Applying federated learning to a Covid-19 anomaly detection pipeline
Year: 2023
Type: dissertation
Abstract: The COVID-19 pandemic of 2020 spread around the world fast, overwhelming healthcare systems and causing millions of deaths all over the globe. Researchers and scientists rushed to find tools to help detect and contain the spread of the disease, including automatic ones powered by machine learning models. Amongst the efforts, several studies focused on exploring differences between biometric signals in people with the disease collected before and after the infection, in search of patterns that can help detect it as soon as possible. In particular, heart rate related signals collected via devices such as smartwatches. These studies have resulted in some detection tools, but they always require users to have data from before the infection occurred in order to be used and often contain personalization based on this healthy data. But what if a new user has not yet collected healthy data? Can a model trained with the data of other individuals successfully detect the illness on a novel one? This work explores that situation by taking an individual based anomaly detection and transporting into a Federated Learning environment in order to see its behavior on the data from individuals that trained the model and novel individuals, including data collected via hospital devices from individuals admitted to the intensive care unit which is unlikely to contain healthy data. Although the resulting models, on average, did not detect more true anomalies than the personalized ones, their performance is similar when applied to the training individuals and novel ones.
Source: Repositório do ISCTE-IUL
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6998608338