Sleep Stage Estimation By Evolutionary Computation Using Heartbeat Data and Body-Movement

Transcription

1 Sleep Stage Estimation By Evolutionary Computation Using Heartbeat Data and Body-Movement 1 Hiroyasu MAtsushima, 2 Kazuyuki Hirose, 3 Kiyohiko Hattori, 4 Hiroyuki Sato, 5 Keiki Takadama 1-5, First Author The University of Electro-Communications, 1,2 {matsushima, hirose}@cas.hc.uec.ac.jp, 3,4 {hattori, sato}@hc.uec.ac.jp, 5 keiki@inf.uec.ac.jp Abstract This paper focuses on distinctive changes of not only the heart rate but also the body movement in REM stage (i.e., light sleep) and Non-REM stage (i.e., deep sleep) and improves our sleep estimation method by employing the feature of such distinctive changes. In particular, the heart rate increases irregularly in REM stage, while the heart rate decreases in Non-REM stage. The body moves intensively in REM stage, while the body does not frequently move in Non-REM stage. Using such distinctive changes, we propose a new fitness function which determines the REM/Non-REM stage and introduce it into for our sleep estimation method based on Genetic Algorithms (GAs), which evolve the sleep stage for each person according to the fitness. To investigate an effectiveness of a new fitness function, we compare the estimated sleep stages of our method employing the proposed fitness function with that of Watanabe s method as the conventional method. The experimental results suggest that our method employing the proposed fitness function has a capability to estimate the sleep stage accurately than Watanabe s method without connecting any devices. 1. Introduction Keywords: Sleep Stage Estimation, Genetic Algorithms, Unrestrained Condition To enjoy and maintain the healthy life, it is essential of the human being to achieve good sleep, eating and egestion. In particular sleep, for example, the sleep shortage becomes the factor which obstructs growth and the learning of the teenager. And good sleep is necessary to maintain health of the aged people to live long. In addition, the sleep is important to the people who protect the safety of citizens and the people who have to work the shift for productivity and efficiency. For all people, as noted above, not only health maintenance but also activity in daily life needs to get enough sleep. The sleep stage estimation is studied so far to investigate how deeply a person sleeps. Conventionally, medical specialists estimate the sleep stage according to Rechtschaffen & Kales s (R&K) method [4] using data such as Electroencephalogram (EEG), Electromyogram (EMG), and Electro-oculogram (EOG). However, this type of estimation is subjective and the conclusions of medical specialists using the R&K criteria may differ among them. Furthermore, this method requires a connection to a device to examinee s body, which causes psychological stress. To reduce such a stress, Watanabe proposed the sleep stage estimation without connecting any devices to human s body and succeeded to estimate the sleep stage from (the middle range of) heart rate measured by the pneumatic approach [8]. However, this method requires some parameters measured from the human subject experiments, which means that the accuracy of the sleep stage estimation depends on such parameters (in our prior experiment, we found that the sleep stage estimation of a certain person is correct, while that of another person is not correct). To tackle this problem, Hirose proposed the sleep stage estimation method that can provide an accurate estimation for each person only connecting a wristwatch-type device to human s body. In particular, Hirose s method learns the specific wave pattern of the heart rate by using Genetic Algorithms [1]. From a comparison of Hirose s method [3, 6] with Watanabe s method, the former method has the great potentials of estimating the sleep stage more accurate than the latter method in each person. However, Hirose s method requires the supervised REM data of wrist-watch type device to estimate the accurate sleep stage. International Journal of Advancements in Computing Technology(IJACT) Volume4, Number 22,December 2012 doi: /ijact.vol4.issue

2 To overcome this problem, this paper focuses on distinctive changes of not only the heart rate but also the body movement data in REM stage (i.e., light sleep) and Non-REM stage (i.e, deep sleep) and improves Hirose s sleep estimation method by employing the feature of such distinctive changes. Concretely, we propose a new fitness function which determines the REM/Non-REM stage according to the distinctive changes of heart rate and body movement, and introduces it into Hirose s method. We employ these two data because these two data are related to sleep stage closely. For example, these two data decrease as sleep becomes deep, and these two data suddenly fluctuate at the time of the REM sleep. This paper organized that as follows. The related works are summarized in Section 2. The fitness function as our proposed method and the system including proposed fitness function are described Section 3, Section 4 provides the experiment, Section 5 shows the experimental results, and analyzes the effectiveness of the proposed method. Finally, our conclusion is given in Section Related work 2.1. Rechtschaffen & Kales s method Historically, the sleep stage of humans is firstly used for medical purpose and its stage is estimated by medical specialists using data such as Electro-encephalogram (EEG), Electro-myogram (EMG), and Electro-oculogram (EOG) of humans as shown in Figure. 1. This approach is based on the Rechtschaen & Kales's (R & K) method [4] and can estimate the sleep stage with high accuracy. What should be noted here, however, is that this type of estimation is subjective in terms of depending on medical specialists. More importantly, as shown in Figure 1, this method requires to connect the devices to the head of humans, which derives the psychological stress. In particular, such restraints directly affect nervous peoples who cannot sleep during measurement. This method is, therefore, difficult or impossible to apply to aged persons in the care house. Figure 1 Sensor attachment position for observation EEG, EMG, and EOG 2.2. Watanabe s method To solve the problem of the R & K method, Watanabe proposed the sleep stage estimation method by using the data obtained from the air mattress[8] in a bed as shown in Figure 2. This method employs the unrestrained pneumatic bio measurement (i.e, heart rate, respiration, snoring, and body movement) in bed. Since several reports suggested that the heart rate has the strong relation to the sleep stage [2, 3, 5], Watanabe s method estimates the sleep stage by using the heart rate obtained from the sensor in bed. Concretely, Watanabe s method divides the heart rate into the low frequency range of the heart rate (which is lower than 135min), the middle frequency range of the heart rate (which is between the low and high frequency range of the heart rate), and the high frequency range of the heart rate (which is higher than the half of averaged time of the peak heart rate rate), and then estimates the sleep stage of humans according to the model of low, middle, and high frequency range of the heart rate. Since Watanabe's method does not require any devices to be connected to human's body, it has the great potential of being applied into the estimation of the sleep stage of aged persons by using the heart rate obtained from the air mattress. However, the models of the low, middle, and high frequency range of the heart rate in this method are determined according to the small number of young healthy subject experiments, i.e, the parameters of the models are calculated and set from the human subject experiments. This causes the following problems: (1) the estimated sleep stage may not be accurate, especially for other persons including aged persons, because the parameters used in this method are 282

3 calculated from data of the small number of young persons; and (2) the sleep stage may not be robust to the bad physical condition, because the parameters used in this method are calculated from data of the persons with the good physical condition (not the bad physical condition). 3. Proposed method Figure 2 Watanabe s method Figure 3 shows distinctive changes of heart rate data and the body movement data in REM stage. In Figure 3, vertical axis indicates heart rate and value of body movement, horizontal axis indicates sleep time a night,. The distinctive changes of heart rate data and the body movement data in REM/Non- REM stage are described as follows. In REM stage the heart rate increases erratically, while in Non- REM stage the heart rate decreases. In REM stage the body moves intensively, while in Non-REM stage the body hardly moves. In addition, the body moves strenuously when the sleep stage changes from Non-REM 3 to Non-REM 1 or from REM stage to Non-REM 1. In view of these distinctive changes of heart rate and body movement, the sleep stage is determined by using following requirements. Req. 1: The standard deviation of body movement among 60 seconds more than one among a night. Req. 2: The linear of heartbeat by least squares method among 10 minutes more than the average of heartbeat among a night. Figure 3 determination REM/Non-REM stage by using heartbeat and body movement Table 1 Distinct change of heart rate and body movement in the REM/Non-REM sleep stages distinctive change in the heart rate distinctive change in the body movement REM Heart rate increases Body movement becomes to be active Non-REM Heart rate decreases Body movement becomes to be calm 283

4 Figure 4 Determinate REM/Non-REM sleep stage by using proposed method If both requirements are satisfied, the sleep stage determined as REM or Wake stages. If both requirements are not satisfied, the sleep stage determined as Non-REM stage 3 which is light sleep stage. If both requirements are not satisfied, the sleep stage is determined as Non-REM stage which is deep sleep (e.g., Non-REM stage 2, 3, or 4). Concretely, as shown Figure 4, by using proposed method, REM/Non-REM sleep stages are determined and used as the fitness function of Genetic Algorithms(GAs)[1] for accurate sleep stage estimation. The GAs are powerful and broadly applicable to stochastic search and optimization techniques based on evolutional computation. It is especially appropriate for problems with large and complex search-spaces, where the global optimum cannot be found within a reasonable amount of time using traditional techniques. Since the calculation cost for sleep stage estimation per filter is huge, the filters are searched efficiently by applying GA Sequence of proposed system To estimate the sleep stage, GA is applied as following sequence and shown Figure 5: (1). The proposed GA as shown in Figure 2 records the heart rate and body movement data from the mattress sensor. (2). The proposed GA calculates the frequency of the heart rate data by Fast Fourier Transfer (FFT) (3). Each individual (corresponding to multiple band-pass filters) learns to determine which frequency of the heart rate should be extracted according to 0/1 representation in the bit string; (4). Inverse Fast Fourier Transfer (IFFT) is executed to derive the wave pattern from the extracted frequency of the heart rate; Figure 5 Overview of sleep stage estimation by using proposed method and Genetic Algorithm 284

5 (5). The fitness of all individuals is calculated by comparing the obtained wave pattern with determined sleep stage by proposed method. (6). The deletion and generation operations in GA process are executed to create appropriate individuals and generation counter add 1. If generation counter is equal to limit of generation, system output filters, and sequence is over. If generation counter is smaller than limit of generation back to (3). Note that fitness is calculated as accuracy rate of estimated sleep stages and fitness function by proposed method Discretize to 6 sleep stage Through the above cycle, the heartbeat data is extracted by the individuals with the best accuracy rate of the sleep stage estimation and the extracted wave data by filtering and IFFT is converted in discrete value as following equation. This equation is extended Watanabe's method[8]. In this equation, Sleep Stage, Raw Data, m i, δ i indicate the sleep stage represented by i (1, 2, 3, 4, 5, and 6 correspond to the wake-up stage, REM sleep stage, Non-REM sleep stages 1, 2, 3, and 4, respectively), the raw data of the extracted heartbeat data, the average of the extracted heartbeat data in the sleep stage i, and the standard deviation of the extracted heartbeat data in the sleep stage i, respectively Fitness calculation To estimate the sleep stage accurately by GA, fitness is calculated as follows. Condition 1: When determinate sleep stage by using proposed method is REM stage in a certain second. If estimated stage by GA is Wake or REM stage, fitness is incremented by 1, or if estimated stage is Non-REM stages, fitness is decremented by 1. Condition 2: When determinate sleep stage by using proposed method is Non-REM stage 1 in a certain second. If estimated stage by GA is Non-REM stage 1, fitness is incremented by 1, or if estimated stage is other stages, fitness is decremented by 1. Condition 3: When determinate sleep stage by using proposed method is Non-REM stage 2, 3, and 4 in a certain second. If estimated stage by GA is Non-REM stage 2, 3, and 4, fitness is incremented by 1, or if estimated stage is other stages, fitness is decremented by 1. Additionally, in each second, added value is weighted based on estimated sleep stages. Concretely, when an estimated sleep stage in a certain second is Wake, REM, Non-REM stage 1, 2, 3, and 4, an added value becomes 6 times, 5 times, 4 times, three times, twice, no weighted value, respectively. Since Non-REM stage 1 and 2 appear frequently, it is difficult to estimate REM, Non-REM 3, or 4 by only simple increment or decrement of value. Above conditions are concerned, and fitness add value. Finally summed fitness is divided by total second. 4. Experiment In order to investigate sleep stage estimation system by using proposed method, we compare estimated sleep stages with Alice PDx as a kind of electro-encephalograph and the following cases are conducted in our experiment. Method 1: Watanabe methods Method 2: extended Watanabe method Method 3: our proposed method 285

6 The evaluation criteria are the accuracy rate of estimated sleep stages, which is compared with the Alice PDx in discretized 6 ({Wake}, {REM}, {Non-REM 1}, {2}, {3}, {4}), 5({Wake}, {REM}, {Non-REM 1}, {2}, {3, 4}), and 3({Wake, REM}, {Non-REM1, 2}, {3, 4}) stages. In all cases, heart rate data and Alice PDx data of 5 subjects who are 20 s man are used (i.e, subject A, B, C, D, and E offers data of 10, 8, 7, 4, 3 days respectively, total 32 days). The parameters of GA and used devices are explained as follows. The parameters of GA are set as Table 2. GA is applied in each day and explores individuals had higher fitness, which are more appropriate multi-band pass filter for an applied day. Population size and children size indicates maximum number of individuals and number of generated child respectively. Genetic operations, i.e., cross over and mutation, are applied based on cross over rate and mutation rate. As shown Table 2, the set parameters mean that 250 children of which gene is different from selected parents are generated. Gene length corresponds to used frequency components from low frequency. Table 2 GA Parameters Population size Children size Cross over rate Mutation rate Gene Length Figure 7 shows Emfit sensor. This sensor, developed by VTT Technical Research Center of Finland (VTT) for care support in the 1990 s, is composed of Emfit film under a mattress. The film is based on the sophisticated pressure sensitive sensor and it can measure the data of the heart rate, respiration, and body movement without connecting any devices to the human body. Furthermore, Emfit sensor can detect when person gets out of the bed, and the computer connected with Emfit sensor makes an alarm in such a case. Alice PDx is a kind of the device which can monitor EEG, EOG, and EMG by attaching sensors as shown Figure 6. It satisfies International portable testing requirements (for eg. A.A.S.M.levels II, III and IV) and provides capabilities of basic screening to advanced diagnostic evaluation. The Alice PDx enables patients to be tested outside of the lab without compromising study results and helps clinicians avoid the costs associated with retesting. Figure 6 Emfit sensor Monitoring EEG, EMG, and EOG by Alice PDx 4.3. Result The table 3 indicates the average and standard deviation of accuracy rate in each subject. The bar graphs as shown Figure 7-11 indicate same results as Table 3. In each figure, vertical axis indicates accuracy rate. From these results, in all subjects, proposed method (case 2, i.e, proposed discretization of sleep stages) and proposed system (case 3) are better than watanabe method (case 1). In addition, the standard deviation of case 3 is smaller than one of case 1 and 2 in many days data (i.e, subject A and B). The mean accuracy of case 1 is, %, %, and % for total 32 days of all subjects in discretized 6 stages, 5 stages, and 3 stages respectively. The mean accuracy of case 2 is, %, %, and % for same data as case 1 in discretized 6 stages, 5 stages, and 3 stages respectively. The mean accuracy of case 3 is, %, %, and % for same data as case 1 in discretized 6 stages, 5 stages, and 3 stages respectively. 286

7 5. Discussion Figure 12 shows the comparison Alice PDx and determined REM by proposed method. As shown Figure 12, the timing of determined REM stages by proposed method seems the same timing of the REM stages by Alice PDx, while the timing of Non-REM sages by proposed method also seem the same timing of the Non-REM stages by Alice PDx. From Figure 12, REM/Non-REM sleep stages by the proposed method are approximated actual sleep cycle. In addition, Table. 4 shows average of appearance rate of sleep stages by Alice PDx and proposed system respectively. From Table 4, Non- REM stage 2 is most appearance rate and Non-REM stage 4 is worst appearance rate in both of appearance rate. Other appearance rate also seems approximately in each sleep stages. Both of appearance rate are approximate to the sleep stage appearance of statistical data [7]. Figure 13 shows comparison Alice PDx and estimated sleep stages by proposed system. As shown Figure 13, in this graph vertical axis indicates sleep stages (i.e, Wake, Rem, Non-Rem 1, 2, 3, and 4, respectively), horizontal axis indicates sleep time. The wave of estimated sleep stage by proposed system approximates wave of Alice PDx visually. From result and these analysis, weighted value based on sleep stage in fitness calculation effective for accurate estimation sleep stage. In addition, from results, proposed system is better performance for sleep stage estimation than watanabe method. However, the sleep stage appearance changes depending on the some various factor of human (e.g., age, condition, meal, and so on). In case of this paper, weighted fitness can guide toward approximation for sleep stage appearance of subjects, but, in case of estimation sleep stage for aged persons, weighted fitness may not be able to guide toward approximation sleep stage appearance because it is reported that sleep stage appearance of the aged person is statistically different from young people as subjects in this paper [7]. 6. Conclusion This paper focused on distinctive changes of not only the heart rate data but also the body movement data in REM stage (i.e, light sleep) and Non-REM stage (i.e, deep sleep) and improved our sleep stage estimation method by employing the feature of such distinctive changes. As the first step toward an accurate sleep stage estimation under unrestrained condition, we proposed a new fitness function which determined the REM/Non-REM stage and introduced it into our sleep estimation method based on Genetic Algorithms, which evolve the sleep stage for each person according to the fitness To investigate a new fitness function, we compare the estimated sleep stages of our method employing the proposed fitness function with that of Watanabe s method as the conventional method. The experimental results suggest that our method employing the proposed fitness function has a capability to estimate sleep stage accurately than Watanabe s method without connecting any devices. In detail, the following implications have been revealed: (1) The mean accuracy of sleep stages by using the proposed fitness function is better 10 % than Watanabe s method; (2) sleep stage estimation by using proposed fitness function is more robust than Watanabe s method and previous work; and (3) our proposed fitness function contribute to increase accuracy of sleep stage estimation by using GA. Table 3 Average and standard deviation of accuracy rate Discretized 6 stages Discretized 5 stages Discretized 3 stages Cases Subjects Case 1 Case 2 Case 3 Case 1 Case 2 Case 3 Case 1 Case 2 Case 3 Subject A (10 days) ± ± ± ± ± ± ± ± ±.064 Subject B (8 days) ± ± ± ± ± ± ± ± ±.054 Subject C (7 days) ± ± ± ± ± ± ± ± ±.097 Subject D (4 days) ± ± ± ± ± ± ± ± ±.064 Subject E (three days) ± ± ± ± ± ± ± ± ±

8 Figure 7 Results of subject A in each case Figure 8 Results of subject B in each case Figure 9 Results of subject A in each case Figure 10 Results of subject B in each case Figure 11 Results of subject A in each case The following issues should be pursed in the near future: (1) the verification of the generality of the obtained implications by increasing the number of human subjects; (2) appropriate fitness guidance for aged person. 288

9 Table 4 Average of appearance rate of sleep stages subject A (10 days) subject B (8 days) subject C (7 days) Alice Case 1 Case 2 Case 3 Alice Case 1 Case 2 Case 3 Alice Case 1 Case 2 Case 3 Wake REM Non-REM Non-REM Non-REM Non-REM subject D (4 days) subject E (three days) Alice Case 1 Case 2 Case 3 Alice Case 1 Case 2 Case 3 Wake REM Non-REM Non-REM Non-REM Non-REM Acknowledgement The part of this work was supported by Grant-in-Aid for the Japan Society for the Promotion of Science (JSPS) Fellows. 8. References Figure 12 Alice PDx and determined REM by proposed method [1] Goldberg, D. E.: Genetic Algorithms in Search, Optimization, and Machine Learning, Addison- Wesley (1989). [2] Harper, R. M., Schechman, V. L. and Kluge, K. A.: Machine classification of infant sleep state using cardiorespiratory measures, Electroencephalogr: Clin. Neaurophysiol., No. 67, pp , [3] Hirose, K., Matsushima, H., Hattori, K., and Takadama K.,: Sleep Stage Estimation by Learning Classifier System towards Care Support for Aged Person, ICROS-SICE International Conference, pp , (2009) 289

10 [4] Rechtschaffen, A. and Kales, A. (Eds): A Manual of Standardized Terminology, Techniques and Scaring System for Sleep Stage of Human Subjects, Pulbic Health Service U. S. Government Printing Office, [5] Shimohira, M., Shiiki, T., Sugimoto, J., Ohsawa, Y., Fukumizu, M., Hasegawa, T., Iwakawa, Y., Nomura, Y., and Segawa, M.: Video Analysis of Gross Body Movements During Sleep, Psychiatry Clin. Neurosci., Vol. 52, No. 2, pp , [6] Takadama, K., Hirose, K., et al.: Learning Multiple Band-Pass Filters for Sleep Stage Estimation: Toward Care Support for Aged Persons, IEICE Transaction on Communications, Vol.E93-B, No.4, pp , April, [7] The Japanese Society of Sleep Research (Eds): Suimingaku, Asakura publishing co. Ltd, [8] Watanabe, T. and Watanabe, K. Noncontact Method for Sleep Stage Estimation, IEEE TRANSACTION ON BIOMEDICAL ENGINEERING, No. 10, Vol. 51, pp (2004). 290