Predicting Sleep Quality Using Lifelog Data with Deep Learning Techniques

Sleep quality significantly impacts overall health. As wearable devices are recently developed and commercialized, a variety of sensors are now available to record lifelog from daily activities. This paper presents a sleep quality estimation method in which health status are classified, including sleep quality, emotions, and stress levels. The proposed method evaluates sleep quality using sensor data such as heart rate, mobile accelerometer, and GPS location. Previous research has shown a close relationship between sleep quality and movements detected from heart rate and accelerometer data during sleep. Our study extends this analysis beyond the sleep period by utilizing data collected throughout the day and integrating multiple data sources. In addition, we employ data augmentation techniques such as time-shifting, adding noise, and addressing class imbalance. By using long short-term memory networks and ensemble methods, we effectively train the proposed model on lifelog sequence data. Extensive experimental results confirm that the proposed method achieves outstanding performance in health status classification using the Human Understanding AI Paper Challenge 2024 dataset.