Communication Efficient Tensor Factorization for Decentralized Healthcare Networks.

Area of Science:

• Computational health informatics
• Machine learning for healthcare
• Data privacy in medical research

Background:

• Tensor factorization is effective for unsupervised learning in health data analysis, particularly for computational phenotyping using Electronic Health Records (EHRs).
• Federated tensor factorization enables multi-hospital phenotype learning while preserving patient privacy but suffers from central server vulnerabilities and communication bottlenecks.

Purpose of the Study:

• To propose CiderTF, a communication-efficient, decentralized generalized tensor factorization method.
• To address the single-point-failure issue and reduce uplink communication costs in federated tensor factorization for EHR analysis.

Main Methods:

• Developed CiderTF, a decentralized generalized tensor factorization framework.
• Implemented a four-level communication reduction strategy tailored for generalized tensor factorization.
• Evaluated the method on two real-world EHR datasets.

Main Results:

• CiderTF achieves comparable convergence to existing methods.
• Demonstrated significant communication cost reduction, up to 99.99%, in experiments.
• Showcased the flexibility of generalized tensor factorization with various tensor distributions and loss functions.

Conclusions:

• CiderTF provides an efficient and robust decentralized solution for federated tensor factorization in healthcare.
• The proposed communication reduction strategy effectively mitigates uplink bandwidth limitations.
• This approach enhances the scalability and security of collaborative EHR analysis for computational phenotyping.