Conference Agenda

Mortality prediction in intensive care units represents a fundamental challenge for clinical decision-making. However, developing robust predictive models requires large volumes of data that are typically fragmented across multiple healthcare institutions, each with strict privacy policies that prevent the centralization of sensitive information. This study implements and evaluates a federated learning system based on the Federated Averaging (FedAvg) algorithm to train mortality prediction models without the need to share clinical data among participating institutions. Thru computational simulation, a multi-institutional scenario was reproduced with five virtual hospitals, each with heterogeneous demographic characteristics and data distributions. The results demonstrate that the federated approach achieves an area under the ROC curve (AUC-ROC) of 0.892 ± 0.008, representing only a 3.6% difference compared to the centralized reference model (AUC-ROC = 0.925 ± 0.005), while reducing the volume of transferred data by 98% and fully preserving institutional privacy. Statistical analysis using a paired Student’s t-test confirms that this difference, although statistically significant (p < 0.001), is clinically acceptable. It is concluded that federated learning constitutes a viable alternative for inter-institutional collaboration in clinical research, enabling the development of high-performance predictive models without compromising patient confidentiality.