. 2021 Nov 22;11:22636. doi: 10.1038/s41598-021-02168-4

Table 4.

Validation of patient-level accuracy in TCGA cases. Patient-level accuracy is shown for EBV + MSI vs. others task with and without data augmentation. The result when a part of TCGA cases (20% of all cases, randomly selected) was added to the training data and the remaining 80% of cases were used as test data is also shown. The highest AUC is indicated in bold. AUC, Area under the curve; CI, Confidence interval; EBV, Epstein-Barr virus; MSI, Microsatellite instability; TCGA, The cancer genome atlas; UT University of Tokyo.

Data augmentation	Use a part of TCGA cases for training	UT test case	TCGA test case
−	−	0.848 (28/33)	1.000 (49/49)	0.934 (0.864–1.000)	0.851 (57/67)	0.480 (85/177)	0.756 (0.686–0.825)
−	+	0.818 (27/33)	0.959 (47/49)	0.943 (0.885–1.000)	0.574 (31/54)	0.852 (121/142)	0.800 (0.729–0.871)
+	−	0.879 (29/33)	0.878 (43/49)	0.947 (0.901–0.992)	0.731 (49/67)	0.876 (155/177)	0.864 (0.811–0.918)
+	+	0.848 (28/33)	0.816 (40/49)	0.939 (0.886–0.991)	0.741 (40/54)	0.873 (124/142)	0.870 (0.809–0.931)

Data augmentation

Use a part of TCGA cases for training

UT test case

TCGA test case

Sensitivity

Specificity

AUC
(95%CI)

Sensitivity

Specificity

AUC
(95%CI)

−

0.848

(28/33)

1.000

(49/49)

0.934

(0.864–1.000)

0.851

(57/67)

0.480

(85/177)

0.756

(0.686–0.825)

−

0.818

(27/33)

0.959

(47/49)

0.943

(0.885–1.000)

0.574

(31/54)

0.852

(121/142)

0.800

(0.729–0.871)

−

0.879

(29/33)

0.878

(43/49)

0.947

(0.901–0.992)

0.731

(49/67)

0.876

(155/177)

0.864

(0.811–0.918)

0.848

(28/33)

0.816

(40/49)

0.939

(0.886–0.991)

0.741

(40/54)

0.873

(124/142)

0.870

(0.809–0.931)