A Novel Sleep Scoring Algorithm-Based Framework and Sleep Pattern Analysis Using Machine Learning Techniques

Introduction

The preeminence of having proper sleep is of the prime aspect for maintaining proper health. In the last two decades, the lifestyle of people has changed rapidly, either because of the customed busy life or because of the fine-tuned arrangement of living. People these days are quite reluctant to maintain or keep a check on the duration or the quality of sleep that they are having. Such kind of delinquency has often led to multiple sleep disorders and often lead to the stage of insomnia, apnea (Balakrishnan et al. (2005)) and a multitude of health intricacies. Therefore, maintaining a check on the quality, duration, and efficiency of sleep is an important routine that each individual needs to follow for seeking a proper health condition (Cheng and Mei, (2012)).

Detection and identification of sleep have always been an active area of research and also the area has been acknowledged enough for different methods and processes to analyze the duration and efficiency of sleep quantitatively. As we know that polysomnography is considered as the “Paragon of Excellence” (Alickovic and Subasi, (2018)) for determining sleep quality and sleep stage classification, but the main complication behind polysomnography is the number of sensors, cables, and devices that are attached to the body and also the complete task needs to be carried out in a specific laboratory environment which even leads to basic distress to the subject who is carrying out the procedure. Therefore, a plethora of smartwatches and wearables are now available in the market for maintaining a proper check on the efficiency of the sleep and also to dictate terms of organizing the time to manage it as a whole.

For the wearable devices (Sathyanarayana et al. (2016) and Yeo et al. (2017)) to determine sleep efficiency and the sleep duration, the developers predominantly used the tri-axial accelerometer that is fetched from the wearable devices and determines a particular state or an epoch as a slept stage or awake stage. This classification of each epoch is computed to determine the efficiency and duration of the complete sleep cycle (Sadeh et al. (1994)). The purpose of this study is actually to develop a more generic algorithm for detecting and classifying every epoch of the data to be slept or awake. Also, the work tends to perform a comparative analysis between previous algorithms and also to perform a cross-technique analysis between rule-based methods and machine learning-based models. The method is completely based on statistical analysis which gives it the leap of behaving as a generic model, which can classify the epochs either for data that is being generated by the sensors present in a mobile phone or smart wearable devices.

The system defined in the work consumes the results from the algorithm as ground truths and further simplifies the complete system into an intelligent engine as it leverages machine learning algorithms to classify a particular epoch as slept or awake. The model is trained with statistical features (Yeo et al. (2017), and Aich et al. (2019)) that have been generated from the data and further trained on different algorithms for fetching higher recall and higher precision.

The paper is structured as follows: The next section shows the previous study and also depicts the types of methods developed for either representing the same problem or using the same sensor technology to address different problems. the third section typically is the one that addresses the complete development mechanism of the system and also explains the flow of control of the system starting from data extraction to the final model. This section also places forward the complete development mechanism of the algorithm and also examines the complexities associated with the algorithm for further research that may rise to concern while performing any further research on it. Moreover, in the particular segment, we also try to devise some feature engineering methods and feature extraction methods for feeding the data to the machine learning model in due course of the paper. The fourth section mainly plots the complete results obtained from the system and also to perform a comparative analysis between previous algorithms and the newly developed algorithm. the fifth section demonstrates the discussion on the developed system and, manages to argue upon the indifferences in other algorithms that developed the motivation for this particular work. The sixth and final section puts forward the conclusion.