Table 1.
List of variables used to predict hospitalization mortality in the federated and local learning approaches.
Table 2.
Hospitals by region in Brazil, with a list of hospitals.
Fig 1.
Schematic visualization of the federated learning approaches.
(A) Parameter averaging framework (Scenario I): In this system, participants (hospitals) compute model coefficients and send them anonymously to the server for aggregation. The server then returns the updated global model to the participants in an iterative process. (B) Decision tree aggregation framework (Scenario II): Here, the server determines the number of decision trees each hospital should build based on its proportional data size. These locally built trees are then sent back to the server, which aggregates them into a single Federated Random Forest model, subsequently distributed to all participating hospitals.
Fig 2.
Comparison of predictive performance between federated and local learning models.
The figure illustrates the performance results for three machine learning models: (A) Logistic Regression, (B) Multi-Layer Perceptron (MLP), and (C) Random Forest. For each model, the left panel compares the mean Area Under the Curve (AUC) of the federated model (red dashed line) against the locally trained model (blue solid line), with shaded areas representing the 95% confidence intervals derived from bootstrap analysis. The right panel displays the performance gain, quantified as the Delta AUC (ΔAUC = AUC - AUC
), where the error bars indicate the 95% CI. The horizontal dotted line represents a performance gain of zero. The x-axis for all plots is on a logarithmic scale and corresponds to the number of patients at each hospital. A consistent trend is observed across all models where federated learning generally outperforms local learning, an advantage that is particularly evident for hospitals with smaller datasets.