Table 4.
Performance comparison of different learning methods (with the same NN architecture and hyperparameters)
| Datasets | MIMIC-III | MIMIC-IV | eICU-CRD | Average accuracy |
|---|---|---|---|---|
| Learning methods | ||||
| Centralized learning | 77.06% | 72.42% | 85.36% | 78.28% |
| Local learning using MIMIC-III | 77.15% | 73.31% | 85.29% | 78.58% |
| Local learning using MIMIC-IV | 67.80% | 83.02% | 48.59% | 66.47% |
| Local learning using eICU-CRD | 78.04% | 72.96% | 85.37% | 78.79% |
| FedAvg [8] | 72.86% | 77.48% | 83.21% | 77.85% |
| Weighted FedAvg [8] | 72.68% | 77.21% | 83.22% | 77.70% |
| Weighted FedProx [10] | 58.56% | 60.44% | 66.79% | 61.93% |
| Weighted Mime Lite [11] | 51.84% | 52.58% | 55.51% | 53.21% |
| Non-IID Weighted FedAvg | 79.04% | 74.63% | 85.87% | 79.85% |
| Non-IID FedAvg | 79.03% | 75.83% | 84.44% | 79.77% |