Machine learning models play a crucial role in predicting in-hospital mortality but face challenges related to data privacy and the inherent complexity of clinical data. This study evaluates five different federated learning strategies for their effectiveness in real-world clinical settings with non-IID data distributions. By employing the large-scale MIMIC-IV dataset, the research creates a realistic environment that mimics the variability found in actual healthcare data.
The findings reveal that the regularization-based strategy, FedProx, yielded the highest F1-Score of 0.8831, suggesting it is more effective at handling the class imbalance inherent to mortality prediction tasks. In contrast, while FedAvg was the most computationally efficient, it performed poorly in terms of predictive accuracy. These results underscore the significance of selecting appropriate federated learning algorithms to ensure reliable and effective clinical predictions in healthcare applications.
👉 Pročitaj original: arXiv AI Papers
