PERSONALIZED FEDERATED LEARNING FOR IN-HOSPITAL

Abstract

When dealing with privately dispersed data, such as in healthcare settings where 
numerous independently owned institutions are present, federated learning (FL) offers 
a potential answer. Nevertheless, institutions may be discouraged from engaging in 
training and FL's performance may be negatively affected due to the data's intrinsic 
imbalance and non-IID (non-identically distributed) characteristics. Keeping the 
original non-IID and imbalanced distribution, this research examines these difficulties 
using real-world multi-center intensive care unit electronic health record data. In order 
to address these concerns, the research first investigates why baseline FL does not 
perform well in these settings. Then, it introduces a personalised FL (PFL) method 
named POLA. POLA is a two-step FL approach that is customised for each 
participant and is aimed to generate high-performance personalised models. Testing 
POLA in comparison to two other PFL approaches shows that it can improve 
prediction accuracy and decrease communication rounds. It also shows promise for 
use in comparable cross-silo FL settings. Due to its scalability and adaptability, it 
might find use in many other fields outside healthcare.