Fig 1.
Process of calculating the wear probability ().
(a) original data recorded by the wearable device (black line) and non-wear time reported by the behavioral record (gray bands) (b) the probability of the nonzero value () appearing in each 60 minutes time window during the forward and backward 60 minutes sliding windows (c) the wear probability (
) at each time was calculated as the mean value of
at each time.
Fig 2.
(a) The probability of the appearance of zero counts relative to the total time of sedentary behavior () was calculated by the total time for each zero count and sedentary behavior. (b) The optimal threshold was estimated as the function of
. Scatters show the relationship between the
and the optimal threshold, and the color map shows the probability of the distribution of the generated data by the simulation for data augmentation.
Fig 3.
The histogram of consecutive zero counts (gray bar), probability density function (solid line), and cumulative probability function (dotted line) during the wear time.
Table 1.
Contingency table of wear/non-wear classification.
Fig 4.
(a) original data (gray) and corrected data by the proposed method (red). Gray bands indicate the non-wear time recorded by the behavioral record. When both the labels (
) indicate wear (b and c), the time that shows zero counts were re-classified as wear time (
) (d).
Fig 5.
Performance (recall, specificity, precision, and accuracy) of the original, current, and proposed methods.
The left panels show the relationship between the probability of the appearance of zero counts and the performance index. The right panels show the box plot of the original, current, and proposed method.
Fig 6.
(a) The original data was recorded under the free-living conditions in a pregnant woman, (b) the data corrected by the current method, and (c) the data corrected by the proposed method. The gray bands indicate the non-wear time recorded by the behavioral record. In the current method, the non-wear time up to 60 minutes was misclassified as the wear time and corrected as the sedentary behavior (b). Indeed, the proposed method classified the actual non-wear time correctly (c). However, the misclassification was observed before 12:00 due to the over 10 minutes of consecutive zero counts (c).