A Machine-learning-based Sleep-wake Detection Method from Physical Acceleration Signals

Abstract

We aimed at developing a novel and accurate sleep-wake state detection method using physical acceleration data measured by a wearable device. In this study, we conducted simultaneously measurement of activity data using an AMI actigraph and a tri-axial acceleration monitor (12 healthy adults; >24 h). Actigraph has an equivalent performance in sleep-wake state detections to traditional sleep polygraph tests. We therefore used outputs of the actigraph as a teacher signal for the succeeding analysis based on a supervised machine learning (support vector machine; SVM). The feature vectors, input data to SVMs, were automatically selected from various statistics extracted from the acceleration signals using neighboring component analysis; signal-magnitude-area and trunk-angle were selected as optimal features. The k-fold and one-leave-out cross-validation method were adopted. We obtained satisfactory results from the constructed model with higher coincidence with actigraph (accuracy: 94.6 ±4 %, specificity: 95.3 ±2 %, sensitivity 95.6 ±3 %).