MORTALITY PREDICTION OF MULTI-CENTER ICU

Abstract

In healthcare settings with numerous independently managed institutions, federated 
learning (FL) offers a viable answer to the issues of applying machine learning (ML) 
to privately dispersed data. Unfortunately, FL's performance and institutions' 
willingness to participate in training might be negatively impacted by the intrinsic 
non-IID (non-identically distributed) and imbalanced character of data distribution. 
Using the initial non-IID and imbalanced distribution, this research examines these 
difficulties using real-world multi-center intensive care unit electronic health record 
data. The study starts by looking at the reasons baseline FL doesn't do well in certain 
situations, and then it introduces a personalised FL (PFL) method named POLA to fix 
these problems. For each participant, POLA generates a high-performance 
personalised model using a one-shot, two-step FL approach. Results from trials 
comparing POLA to two other PFL approaches show that it improves prediction 
accuracy, decreases communication rounds, and might be used in comparable cross- 
silo FL settings. Its adaptability and scalability indicate that it might find use in a 
variety of fields outside of healthcare.