ISEP - DM – Engenharia de Inteligência Artificial
Permanent URI for this collection
Browse
Browsing ISEP - DM – Engenharia de Inteligência Artificial by Subject "Acute pancreatitis"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- A federated learning approach for data privacy in healthcare applicationsPublication . Vieira, Pedro Manuel Ribeiro; Maia, Eva Catarina GomesIn healthcare, actions tend to generate a vast amount of sensitive patient data, which is useful for scientific advancements and new applications, but also presents privacy and security challenges. Artificial intelligence can significantly benefit from this data, but traditional Machine Learning (ML) techniques in collaborative environments expose it excessively. Federated Learning (FL) emerges as a solution, enabling model training without directly sharing patient information, thus reducing the risk of data exposure. This thesis has three main goals. It aims to understand the most common FL tools in the state of the art, analyzing their advantages and disadvantages to select the most appropriate one. This is due to the need to identify tools that can be effectively applied to ensure both learning efficiency and data security, as well as applicability to the theme at hand. It also addresses the need to understand the most common FL scenarios in the healthcare domain presented in the literature, as it helps to identify best practices and specific challenges in this sector. The last goal is to suggest an effective FL approach that ensures data privacy. This goal is driven by the growing need for solutions that can ensure compliance with privacy regulations while enabling model training in a collaborative environment. Regarding the first objective, it was concluded that Flower is the most suitable tool for the purpose of this thesis. Although other tools, such as PySyft, stood out, Flower was the one that best met the needs of the work. Next, four major technical problems commonly encountered when working with FL were identified: scalability, security, the particularities of each type of FL partition, and data distribution. To deal with some of these technical challenges, techniques such as undersampling were employed. Furthermore, through this investigation, it became clear that a network of hospitals is one of the most common scenarios when it comes to FL in healthcare. A solution was finally proposed, and an FL scenario was designed with three hospitals collaborating to train a global model. First, the robustness and effectiveness of FL compared to traditional ML were analyzed, noting no significant loss in most models. Next, the performance of aggregation algorithms (FedAvg, FedAdam, FedAdagrad) was compared, with FedAvg standing out. Finally, the training time between the various models was compared. This performance analysis derived from two case studies: predicting mortality in patients with Acute Pancreatitis and predicting mortality in patients in Intensive Care Units (ICU) with various diseases. Thus, all the three proposed objectives were completely fulfilled.