ENHANCING IN-HOSPITAL MORTALITY PREDICTION VIA PERSONALIZED FEDERATED LEARNING IN MULTI INSTITUTIONAL ICUS

Abstract

A fresh and promising plan to advance machine learning (ML) at many independently dispersed healthcare facilities is provided by federated learning (FL), a paradigm for resolving the difficulties of applying ML to private distributed data. Its effectiveness may be hampered, however, by the non-IID and uneven distribution of the data, which may even make the institutions less inclined to take part in its training. In order to retain the original non-IID and unbalanced data distribution, this research investigated the issue of an in-hospital mortality prediction job under a real multi-center ICU electronic health record database. Prior to proposing a personalised FL (PFL) solution called POLA to address the issue, it first examined the cause of the baseline FL's performance decrease in this data setting. POLA is a customised one-shot, two-step FL technique that may provide very effective customised models for every individual participant. In studies, the suggested POLA approach was contrasted with two other PFL techniques, and the findings show that it not only considerably lowers the number of communication cycles but also enhances FL's prediction performance. Additionally, because of its generality and flexibility, it may be expanded to other comparable cross-silo FL application contexts.