An IoT and Machine Learning-based Neonatal Sleep Stage Classification

Abstract

Sleep, in neonates, is used to access the quality of brain and physical development. Typically, neonatal sleep has been divided into three stages: active sleep (AS), quiet sleep (QS), and intermediate sleep (IS). Polysomnography (PSG) is considered a gold standard to classify sleep. To address this issue, over the past two decades, researchers proposed multiple algorithms for automatic sleep stage classification. These algorithms work achieved outstanding results i.e. quiet sleep detection still, lacks in many aspects. One major drawback of the existing research is amalgamation of awake and active sleep into low voltage irregular (LVI) state. This amalgamation corrupts 40% of the overall EEG signal. For this reason, we proposed an algorithm for neonatal sleep-wake classification using machine learning. The proposed research is divided into three steps. Firstly, the EEG signal was pre-processed using finite impulse response filter to remove the noise and artifacts. Clean EEG signal is then divided into 4560 30-sec segments. Then, twenty prominent EEG features were extracted from time, frequency, and spatial domain. After feature extraction, support vector machine was used for sleep stage classification. The propounded study outperformed all the existing algorithms for sleep-wake classification with a mean accuracy of 83.7%. Four-fold cross-validation was used to validate the overall dataset. Multiple other performance matrices i.e. sensitivity, specificity, Kappa were calculated to prove the efficacy of the proposed study. Statistical results show that the proposed study can be used as a real-time neonatal sleep and Awake classification algorithm, as this did not use prior post-processing techniques.