• Implemented a three-level communication reduction strategy to minimize uplink costs.
• Theoretically analyzed the algorithm's convergence speed under aggressive communication compression.

Main Results:

• Achieved up to 99.90% reduction in uplink communication cost.
• Demonstrated that the proposed method does not compromise convergence speed.
• Validated efficiency improvements in computation and communication on two real-world electronic health record datasets.

Conclusions:

• The proposed communication-efficient federated generalized tensor factorization is a viable and effective approach for collaborative phenotyping.
• This method offers significant improvements in communication efficiency without sacrificing accuracy or convergence.
• The GFT framework provides flexibility in choosing appropriate loss functions for diverse healthcare data.