Skip to main content

Table 4 Contingency table of classifier performance on test cohort

From: Sparse feature selection for classification and prediction of metastasis in endometrial cancer

Actual/Classification

Positive

Negative

Total

Positive

Negative

Total

Node-Positive

8

1

9

7

2

9

Node-Negative

4

15

19

4

15

19

Total

12

16

28

11

17

28

Accuracy

0.8214

0.7857

Sensitivity

0.8889

0.7778

Specificity

0.7895

0.7895

False Discovery Rate

0.0625

0.1174

P-Value (Fisher)

0.0012

0.0104

P-Value (Barnard)

0.0004

0.0037

  1. (The performance of the classifier on the 86 training cohort is not shown as it was 100%.) The left part of the table corresponds to sample #198 treated as node-positive, while the right part of the table corresponds to sample #198 treated as node-negative. When sample #198 is treate as node-positive, the classifier has accuracy of 82.14%, with 23 out of 28 tumors being correctly classified; sensitivity of 88.89% with 8 out of 9 lymph-positive tumors being correctly classified; and specificity of 78.95%, with 15 out of 19 lymph-negative tumors being correctly classified. The P-value of obtaining these values purely by chance was computed using the Fisher exact test at 0.0012 and as 0.0004 using the more powerful Barnnard exact test. The corresponding figures with sample #198 treated as node-negative are shown for comparison. It can be seen that even this case, all P-values are far lower than the widely accepted threshold of 0.05