Table 3 Real-time sleep staging performance across different input segment durations

From: Advancing sleep health equity through deep learning on large-scale nocturnal respiratory signals

Dataset

Segment

Accuracy

F1-score

Kappa

Sensitivity (%)

Inference Time*

MFLOPs*

Training Time*

  

(%)

  

Wake

Light

Deep

REM

/ Sleep Period

 

/ Night

ClinHuaiAn

1 min

78.44

0.7423

0.6585

87.75

82.10

51.85

71.22

60.70ms

286.49

25.34s

 

2 min

78.97

0.7475

0.6666

87.62

82.67

50.54

73.97

61.51ms

545.46

30.54s

 

3 min

79.72

0.7584

0.6790

88.96

82.95

53.54

73.92

62.01ms

806.26

37.14s

 

4 min

79.75

0.7586

0.6792

88.79

83.06

53.51

73.99

62.36ms

1065.23

45.85s

 

5 min

79.73

0.7584

0.6789

88.72

83.06

53.51

73.99

62.46ms

1329.70

54.89s

ClinRadar ClinRadar

1 min

76.09

0.7164

0.6227

71.96

85.53

46.36

72.19

60.70ms

286.49

25.34s

 

2 min

76.30

0.7198

0.6263

72.50

85.45

47.39

72.21

61.51ms

545.46

30.54s

 

3 min

76.27

0.7196

0.6256

72.35

85.44

47.42

72.25

62.01ms

806.26

37.14s

 

4 min

76.26

0.7194

0.6251

72.13

85.49

47.44

72.28

62.36ms

1065.23

45.85s

 

5 min

76.19

0.7189

0.6240

72.09

85.34

47.49

72.29

62.46ms

1329.70

54.89s

  1. Results are reported for respiratory belt-derived respiratory signals (ClinHuaiAn) and radar-derived respiratory signals (ClinRadar).
  2. *Inference Time: The model inference time for a single sleep epoch (30 seconds).
  3. *MFLOPs Million Floating Point Operations, a measure of the computational complexity of a neural network.
  4. *Training Time: Training duration of a single epoch when the model takes a whole night as input.
  5. The model was trained on: GPU – A800-80GB, CPU – 14 vCPU Intel Xeon Gold 6348 @ 2.60 GHz.
  6. Inference time calculations were performed on: GPU – RTX 4090, CPU – 16 vCPU Intel Xeon Gold 6430.
Back to article page