Personalized On-Device E-Health Analytics With Decentralized Block Coordinate Descent.

Area of Science:

• E-health
• Machine Learning
• Decentralized Learning

Background:

• E-health analytics are increasingly popular for personal healthcare, with machine learning enhancing medical diagnosis.
• Centralized cloud-based E-health models raise privacy concerns and introduce time delays due to data aggregation.
• Existing decentralized methods like D-SGD face gradient vanishing and slow initial training, impacting efficiency and fairness.

Purpose of the Study:

• To propose a novel Decentralized Block Coordinate Descent (D-BCD) learning framework for optimizing decentralized deep neural network models in E-health.
• To address the limitations of gradient-based methods, including gradient vanishing and slow early-stage convergence.
• To enhance fairness and accuracy in E-health analytics, particularly for minority groups with sparse data.

Main Methods:

• Implemented a gradient-free Block Coordinate Descent (BCD) optimization method for decentralized learning.
• Introduced similarity-based model aggregation to leverage knowledge from similar on-device models.
• Developed a D-BCD framework for optimizing deep neural networks on decentralized devices for E-health analytics.

Main Results:

• The proposed D-BCD framework demonstrated effectiveness and practicality in benchmarking experiments on three real-world datasets.
• D-BCD mitigates the gradient vanishing issue and shows faster convergence in the early training stages compared to gradient-based methods.
• Similarity-based aggregation improved personalization and accuracy, addressing data scarcity issues.

Conclusions:

• The D-BCD framework offers a superior approach for decentralized E-health analytics, enhancing model optimization and training efficiency.
• The method provides a practical solution for privacy-preserving E-health analytics, improving fairness and accuracy.
• D-BCD shows strong applicability in real-life E-health scenarios, paving the way for more robust decentralized healthcare solutions.