Enhancing sleep postures classification by incorporating acceleration sensor and LSTM model

It has been well established that sleep posture plays a key role in sleep quality monitoring. Consequently, many noncontact and wearable devices, whose systems rely on sensors such as cameras, radar, wireless, and accelerometers, have been developed to classify sleep positions and postures. However, noncontact systems were often unsuccessful when facing limited conditions such as low light and physical obstacles. On the other hand, other systems currently in research, which involves wearable devices, may have used machine learning models but have not competently exploited other more accurate deep learning models. Recognizing scope for improvement, we propose an enhanced five-sleep-posture classification system (5-SPCS) where a novel integration of accelerometer and an LSTM deep learning model can classify sleep postures more efficiently than either one of them does separately. Our experiments showed that the 5-SPCS was capable of outperforming the baselines of existing machine learning-accelerometer systems at 99.6% accuracy.