PROJECT TITLE :

Personalized On-Device E-health Analytics with Decentralized Block Coordinate Descent

ABSTRACT:

The proliferation of interest in e-health is being driven in large part by the increased focus on individual Healthcare as well as the ongoing epidemic. In today's world, improvements to medical diagnosis achieved through the use of Machine Learning models have proven to be highly effective in many facets of e-health analytics. Nevertheless, in the traditional cloud-based and centralized e-health paradigms, all of the data will be centrally stored on the server in order to make model training more accessible. This will inevitably raise concerns regarding the confidentiality of the data as well as a significant delay in its processing. Distributed solutions such as the Decentralized Stochastic Gradient Descent (D-SGD) have been proposed in order to provide diagnostic results that are secure and up to date based on personal devices. Methods such as D-SGD, on the other hand, are susceptible to an issue known as gradient vanishing and typically move slowly during the early stages of training. This hinders both the effectiveness and the efficiency of the training process. Additionally, existing methodologies are prone to learning models that are biased towards users who have dense data, which compromises the fairness of providing E-health analytics to minority groups. In this paper, we propose a Decentralized Block Coordinate Descent (D-BCD) learning framework that, for the purpose of E-health analytics, can better optimize deep neural network-based models that are distributed on decentralized devices. When compared to traditional gradient-based optimization, the Block Coordinate Descent (BCD) method of optimization, which does not use gradients, solves the problem of vanishing gradients and converges more quickly during the early stages of the process. We propose a similarity-based model aggregation as a solution to the potential data scarcity issues for users' local data. This solution enables each on-device model to leverage knowledge from similar neighbor models in order to achieve both high accuracy and personalization for the learned models. Experiments on three real-world datasets were used to demonstrate the efficacy and applicability of our proposed D-BCD. In addition, a simulation study demonstrated the strong applicability of D-BCD in real-life E-health scenarios, demonstrating the strong applicability of D-BCD in real-life E-health scenarios.