Table 2.

Comparison of the experimental results for the five different MNIST cases described in the Methods.^a

Experiments	AUROCb (95% CI)	F1-score (95% CI)	Precision (95% CI)	Recall (95% CI)
CMLc	0.999 (0.999-0.999)	0.981 (0.978-0.983)	0.981 (0.972-0.989)	0.981 (0.971-0.989)
Basic FLd	0.997 (0.996-0.998)	0.946 (0.941-0.950)	0.945 (0.929-0.959)	0.945 (0.930-0.959)
Imbalanced FL	0.995 (0.994-0.995)	0.921 (0.917-0.927)	0.920 (0.904-0.937)	0.920 (0.903-0.937)
Skewed FL	0.992 (0.991-0.993)	0.905 (0.899-0.911)	0.905 (0.885-0.922)	0.904 (0.885-0.920)
Imbalanced and skewed FL	0.990 (0.989-0.991)	0.891 (0.884-0.896)	0.890 (0.869-0.909)	0.889 (0.868-0.908)

^aAll experiments used the same model and hyperparameters. All results are presented with a 95% CI by resampling the validation task 100 times.

^bAUROC: area under the receiver operating characteristic curve.

^cCML: centralized traditional machine-learning method.

^dFL: federated learning.