Table 3.
Sleep/wake differentiation performance by k-nearest neighbors across different feature inputs in the Apple Watch (PPG, MEMS) dataset
Accuracy | Wake correct (specificity) | Sleep correct (sensitivity) | κ | AUC | |
---|---|---|---|---|---|
Motion | 0.789 | 0.672 | 0.8 | 0.255 | 0.803 |
0.866 | 0.483 | 0.9 | 0.3 | ||
0.887 | 0.405 | 0.93 | 0.307 | ||
0.9 | 0.345 | 0.95 | 0.307 | ||
HR | 0.768 | 0.406 | 0.8 | 0.117 | 0.682 |
0.845 | 0.237 | 0.9 | 0.117 | ||
0.868 | 0.172 | 0.93 | 0.103 | ||
0.882 | 0.12 | 0.95 | 0.082 | ||
Motion, HR | 0.79 | 0.678 | 0.8 | 0.255 | 0.81 |
0.867 | 0.496 | 0.9 | 0.308 | ||
0.889 | 0.431 | 0.93 | 0.327 | ||
0.903 | 0.38 | 0.95 | 0.338 | ||
Motion, HR, and Clock Proxy | 0.8 | 0.797 | 0.8 | 0.309 | 0.868 |
0.877 | 0.627 | 0.9 | 0.391 | ||
0.897 | 0.535 | 0.93 | 0.404 | ||
0.909 | 0.458 | 0.95 | 0.402 |
Fraction of wake correct, fraction of sleep correct, κ, and AUC for sleep-wake predictions of k-nearest neighbor classifier with use of motion, HR, clock proxy, or combination of features. HR, heart rate.