A Multi-Task Cardiovascular Disease Classification Method via Adaptive Update

EasyChair Preprint 5895

Abstract

Federate learning is a generic method to augment medical data by training a global model with distributed data located in regional medical environment. Nevertheless, there exists a significant distinction with respect to the distribution of medical cases between primary hospitals and tertiary hospitals, which lead to the state that the global model is incapable of satisfying individual medical needs. We propose a novel multi-task cardiovascular disease classification method based on adaptive update of sample masking and the adaptive optimization of weights. First, a mask is designed to fit local data, and it can guide the weight updated with accordance to hard samples. Then, the global model is focused on specific task with adaptive optimization of weights. Experimental evaluation on our collected data show great improvement. The F1 metric increases from 49.54% to 75.5% by adopting the proposed method.

Keyphrases: Adaptive update, Federated Learning, personalized model

@booklet{EasyChair:5895,
  author    = {Renjie Tang and Junzhou Luo and Junbo Qian},
  title     = {A Multi-Task Cardiovascular Disease Classification Method via Adaptive Update},
  howpublished = {EasyChair Preprint 5895},
  year      = {EasyChair, 2021}}