Table 2.
Classification performance statistics
| Model type* | Dataset | Sensor type/placement | Accuracy [%] |
Specificity [%] |
Sensitivity [%] |
Cohen's kappa*100 |
|---|---|---|---|---|---|---|
| Four past, two future | Healthy volunteers test set† |
Radar nightstand | 94.8 (1.7) | 89.8 (7.1) | 96.6 (1.6) | 86.9 (5.4) |
| Radar ceiling | 93.8 (2.3) | 86.3 (7.4) | 96.6 (2.1) | 84.3 (6.6) | ||
| Actigraphy | 93.1 (1.2) | 85.4 (7.4) | 96.0 (2.8) | 82.6 (3.2) | ||
| Five past, zero future (real time) | Healthy volunteers test set |
Radar nightstand | 94.5 (1.6) | 89.5 (6.4) | 96.3 (1.5) | 86.3 (4.9) |
| Radar ceiling | 93.3 (2.4) | 85.4 (7.7) | 96.3 (1.9) | 83.1 (6.8) | ||
| Actigraphy | 92.7 (1.1) | 83.9 (7.7) | 96.1 (2.6) | 81.5 (3.2) | ||
| Four past, two future | Patients with sleep disorders ‡ | Radar nightstand | 80.9 (15.7) | 53.7 (18.4) | 89.5 (16.9) | 44.8 (25.6) |
| Actigraph | 83.8 (9.0) | 74.3 (20.0) | 89.4 (7.3) | 53.3 (15.8) | ||
| Five past, zero future (real time) | Patients with sleep disorders | Radar nightstand | 80.9 (15.3) | 53.4 (18.7) | 89.7 (16.5) | 44.3 (24.8) |
| Actigraphy | 84.1 (9.0) | 74.0 (20.2) | 89.9 (6.9) | 53.8 (16.2) |
Epoch-by-epoch classification performance statistics for two models, both with the heuristic Cole–Kripke rescoring rules applied, compared to PSG§-determined sleep/wake. Mean (SD) over the participants in the datasets.
*The models are defined by the number of preceding (past) and succeeding (future) epochs used to score a present epoch.
† n = 12, mean age ± SD: 23.0 ± 3.1 years, 5 male, 4 nights of PSG + actigraphy + two radars per participant. The participants were randomly assigned into a training set for model development (n = 24/22 for nightstand/ceiling), and a testing set for validation (n = 23/21 for nightstand/ceiling). Values were calculated for all epochs from each participant first, then averaged together.
‡Ambulatory sleep disorder patients. n = 28, mean age ± SD: 46.25 ± 13.98 years, 19 male.
§PSG = polysomnography.