Decentralized Machine Learning Approach on ICU Admission Prediction for Enhanced Patient Care Using COVID-19 Data

Abstract

The Intensive Care Unit (ICU) represents a constrained healthcare resource, involving invasive procedures and high costs, with significant psychological effects on patients and their families. The traditional approach to ICU admissions relies on observable behavioral indicators like breathing patterns and consciousness levels, which may lead to delayed critical care due to deteriorating conditions. Therefore, in the ever-evolving healthcare landscape, predicting whether patients will require admission to the ICU plays a pivotal role in optimizing resource allocation, improving patient outcomes, and reducing healthcare costs. Essentially, in the context of the post-COVID-19 pandemic, aside from many other diseases, this prediction not only forecasts the likelihood of ICU admission but also identifies patients at an earlier stage, allowing for timely interventions that can potentially mitigate the need for ICU care, thereby improving overall patient outcomes and healthcare resource utilization. However, this task usually requires a lot of diverse data from different healthcare institutions for a good predictive model, leading to concerns regarding sensitive data privacy. This paper aims to build a decentralized model using deep learning techniques while maintaining data privacy among different institutions to address these challenges.