The AAAI 2017 Spring Symposium on Wellbeing AI: From Machine Learning to Subjectivity Oriented Computing Technical Report SS-17-08 Sleep Stage Estimation Based on Approximate Heartrate Calculated from Other Persons Yusuke Tajima Graduate School of Informatics and Engineering, The University of Electro-Communications y_tajima@cas.hc.uec.ac.jp Tomohiro Harada College of Information Science and Engineering, Ritsumeikan University harada@ci.ritsumei.ac.jp Keiki Takadama Graduate School of Informatics and Engineering, The University of Electro-Communications keiki@inf.uec.ac.jp Abstract This paper focuses on the sleep stage estimation method based on the approximate heartrate calculated from one person, and improves its estimation accuracy by employing the approximate heartrate calculated from other persons. Concretely, the proposed approximate heartrate is a weighted summation of the approximate heartrate calculated from other persons. Through the human subject experiments, the following implications have been revealed: (1) the accuracy of the sleep stage estimation method based on the approximate heartrate calculated from other persons is higher than that of the conventional method based on the approximate heartrate calculated from one person; and (2) the accuracy of the sleep stage estimation increases as the number of the heartbeat data of other persons for calculating the approximate heartrate increases, i.e., the accuracy of the sleep stage estimation employing the heartbeat data of the six other persons is higher than of two other persons, one other person, and none of other person Introduction Since sleep is a fundamental activity which accounts for about one-third of the activity in a day, the sleep quality affects condition of both the physical and mental health. From this fact, sleep is one of the most important activities for human beings, and understanding of our sleep is indispensable to check whether we do not have some sleep troubles. For this purpose, the sleep stage is useful to understand our sleep, and the Rechtschaffen and Kales (R&K) method (Kales and Rechtschaffen 1986) as the gold Copyright 2017, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. standard method is widely employed in polysomnogram. In this method, the sleep stage is determined according to the characteristics of vital data obtained from the electroencephalogram (EEG), the electromyogram (EMG), and the electrooculography (EOG). The R&K method is useful to provide the sleep stage, but it requires not only a connection of the highly constrained devices to the human body but also require a support of medical experts to determine the sleep stage. This indicates that daily sleep is hard to be evaluated due to a connection of many devices and we cannot understand our sleep without medical experts. To tackle this problem, Watanabe developed an air mattress biosensor which can acquire the vital data (such as heartbeat, body movement, and respiration) without connecting any devices to human body and can roughly estimate the sleep stage without medical experts (Watanabe 2004). In detail, the sleep stage is mainly estimated according to the middle frequency range of the obtained heartbeat data (note that body movement data is also used to estimate the sleep stage). To increase an accuracy of the sleep stage, Takadama proposed the sleep stage method which can adjust its stage for each person by estimating the appropriate middle frequency range of the heartbeat data for each person (Takadama 2010). Since these methods cannot estimate the sleep stage in real-time, Harada proposed the method which can achieve it by the trigonometric function approximation (Harada 2016). The above three methods are very useful from the viewpoint of the sleep stage estuation with neither a connection of any devices to human body nor a support of medical experts, but what should be noted here is that the accuracy of these three methods is not enough high in comparison with 734
the global standard R&K method. To tackle this problem, this paper focuses on the Harada s sleep stage estimation method which is based on the approximate heartrate calculated from one person, and improves its estimation accuracy by employing the approximate heartrate calculated from other persons. The remaining of this paper is organized as follows. First, the previous works related to the sleep stage estimation are introduced in Section 2. Section 3 shows the proposed method that employs the approximate heartrate calculated from other persons. Section 4 shows the conducted subject experiment and its result is shown. Finally, the conclusion of this paper is given in the final section. Sleep Stage and its estimation method The sleep stage is an index of the sleeping activity and it is represented by a numerical value meaning the depth of sleep. As shown in Figure 1, the sleep stage is classified into six types as follows: (i) awake (W) as the lightest sleep; (ii) REM sleep (R) which is deeper sleep than awake and it is generally occurred in dreaming time; (iii) Non-REM sleep 1~4 (Non1~4) as deep sleep (note that the depth of sleep becomes deeper from Non1 to Non4). Figure.1 shows the sleep stage diagram of one subject during sleep. In this figure, the horizontal axis indicates the sleeping time, while the vertical axis indicates the sleep stage. or respiration measured by mattress biosensor without connecting any devices to human body (e.g., Watanabe s method, Takadama s method, and Harada s method). The two methods from each category are explained in the following subsections. Rechtschaffen and Kales method Rechtschaffen and Kales (R&K) method was proposed to calculate the quality and quantity of sleep as an objective numerical value (Kales and Rechtschaffen 1986). Concretely, the medical experts determine the sleep state which is classified into the six stages from a viewpoint of depth of sleep estimated by the electroencephalogram (EEG), the electromyogram (EMG), and the electrooculography (EOG). Because the accuracy of the sleep stage in this method is enough for a medical purpose, this method has been widely employed as the global standard method. But as described in Section 1, this method requires to connect many devices (including many electrodes and codes) to human body, which increases the stress of human subjects. Trigonometric Function Approximation Method In order to estimate the sleep stage from heartrate, the Harada s method starts to approximate the heartrate as the trigonometric function and estimates the sleep stage from the intermediate frequency range of the approximate heartrate. The approximate heartrate is modeled as follows, where h(, ) denotes the estimated heartrate at time with the model parameter, L denotes the maximum period of the intermediate frequency component, N denotes the number of composed trigonometric functions, and is the model parameter = {,, } (n {1,, N}). Figure.1 Sleep stage in one day Sleep Estimation methods As described in Section 1, the sleep stage determination /estimation methods are classified into the following category: (i) the sleep stage determined according to vital data obtained from the electroencephalogram, eye movement, and electromyogram through a connection of devices to human body (e.g., R&K method); (ii) the sleep stage estimated according to vital data of heartbeat, body movement, The model parameters = {,, } are calculated by the maximum likelihood estimation method that minimizes the following likelihood function, where T denotes the elapsed time after falling asleep, HR(t) denotes the obtained actual heartrate at time t. Note that the second term denotes the regularization term, which contributes to not find the large parameters which only fit to the current actual heartrate as specialized parameters. After calculating the parameters, the sleep stage is estimated by discretizing the approximate heartrate h(t, ) according to the following equation, where s(t) denotes the 735
sleep stage at time t, ave and stdev denotes the average and the standard deviation of the approximate heartrate h(t, ), denotes the ceiling function that returns the minimum integer value which is equal to or larger than x, and from 5 to 0 correspond to WAK, REM, Non1, Non2, Non3, and Non4, respectively. This discretization formula is based on the previous research (Takadama 2010). Proposed method As described in Section 1, this paper improves the accuracy of the sleep stage estimation of Harada s method (which is based on the approximate heartrate calculated from one person) by employing the approximate heartrate calculated from other persons. Estimating using by Coefficient The Harada s method can estimate the sleep stage in realtime, it has a problem of over fitting to the actual heartrate especially in the case of a small number of heartbeat data. In order to overcome this problem, this paper focus on the model parameter = {,, } of other persons as the prior knowledge. This is because such knowledge has a potential of determining the rough sleep stage. However, the parameters of other persons are not always to be suit to all persons. From this fact, it is important to create appropriate parameters for some persons from the parameters of other person. Approximate heartrate minimum method In order to decide appropriate the prior knowledge regard less of subjects and estimation dates, this paper proposed the approximate heartrate minimum method. Figure 2 shows the overview of this method. This method calculates the similarity between the shape of estimating currently sleep stage and the shape of the sleep stage measured in the past. The calculation formula of the similarity is as shown in the formula 1. f(x) is the shape of estimating currently sleep stage, and g(x) is the shape of the sleep stage measured in the past. L means the data length of the measured heart rate, n means the date of sleep data. The weighted average of the degree of the similarity to each sleep is calculated by the formula 2 from the calculated similarity. In the formula, m is the total number of sleeping days. By multiplying this weighted average value and the coefficient indicating each the sleeping shape, it is possible to obtain the prior knowledge in considering the degree of similarity. In this method, estimating the sleep stage using this prior knowledge. Experiment Settings To investigate the effectiveness of the proposed method we conduct the subjects of experiments. Table. 1 shows details of three subjects. Figure.2 Overview of the proposed method 736
Table. 1 Details of subjects age Sex Subject A 20 Male Subject B 40 Male Subject C 60 Female This experiment uses Alice PDx and EMFit which are the biometric sensor to obtain the sleep data. In figure. 4, the above figure shows Alice PDx which is a kind of the electro-encephalograph, the below figure shows EMFit which is non-contact biosensor to measure subject's heart rate, body movement, respiration by being laid under a bed mattress. The data using by the Alice PDx is used to calculate the sleep stage by R&K method, the data using by the EM- Fit is used to estimate the sleep stage by proposed method and trigonometric function approximation method. REM sleep 4 are combined, in other word, this experiment evaluates in 4 stages instead of 6 stages. Results Table. 2 Parameters in the trigonometric function approximation method Parameter Value 1 60[sec] = 1[min] 13 Figure 5 shows the results of three days of same subject and figure 6 shows the results of one day of three subjects. In each figure, it shows not using the knowledge, the using one other's knowledge, using the two other's knowledge, and using six other's knowledge. The value on the vertical axis is match rate. In the figure 4, the more you use sleeping knowledge, the higher the accuracy is. In the figure 5, it turns out that the result of the same trend appears. From the above, it can be seen that the degree of sleep information affects the estimation accuracy regardless of the subject / estimated date. Figure.3 The experimental tool in the experiments In the trigonometric function approximation method, parameters for estimation are determined as Table. 2. In the proposed method, this experiment is conducted with two viewpoints. 1) The influence of the number of data used in the same subject, 2) The influence of the number of data used in different subjects. In particular, this experiment verifies the difference of precision by used one sleep data, two sleep data, or six sleep data. Evaluation criteria The sleep stage which is estimated by the proposed method compared with the sleep stage is estimated by the R&K method. Because the R&K method is the gold standard method all over the world, the sleep stage estimated by this method is treat correct. In this experiment, the sleep stage which is derived by proposed method compares the correct sleep stage. In comparison, REM sleep 1 and REM sleep 2, REM sleep 3 and Discussion Consider a big factor that leads to good estimation by finding coefficients using more knowledge of others. The knowledge of others is less similar than their own knowledge, and the possibility of applying it is low. Therefore, it is considered that it is caused by not being adapted to the estimator when only one knowledge of others is used. Conversely, when many data are used, it is considered that the characteristic part of another person disappears by the weighted average formula, and it is possible to derive human's own sleep information. Therefore, highly accurate estimation can be made for other people without similarity. Conclusion This paper focused on the sleep stage estimation method based on the approximate heartrate calculated from one person, and improved its estimation accuracy by employing the approximate heartrate calculated from other persons. Concretely, the proposed approximate heartrate is a weighted summation of the approximate heartrate calcu- 737
Figure.5 Estimation results of multiple days in the same subject lated from other persons. Through the human subject experiments, the following implications have been revealed: (1) the accuracy of the sleep stage estimation method based on the approximate heartrate calculated from other persons is higher than that of the conventional method based on the approximate heartrate calculated from one person; and (2) the accuracy of the sleep stage estimation increases as the number of the heartbeat data of other persons for calculating the approximate heartrate increases, i.e., the accuracy of the sleep stage estimation employing the heartbeat data of the six other persons is higher than of two other persons, one other person, and none of other person. Figure.6 Estimation results of each day in the three subjects What should be noted here is that the above results have only been shown from an insufficient number of human subjects. This suggests that further careful qualifications and justifications, such as an increase of the number of human subjects, are needed to generalize our results. Such important directions must be pursued in the near future in addition to the following future research: (1) The effect of re-estimation at places where sleep rhythm changes; and (2) The effect of body movement and respiration data to sleep stage. 738
Acknowledgments The research reported here was supported in part by a Grant-in-Aid for Scientific Research (A, No. 15H01720) from the Ministry of Education, Culture, Sports, Science and Technology (MEXT). References Harada, T., Uwano, F., Komine, T., Tajima, Y., Kawashima, T., Morishima, M., and Takadama, K 2016. ``Real-time Sleep Stage Estimation from Biological Data with Trigonometric Function Regression Model,'' The AAAI 2016 Spring Symposia, Well- Being Computing: AI Meets Health and Happiness Science, AAAI The Association for the Advancement of Artificial Intelligence), pp. 348-353. Harper, R. M.; Schechtman, V. L.; and Kluge, K. A. 1987. Machine classification of infant sleep state using cardiorespiratory measures. Electroencephalography and Clinical Neurophysiology 67(4):379 387. Kales, A., and Rechtschaffen, A., eds. 1968. A manual of standardized terminology, techniques and scoring system for sleep stages of human subjects. Bethesda, Md.: U.S. National Institute of Neurological Diseases and Blindness, Neurological Information Network. Shimohira, M.; Shiiki, T.; Sugimoto, J.; Ohsawa, Y.; Fukumizu, M.; Hasegawa, T.; Iwakawa, Y.; Nomura, Y.; and Segawa, M. 1998. Video analysis of gross body movements during sleep. Psychiatry Clin Neurosci 52(2):176 177. Takadama, K.; Hirose, K.; Matsushima, H.; Hattori, K.; and Nakajima, N. 2010. Learning multiple band-pass filters for sleep stage estimation: Towards care support for aged persons. IEICE Transactions on Communications E93.B(4):811 818. Watanabe, T., and Watanabe, K. 2004. Noncontact method for sleep stage estimation. Biomedical Engineering, IEEE Transactions on 51(10):1735 1748. 739