Feature Analysis to Estimate Sleep Time Based on Simple Measurement of Biological Information after Awakening

Currently, many people wear a wristband type device while sleeping to automatically record how many hours they sleep. Even a system without a wearing device, such as a smartphone application, needs to be set in advance. Therefore, automatic recording of sleep time cannot be realized without advanced measurement preparation. In this study, we propose a method to estimate sleep time without advanced preparation based on a simple measurement of biological information after awakening. We extracted 97 types of features from sensor data that were measured using wearable devices. We analyzed whether significant differences between each feature appear according to the previous sleep time. Furthermore, we evaluated the accuracy when the sleep time is estimated by machine learning using features with a significant difference. We adopted Support Vector Machine (SVM) as a machine learning algorithm and Leave-One-Session-Out Cross Validation (LOSO-CV) as an evaluation method. Consequently, there were seven features with significant differences when the biological information was measured one hour after awakening. By using machine learning, the accuracy of the previous sleep time (three sleep time categories: short, medium, or long) was estimated to be 62.5%.