Non-Contact REM Sleep Estimation Correction by Time-Series Confidence of Predictions: From Binary to Continuous...

This paper focuses on the REM sleep estimation with bio-vibration data acquired from mattress sensor, and proposes its "correction" method based on Time-Series Confidence (TSC) of the REM sleep prediction calculated by Random Forest (RF) as one of the Machine Learnings (MLs). Unlike the conventional MLs that classify whether the REM sleep or not as its binary prediction, the proposed method determines whether the estimated REM sleep should be corrected or not from its continuous prediction. Concretely, the proposed method computes the REM sleep prediction as the percentage of trees that classify the REM sleep for each epoch (30 seconds), calculates TSC of the REM sleep prediction by windowing the REM sleep prediction of a certain number of epochs to smooth them, and the REM sleep estimated by other MLs is corrected when TSC is lower than a certain threshold. Through the human subject experiments, the following implications have been revealed: (1) the proposed method shows a small TSC in the sudden wrong REM sleep estimation, which contributes to correct it; and (2) because of this feature of the proposed method, the number of False-Positive of the REM sleep estimation is successfully reduced, which improves Precision from 51.4% (w/o TSC) to 59.4% (w/ TSC).