Contents I. Introduction
Polysomnography is the most effective way to evaluate a sleep quality. This test consists of EEG, ECG, EOG, EMG and videotaping. However, it is conducted by medical specialists and requires overnight hospital stay, which consequently leads to high cost. Therefore, less expensive and user friendly wearable devices are desirable as readily available consumer electronic products. Among them, accelerometer based wrist-worn device has been introduced because it is known to users as reliable and convenient to use as data collection can be continuously made [1]. It enables a user to collect data without any disruptions. In this paper, an effective method to estimate sleeping time using a tri-axial accelerometer device is proposed. Most conventional methods for sleep and wake time estimation are based on regression. Regression based methods calculate weighted sum of feature values of certain epoch (e.g. 20 seconds or 1 minute) [2], [3]. For regression, some useful features have been suggested including Proportional Integration Mode (PIM), Zero Crossing Mode (ZCM), and Time Above Threshold (TAT) [4]. In the conventional framework, once the feature value exceeds the threshold the state of the epoch is determined as “wakefulness” and otherwise is “sleep”. This is essentially a static classification with 2-state binary decision. A weakness of this method is that it determines a state of epoch according to the corresponding feature value. So, if a person wearing accelerometer intermittently moves while sleeping or slightly moves while awake, it is likely to misclassify the state. In this paper, a robust dynamic classifier is proposed to address the dynamic nature of intermittent motions within both the “wakefulness” and “sleep” states by adding a classification stage which considers similarity between scores of adjacent epochs acquired by the static classifier.
