Table 1 Comparison of baseline models to the weakly supervised CNN using the F1-score and FPR metrics.

From: Weak supervision as an efficient approach for automated seizure detection in electroencephalography

Performance comparison of baseline models to the weakly supervised model

 

Pediatric (F1-score, FPR)

Adult (F1-score, FPR)

 

12-s

60-s

12-s

60-s

Logistic regression

0.25, 0.36

0.37, 0.50

0.24, 0.56

0.38, 0.55

Random forest

0.38, 0.19

0.56, 0.52

0.37, 0.61

0.58. 0.52

Persyst-13

0.07, 0.015

0.61, 0.10

0, 0

0.65, 0.054

Dense-CNN (small gold-standard set)

0.29, 0.39

0.60, 0.20

0.28, 0.47

0.41, 0.16

Dense-CNN (weak annotations)

0.67, 0.13

0.77, 0.10

0.49, 0.14

0.76, 0.079

  1. The F1-score of our weakly supervised CNN outperforms the baselines by substantial margins. The small gold-standard training dataset consisted of 241 clips for pediatrics and 246 clips for adults. The weakly annotated training dataset consisted of 25,386 clips for pediatrics and 32,596 clips for adults.
Back to article page