Skip to main content

Table 2 Epoch-per-epoch agreement between predicted sleep stages based on PPGĀ and ground-truth for different classification tasks

From: Direct application of an ECG-based sleep staging algorithm on reflective photoplethysmography data decreases performance

Task

kappa (ā€“)

Accuracy (%)

Sensitivity (%)

Specificity (%)

PPV (%)

Wake/N1ā€“2/N3/REM

0.56ā€‰Ā±ā€‰0.15

73.0ā€‰Ā±ā€‰9.4

n/a

n/a

n/a

Wake/NREM/REM

0.62ā€‰Ā±ā€‰0.16

81.4ā€‰Ā±ā€‰8.5

n/a

n/a

n/a

Wake/sleepa

0.57ā€‰Ā±ā€‰0.18

87.7ā€‰Ā±ā€‰8.1

67.8ā€‰Ā±ā€‰19.9

91.9ā€‰Ā±ā€‰8.4

68.4ā€‰Ā±ā€‰19.6

N1ā€“2a

0.49ā€‰Ā±ā€‰0.16

75.1ā€‰Ā±ā€‰8.3

77.1ā€‰Ā±ā€‰10.9

72.6ā€‰Ā±ā€‰13.6

75.9ā€‰Ā±ā€‰12.3

N3a

0.51ā€‰Ā±ā€‰0.24

91.2ā€‰Ā±ā€‰5.2

50.7ā€‰Ā±ā€‰26.4

97.6ā€‰Ā±ā€‰3.2

75.5ā€‰Ā±ā€‰26.6

REMa

0.64ā€‰Ā±ā€‰0.22

91.9ā€‰Ā±ā€‰4.5

79.8ā€‰Ā±ā€‰21.8

93.6ā€‰Ā±ā€‰4.1

64.6ā€‰Ā±ā€‰21.0

  1. PPV positive predictive value
  2. aBinary classification tasks were evaluated in a one vs. rest strategy, where one single class (wake, N1ā€“N2, or N3, or REM) was considered the ā€˜positiveā€™ class, and the remaining classes were merged in a single ā€˜negativeā€™ class. All results are presented as meanā€‰Ā±ā€‰SD