Robust Sleep Quality Quantification Method for a Personal Handheld Device
Hangsik Shin, PhD,1 Byunghun Choi, PhD,2 Doyoon Kim, PhD,2 and Jaegeol Cho, PhD2 1
Department of Biomedical Engineering, Chonnam National University, Yeosu-si, Jeollanam-do, Republic of Korea. 2 Digital Media and Communication R&D Center, Samsung Electronics Co. Ltd., Suwon-si, Gyeonggi-do, Republic of Korea.
Abstract Objective: The purpose of this study was to develop and validate a novel method for sleep quality quantification using personal handheld devices. Materials and Methods: The proposed method used 3- or 6-axes signals, including acceleration and angular velocity, obtained from built-in sensors in a smartphone and applied a realtime wavelet denoising technique to minimize the nonstationary noise. Sleep or wake status was decided on each axis, and the totals were finally summed to calculate sleep efficiency (SE), regarded as sleep quality in general. The sleep experiment was carried out for performance evaluation of the proposed method, and 14 subjects participated. An experimental protocol was designed for comparative analysis. The activity during sleep was recorded not only by the proposed method but also by well-known commercial applications simultaneously; moreover, activity was recorded on different mattresses and locations to verify the reliability in practical use. Every calculated SE was compared with the SE of a clinically certified medical device, the Philips (Amsterdam, The Netherlands) Actiwatch. Results: In these experiments, the proposed method proved its reliability in quantifying sleep quality. Compared with the Actiwatch, accuracy and average bias error of SE calculated by the proposed method were 96.50% and - 1.91%, respectively. Conclusions: The proposed method was vastly superior to other comparative applications with at least 11.41% in average accuracy and at least 6.10% in average bias; average accuracy and average absolute bias error of comparative applications were 76.33% and 17.52%, respectively. Key words: sleep monitoring, sleep efficiency, sleep quality, wavelet, healthcare application
Introduction
G
ood sleep is essential for human health and well-being. Sleep is a restorative activity of the brain; thus, sleep deprivation increases the risk of mental disorders such as depression and anxiety. Moreover, insufficient sleep
522 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
reduces the desire and the motivation for physical activity, contributing to weight gain, obesity, and associated disorders.1–4 Sleep is also related to memory and learning. During slow wave sleep, memory is consolidated, which means that the memory pieces of what we have learned during the day come together coherently so that the information can be accessed when needed. In particular, sleep disruption could delay development of children or infant because growth hormone is secreted during slow wave sleep.5 Furthermore, sleep disruption is associated with a higher incidence of behavioral problems in children such as attention deficit and hyperactivity disorder.6 In spite of the importance of good sleeping, millions of people do not get enough sleep, and many suffer from lack of sleep. For example, surveys (see www.apa.org/topics/sleep/ why.aspx) conducted by the National Sleep Foundation (1999–2004) reveal that at least 40 million Americans suffer from over 70 different sleep disorders, and 60% of adults report having sleep problems a few nights a week or more. Most of those with these problems go undiagnosed and untreated. In addition, more than 40% of adults experience daytime sleepiness severe enough to interfere with their daily activities at least a few days each month, with 20% reporting problem sleepiness a few days a week or more. Furthermore, 69% of children experience one or more sleep problems a few nights or more during a week. To improve sleep habit and quality, sleep monitoring techniques have undergone continuous development. Polysomnography (PSG), a representative sleep monitoring technique, measures multiple parameters during sleep, including the electroencephalogram, electromyogram, electrocardiogram, and respiration, and analyzes sleep in detail. For example, PSG reports parameters such as quality of sleep, muscle tensing, cardiac activity, respiratory disorder, sleep apnea test, oxygen saturation, and leg movement or body position during actual sleep.7 Although PSG is a reference technique in analyzing sleep, several limitations still remain in practical daily use. First of all, PSG requires numerous electrical leads, which disturb sleep, in order to measure physiological signal, and circumstances such as temperature, humidity, sound level, or light intensity must be controlled to perform PSG properly. Most of all, PSG is hard to conduct individually and daily because the PSG equipment is high-priced and complex to set up. For these reasons, PSG has just been conducted by specialized agencies, and consecutive measurement is almost impossible. As an alternative method of PSG, several studies have been carried out to validate the availability of actimetry for assessing sleep and waking. Actimetry is essentially designed for ambulatory activity, recording acceleration signals, and it is also used for sleep
DOI: 10.1089/tmj.2013.0216
SLEEP QUALITY MEASUREMENT TECHNIQUE
monitoring. Actigraphy is less expensive, noninvasive, and more conducive to repeated measures in comparison with PSG.8,9 In the last two decades, actimetry has found good agreement (97%) for total sleep time10 and overall agreement rates of 91–93%11,12 compared with PSG. Moreover, it is strongly correlated in normal sleepers and in sleepers with apnea, with correlation coefficients ranging from 0.89 to 0.98.9,13–17 Additionally, there is much evidence related to the clinical usefulness of actigraphy in many different fields of sleep study.18–26 Actimetry dramatically reduces the complexity of sleep monitoring; it is not only simple to use, but also reasonably priced. However, actimetry still necessarily requires PC-based software for device setup and checking up on the result; moreover, it is hard to use without additional training for operation. Also, consumers may not want to purchase the device because the functions of actimetry are quite limited to actigraph recording. Recently, various types of sleep management applications have been unveiled based on personal smart devices. These applications calculate the sleep efficiency (SE) using an acceleration signal measured by a device-embedded accelerometer and provide the result with an intuitive user interface. Some applications estimate the sleep stage and control the alarm to wake up at the proper time. The number of downloads of sleep management applications is increasing consistently because of the ease of approach by the application store, the low price, and the ease of use with an intuitive user interface. Likewise, the sleep monitoring technique has rapidly developed with personal handheld devices. However, we should emphasize that there are no reports to evaluate the clinical suitability of these methods, and most sleep monitoring applications have strict operation guidelines: the device should be located in a designated location such as a side of the head or should be placed in a specified direction. Functional examination of the sleep monitoring application should be designed to give practical assistance to the user. The purpose of this article can be summarized as follows: the first is to develop an improved method for calculating SE based on a smart device, the second is to evaluate the representative smart device–based sleep monitoring applications in practical situations, and the last is to verify the performance of each application compared with a reference device, the Actiwatch (Philips, Amsterdam, The Netherlands).
Materials and Methods
were validated for significance based on the paired t test. SE was calculated by the following procedure: (1) movement sensing with the 3-axes accelerometer and the 3-axes gyroscope (total 6-axes movement signal), (2) noise reduction with the discrete wavelet transform–based wavelet denoising technique, (3) sleep–wake decision with empirical threshold, and (4) SE calculation. This procedure is described in Figure 1.
MOVEMENT SIGNAL ACQUISITION We used two types of sensors in movement recording: the accelerometer and the gyroscope. These sensors are similar to recording motion, but they have different characteristics and specifications. The accelerometer measures the acceleration, including gravity. Thus the acceleration value keeps a steady constant, and it changes with movement and sensor position. In other words, the value of acceleration could be changed by phone direction and tilting without motion. The gyroscope is a sensor for measuring angular velocity. It has zero value basically and just is a constant when motion is occurring. The output value, of course, could change in sensor direction; it is not important in recognizing whether movement is occurring or not. We have measured both signals to distinguish wake status from sleep. Acceleration signal and angular velocity were recorded with built-in Samsung GT-I9300 sensors, and sampling frequency was 16.6667 Hz.
NOISE REDUCTION BASED ON REAL-TIME DISCRETE WAVELET TRANSFORM Acceleration and angular velocity measured from sensors included small variations from manufacturing characteristics, and such variation could be regarded as a noise in discriminating motion. Generally, this natural noise has no effect in detecting obvious movement because it has low amplitude near the baseline and has random characteristics. However, in some cases, the motion-induced signal is very small, so it is hard to discriminate movement signal from natural noise components. Therefore, a noise reduction algorithm was necessary to improve the signal-to-noise ratio of the measured signal. We use wavelet transform to reduce the occurrence of noise naturally. The wavelet is a useful mathematical function and is usually used in digital signal processing and image compression. Also, the wavelet could be used for spatial–temporal analysis of the signal based on mother function and scale functions. From these characteristics, the wavelet transform is very useful to analyze a nonstationary signal and becomes an alternative of short-term Fourier transform and Gabor transform. We have chosen the discrete Haar wavelet, which is the most general and simple, to discriminate movement signal and noise.
In our experiment, the Samsung (Gyeonggi-do, Republic of Korea) GT-I9300 (Galaxy S3) was used for signal acquisition and analysis. The signal processing algorithm was developed with MathWorks (Natick, MA) MATLAB 2011b and converted to Java programming language for loading to the Android (Google, Mountainview, CA) OS. The analysis application was designed to record and to analyze real-time, and it was installed on the GT-I9300 for the experiment. IBM (Armonk, NY) SPSS Statistics version 20 was used for statistical analysis, and data Fig. 1. Procedure of sleep efficiency calculation. DWT, discrete wavelet transform.
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 523
SHIN ET AL.
Discrete Haar wavelet transform is composed with two filters: lowpass filter and high-pass filter. The outputs are divided into the approximation coefficients of the low-pass filter (gn[n]) and teh detail coefficients of the high-pass filter (hn[n]). Each filter can be described in as Eqs. 1 and 2, respectively: 1 gk [n] = pffiffiffi (gj - 1 (2k) + gj - 1 (2k + 1)) 2
(1)
1 hk [n] = pffiffiffi (gj - 1 (2k) + gj - 1 (2k + 1)) 2
(2)
We decompose an input signal into six levels and denoise with threshold. The penalized threshold was used to set a noise rejection threshold. The penalized threshold is a level-wise threshold method in which the detail coefficients could be sorted in descending order. The threshold k could be computed with Eq. 327: " # t n 2 2 k = ar gmin + d + 2r t a + ln , t = 1, . . . , n (3) t k t k=1 where d is the detailed coefficient, r is the noise standard deviation, n is the signal length, and a > 1 is the sparsity parameter. Figure 2 shows the decomposition procedure with the six-level filter bank. In the case of a sleep monitoring system, the power consumption and short delay are very important issues because the handheld device is always turned on during monitoring; moreover, the resources of handheld devices are severely limited. Therefore, we have implemented a real-time wavelet filter for reducing the computational load. By sharing low- and high-pass filtering in every cycle, the computational load of the reporting cycle, which means every 32 cycles, could be decreased. Related contents are briefly shown in Figure 2.
SLEEP AND WAKE DECISION We have postulated that the sleep and wake status depends on movement during sleep. Thus, we have defined ‘‘sleep’’ as the ‘‘movement signal was not detected from all sensors’’ and ‘‘wake’’ as the ‘‘movement signal was detected from at least one sensor.’’ To determine if a movement has occurred or not, the empirical threshold was applied to the wavelet denoised signal on each axis. The threshold could be adjusted by changing the environment such as bed type and sensor position. In our experiment, we set *10–3 as a threshold value. Sleep and wake detection was performed with the following steps: (1) detect movement with the threshold from each x-, y-, z-axis acceleration signal and each x-, y-, z-axis angular velocity (a total of 6 axes), (2) summing up the detection results of 6 axes, and (3) decision sleep and wake with summing up the result.
SE CALCULATION SE, which is widely used to evaluate the quality of one’s sleep, is the ratio of time spent asleep (actual sleep time) to the amount of time spent in bed, and it could be calculated as the number of minutes of sleep divided by the number of minutes in bed. Generally, SE is approximately 85–90% or higher in normal adults, but the ‘‘laboratory effect’’ allows for a 75% efficiency in the sleep lab environment. Equation 4 describes the formula to calculate SE: Tsleep Twake SE(%) = · 100 (4) · 100 = 1 Trec Trec where Tsleep is the total sleep time, Twake is the total wake time, and Trec is the total recording time. In this study, Eq. 4 can be modified to Eq. 5 because sleep and wake status values were reported in every 1-min epoch: # epochsleep · 100 # epochrec # epochwake = 1· 100 # epochrec
SE(%) =
(5)
where #epochsleep is the number of total epochs of sleep, #epochwake is the number of total epochs of wake, and #epochrec is the number of total epochs of recording.
PERFORMANCE EVALUATION Performance of a sleep monitoring application depends on the accuracy of SE. We used mean absolute error (MAE) and bias to determine the accuracy. MAE is a quantity used to measure how close forecasts or predictions are to the eventual outcomes. The formula to calculate MAE is described in Eq. 6: Fig. 2. Schematic diagram of the wavelet decomposition process. This diagram schematically shows how the wavelet decomposition is processed. The hn[n] and gn[n] appearing in the diagram represent the detail coefficients and approximation coefficients, respectively.
524 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
MAE =
1 N + SEapp, k - SEref N k=1
(6)
SLEEP QUALITY MEASUREMENT TECHNIQUE
where SEapp is the SE measured by the handheld device application and SEref is the reference SE measured by the Actiwatch. Calculated SE could be biased by its detection method. Bias means an average tendency of estimation; if bias is positive, the monitoring algorithm overestimates the result, and if bias is negative, the result is underestimated. We calculated bias as an average of the difference between the reference value and the result of each application. Equation 7 shows the equation for calculating bias:
Table 2. Demographics for Subjects Who Participated in the Sleep Study GENDER
AGE (YEARS)
HEIGHT (CM)
WEIGHT (KG)
BMI (KG/M2)
1
F
27
163
52
19.6
2
M
29
183
76
22.7
3
M
28
173
60
20.0
4
M
27
172
65
22.0
5
F
26
155
48
19.9
6
M
25
170
66
22.8
Experimental Results
7
M
30
172
84
28.4
EXPERIMENTAL SETUP
8
M
28
186
82
23.7
9
F
29
165
56
20.6
10
M
26
181
77
23.5
11
M
31
170
79
27.3
12
M
30
178
72
22.7
13
M
30
177
57
18.2
14
M
34
164
60
22.3
Bias =
1 N + (SEapp, k - SEref ) N k=1
(7)
To evaluate the proposed method, we designed an experimental protocol and have compared our method with commercialized sleep monitoring applications. We implemented the proposed method and have installed it on the GT-I9300 to test in actual sleep conditions with various commercial applications simultaneously. We tested the most popular commercial applications, Sleep Cycle Alarm Clock and Motion X, with the proposed method and compared them with the result with the Actiwatch. The proposed algorithm is installed on the GT-I9300; however, Sleep Cycle Alarm Clock and Motion X are tested using the Apple (Cupertino, CA) iPhone 4s/iOS6 because these applications only support Apple iOS. The test protocol includes not only changes of mattress but also changes in measurement position. From self-investigation of existing sleep monitoring applications, we found that the major complaints are as follows: dysfunction in latex or Tempur-Pedic (Lexington, KY) mattress usage and discomfort and inaccuracy from the monitoring position. To guarantee more practical and universal usage, we performed the comparative test using spring and latex mattresses, as well as locating sensors beside both the head and the body. Test results were evaluated with the reference actimetry device, the Philips Actiwatch 2. Although its intersite reliability varies with activity type, the Actiwatch 2 has excellent interunit
SUBJECT NUMBER
No subject had a reported disease, including sleep disorder, and any kind of medication was restricted before sleep. BMI, body mass index; F, female; M, male.
Table 1. Experimental Conditions for Comparative Study of Sleep Monitoring Applications on Various Mattress Types and Measuring Locations CONDITION
VARIABLE
Reference
Phillips Actiwatch 2
Comparative sleep monitoring applications
Sleep Cycle Alarm Clock Motion X
Mattress type
Spring Latex
Measuring location
Head Body
Fig. 3. Experimental setting. Multiple devices including the Actiwatch are simultaneously used to record a movement signal of different locations using various sleep monitoring applications.
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 525
SHIN ET AL.
out of 10 times on body-side measurements. The proposed method showed minimum average error (3.50%) compared with commercial sleep monitoring applications. The proposed method also showed the best performance in the bias test. In the bias test, the bias of the proposed method is a mere - 1.91%, whereas those of the other applications are relatively high. The Sleep Cycle Alarm Clock Fig. 4. Performance comparison of calculated sleep efficiency in various applications: comparhas a very large error not only in the SE ison of (a) average error and (b) average bias between each application and the Actiwatch. MX, Motion X; SCAC, Sleep Cycle Alarm Clock. test but also in the bias test. Both absolute value of error and bias are over 30%. Average bias of the Sleep Cycle Alarm Clock is represented as reliability and good criterion validity in measuring SE. Therefore, - 32.43%. Here, the negative value means that the result is underin optimal application of sleep research, the Actiwatch’s results estimated compared with the reference. On the other hands, Motion X are regarded as valuable data. Experimental conditions are sumshows 8.01% bias except in detection failure cases. Motion X shows marized in Table 1. better performance than the Sleep Cycle Alarm Clock, but it is not up Twenty experiments were conducted to record sleep data, and, in to the proposed method’s performance (–1.91%). Figure 4b shows the total, 14 subjects participated. Every experiment was conducted in an comparison of bias test results. ordinary sleep environment without any control; most cases of experiments were conducted at home and in a private bed, but in some Comparative study. For more detailed analysis, we compared recases data were recorded in the laboratory with a common bed. sults in specified conditions: mattress type changes and various Subjects included 3 women and 11 men, with an average age of 28.57 – 2.38 years (range, 25–34 years) and body mass index of 22.41 – 2.84 kg/m2 (range, 18.19–28.39 kg/m2); no subject had any reported diseases, including sleep disorder, and any kind of medication was restricted. Subject information is presented in Table 2. Figure 3 shows an example of the experimental setup by sensor locations. In every experiment, subjects wore the Actiwatch on their left wrist, and the phone was situated on the bed near the subject. We set a default phone position as the vicinity of the head because all of sleep monitoring applications are based on the head-side recording. In some applications, from ithe operating manual, we located the phone under the pillow. For recording the signal of the head and body simultaneously, two or more devices were used with the Actiwatch.
SE Overall performance. The performance of the proposed method was determined by the accuracy of the SE calculation. We tried to design the experiment to be as real as possible because various events could occur in sleep conditions. For example, measurement conditions or the environment could be changed by phone position, phone direction, mattress type, sleep partner, and event off-bed conditions. Figure 4a represents the overall error of calculated SE, including the various sleep conditions, compared with the reference device, the Actiwatch. This result shows that the Sleep Cycle Alarm Clock has a very high error in calculating SE (32.43%). Although Motion X shows better results (14.91%) relatively, it failed 3
526 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
Fig. 5. Performance comparison of calculated sleep efficiency by changing the mattress type: (a) overall difference of measurement error between a spring mattress and a latex mattress, (b) difference of error in the Sleep Cycle Alarm Clock (SCAC), (c) difference of error in Motion X (MX), and (d) differences of error in the the proposed method.
SLEEP QUALITY MEASUREMENT TECHNIQUE
measuring locations. The comparative study consisted of two different focuses: one is the study related to mattress type, and another is the study related to the measurement location. First, in an experiment using various types of mattress, the error of SE was reduced in the latex bed compared with the spring bed except for the Sleep Cycle Alarm Clock. Figure 5 shows the result of comparative study by changing the mattress. Overally, performance of SE estimation was significantly ( p < 0.05) changed according to mattress type. Figure 5a shows that the detection error was decreased by 6.24% on the latex mattress compared with the spring mattress (68.91% of spring mattress error). With the latex mattress, errors of SE are dramatically reduced by 13.64% ( p = 0.098) and 1.25% ( p = 0.090) in Motion X (Fig. 5c) and the proposed method (Fig. 5d), respectively. That would amount to 62.8% and 30.2% of errors in spring mattress experiments. On the other hand, in the case of the Sleep Cycle Alarm Clock (Fig. 5b), errors are slightly decreased by 3.84% ( p = 0.358), and it would amount to 11.2% of the spring mattress error. However, significant differences were not found in each application, although the average detection errors were decreased according to mattress changes. The average differences of head and body measurement are represented as 0.85%, 15.23%, and 2.66% in the Sleep Cycle Alarm Clock, Motion X, and the proposed method, respectively. From these results, we could postulate that the Sleep Cycle Alarm Clock and the proposed method have less deviation to the measuring location
change compared with Motion X. These results are represented in Figure 6. Figure 6a represents the overall error changes according to changing measuring location; here, the overall difference is 3.90%, but it is not significant. Figure 6b and c show the error changes of the Sleep Cycle Alarm Clock and Motion X, respectively. From Figure 6b and c, we find that errors were varied by - 0.85% in the Sleep Cycle Alarm Clock and by 15.23% in Motion X by changing the measuring location. In both cases, there were no significant differences. However, in Figure 6d, a significant difference is found with the proposed method ( p < 0.05); however, it may not be considerable because the deviation is quite a bit smaller than the other application’s result. Detailed information is presented in Table 3. We find that the Sleep Cycle Alarm Clock underestimates SE in every case. On the other hand, Motion X always overestimates the SE with three times the failure rate. The proposed method showed a balanced estimation result relatively.
Discussion In this article, we evaluate the availability of sleep–wake detection using handheld devices. Handheld devices have many limitations in practical use compared with the existing wristband-type commercial sleep monitor. In other words, the handheld device can measure the human motion indirectly because it works on the bed, whereas the wrist-type sleep monitor is always in contact with the human body. There are many considerable factors of indirect measurement in sleep monitoring. For example, partner or pet movement can be measured as artifacts because indirect measurement measures the whole movement generated by anything on the mattress. In contrast, this type of sleep monitor can misdetect the off-bed condition as a deep-sleep condition because there is not even the least movement. Therefore, good care should be taken not to under- or overestimate the SE in using indirect sleep monitoring with a handheld device. Other important issues raised by commercial application users are the inaccuracy on a latex or Tempur-Pedic mattress and the uncomfortableness of measurement position. Considering that the most important advantages of the handheld device monitoring are the unconstrained measurement and usage of a familiar device, these inconveniences should be improved for practical use. From these points, we designed the experiment to validate the usefulness of a handheld device; specifically, we investigated the overall accuracy of SE calculation and effects of measuring location changes and mattress type changes.
ACCURACY OF SE
Fig. 6. Performance comparison of calculated sleep efficiency by changing the measuring position: (a) overall difference of measurement error between the head side and the body side, (b) difference of error in the Sleep Cycle Alarm Clock (SCAC), (c) difference of error in Motion X (MX), and (d) differences of error in the proposed method.
Overall performance of accuracy can be summarized in two noticeable results: one is the quite good performance of the wavelet-based sleep–wake detection method, and the other is the poor performance of existing commercialized sleep–wake detection applications. Every method has the same principle for detection of ‘‘awake.’’ They measure the movement signal from an accelerometer, filter the signal, decide awake with a specific threshold, and post-process the
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 527
SHIN ET AL.
Table 3. Calculated Sleep Efficiency and Error Compared with the Actiwatch in Different Bed Types and Measuring Locations BED TYPE, TEST NUMBER
RECORDED SE (%)
DIFFERENCE VERSUS REFERENCE (%)
POSITION
ACTI
SCAC
MX
PROPOSED
SCAC
MX
PROPOSED
Body
89.73
65
Fail
89
- 24.73
Fail
Head
89.73
66
91
92
- 23.73
Body
92.33
69
Fail
87
- 23.33
Head
92.33
68
99
94
- 24.33
Body
93.49
47
Fail
92
- 46.49
Head
93.49
46
96
97
- 47.49
2.51
3.51
Body
92.2
56
96
95
- 36.2
3.8
2.8
Head
92.2
57
97
92
- 35.2
4.8
- 0.2
Body
88.01
56
96
92
- 32.01
7.99
3.99
Head
88.01
56
91
93
- 32.01
2.99
4.99
Body
87.35
41
97
86
- 46.35
9.65
- 1.35
Head
87.35
35
98
75
- 52.35
10.65
- 12.35
Body
88.61
33
98
85
- 55.61
9.39
- 3.61
Head
88.61
34
94
76
- 54.61
5.39
- 12.61
Body
85.17
69
97
83
- 16.17
11.83
- 2.17
Head
85.17
67
97
77
- 18.17
11.83
- 8.17
Body
85.29
54
99
84
- 31.29
13.71
- 1.29
Head
85.29
56
99
79
- 29.29
13.71
- 6.29
Body
86.84
58
96
87
- 28.84
9.16
0.16
Head
86.84
58
96
79
- 28.84
9.16
- 7.84
Body
89.16
45
92
91
- 44.16
2.84
1.84
Head
89.16
54
90
86
- 35.16
0.84
- 3.16
Body
90.49
54
99
90
- 36.49
8.51
- 0.49
Head
90.49
41
93
92
- 49.49
2.51
1.51
Body
91.12
54
96
91
- 37.12
4.88
- 0.12
Head
91.12
51
95
92
- 40.12
3.88
0.88
Body
91.87
55
97
90
- 36.87
5.13
- 1.87
Head
91.87
70
97
92
- 21.87
5.13
0.13
Body
94.43
72
99
94
- 22.43
4.57
- 0.43
Head
94.43
67
95
95
- 27.43
0.57
0.57
Body
85.00
69
96
82
- 16.00
11.00
- 3.00
Head
85.00
55
99
77
- 30.00
14.00
- 8.00
Body
84.87
56
96
85
- 28.87
11.13
0.13
Head
84.87
57
96
79
- 27.87
11.13
- 5.87
Spring 1
2
3
4
5
6
7
8
9
10
1.27 Fail 6.67 Fail
- 0.73 2.27 - 5.33 1.67 - 1.49
Latex 11
12
13
14
15
16
17
continued "
528 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
SLEEP QUALITY MEASUREMENT TECHNIQUE
Table 3. Calculated Sleep Efficiency and Error Compared with the Actiwatch in Different Bed Types and Measuring Locations continued
BED TYPE, TEST NUMBER 18
19
20
RECORDED SE (%)
DIFFERENCE VERSUS REFERENCE (%)
POSITION
ACTI
SCAC
MX
PROPOSED
SCAC
Body
87.12
46
94
84
- 41.12
MX 6.88
PROPOSED - 3.12
Head
87.12
45
95
80
- 42.12
7.88
- 7.12
Body
76.15
47
97
83
- 29.15
20.85
6.85
Head
76.15
50
97
77
- 26.15
20.85
0.85
Body
86.93
80
96
84
- 6.93
9.07
- 2.93
Head
86.93
76
97
78
- 10.93
10.07
- 8.93
ACTI, Actiwatch; MX, Motion X; SCAC, Sleep Cycle Alarm Clock; SE, sleep efficiency.
result. In spite of these similar processes, each method shows a different performance. The most significant difference of the proposed method is the wavelet denoising. General acceleration or angular velocity signal contains natural variations, and they have a nonstationary characteristic. Because the motion-generated signal has a large amplitude and specified frequency range, most of the movement could be detected by simple bandpass filtering. However, in practical sleep, there are many unexpected movements such as turning over in bed, arm movement, head movement, leg movement, cough, and so on, and sometimes these movements yield just a small acceleration signal, which is hard to distinguish from natural noise. As we mentioned above, wavelet filtering is more powerful in suppressing nonstationary noise. Thus, we postulate that it is useful to remove the noise from weak movement and farfield noise such as leg movement, and these are proved by the experimental result. Without any post-processing, the proposed wavelet denoising–based method reduces the error of SE by 28.93% and 11.41% in the Sleep Cycle Alarm Clock and Motion X, respectively. As a result, the proposed method shows superior performance over other applications.
MATTRESS TYPE In tests with a mattress change, the result was contrary to our expectation. Most of the complaints about a sleep monitoring application involve the malfunction on the latex or Tempur-Pedic bed. However, in our experiment, the overall SE error is significantly less ( p < 0.05) with a latex mattress than with a spring mattress. We suppose that this result is closely related to the mattress cover. Because the latex mattress is generally used with its cover, the smartphone may sense the cover movement induced by human movement during sleep.
MEASURING LOCATION Existing commercial applications provide the measurement guideline for sleep monitoring, and all of them recommend that the phone should be placed at the side of the head. Some applications
even require that the phone should be located under the pillow. Moreover, most users express their discomfort because most of the applications need to change during sleep monitoring, and the device is located close to the pillow. Users are afraid of exposing their head to electromagnetic waves. Therefore, measuring at the unconstrained location is quite important to improve usability. Every monitoring algorithm has differences according to the measurement location. In the case of both mattresses, the difference in measured SE is less than 2% in the Sleep Cycle Alarm Clock and the proposed method; however, Motion X shows a relatively large difference with changes of measuring location. In particular, Motion X shows unstable measurement performance on body-side measurement with three times the failure rate. The above results imply that location dependency is present and that it is different in each application, and some applications have the limitation of measuring the location in spite of the good SE-measuring performance with the designated position.
Conclusions This research is focused on the validation of a novel sleep monitoring algorithm compared with existing commercialized applications. Specifically, in our experiment, we intensively investigated the sleep condition, including mattress changes and phone location changes. Therefore, we calculated overall performance including all of the SE calculation results from the different mattresses (spring/ latex) and different measuring locations. The result of this research could be summarized as follows: (1) the proposed wavelet-based SE calculation method is more accurate than commercial popular sleep monitoring applications; (2) the proposed SE calculation method has a significant difference by changing the measuring location (approximately 2.66%); however, it would not be a problem because absolute error is definitely smaller than with any other comparative applications; and (3) generally, SE measurement is more accurate on a latex mattress than on a spring mattress. These results provide a promising approach to measure sleep quality using general handheld devices instead of a pricey medical-specified device. This kind of
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 529
SHIN ET AL.
approach is very important to accomplish daily healthcare because most people want to check their health condition 24/7 using ordinary devices without any additional add-ons. Our belief is that ordinary consumer electronics such as a smartphone or a television can play an important role as an innovative solution for practical and efficient daily healthcare or wellness.
Disclosure Statement
14. El Baz M, et al. Evaluation of actiwatch actimeter vs polysomnography in 29 patients with obstructive sleep apnea syndrome [abstract 149]. J Sleep Res 1998;7(Suppl 2):75. 15. Jean-Louis G, et al. The actigraph data analysis software: I. A novel approach to scoring and interpreting sleep-wake activity. Percept Motor Skills 1997;85:207–216. 16. Mullaney DJ, Kripke DF, Messin S. Wrist-actigraphic estimation of sleep time. Sleep 1980;3:83–92. 17. Sadeh A, et al. The role of actigraphy in the evaluation of sleep disorders. Sleep 1995;18:288–302.
B.C., D.K., and J.C. are employees of Samsung Electronics Co. Ltd. H.S. declares no competing financial interests exist.
18. King MA, et al. The validation of a new actigraphy system for the measurement of periodic leg movements in sleep. Sleep Med 2005;6:507–513.
REFERENCES
19. Morgenthaler T, et al. Practice parameters for the use of actigraphy in the assessment of sleep and sleep disorders: An update for 2007. Sleep 2007;30:519–529.
1. Colten HR, Altevogt BM, eds. Sleep disorders and sleep deprivation: An unmet public health problem. Washington, DC: National Academies Press, 2006.
20. Gironda RJ, et al. Preliminary evaluation of reliability and criterion validity of Actiwatch-score. J Rehabil Res Dev 2007;44:223–230.
2. Owens JA, Witmans M. Sleep problems. Curr Probl Pediatr Adolesc Health Care 2004;34:154–179.
21. Hyde M, et al. Validation of actigraphy for determining sleep and wake in children with sleep disordered breathing. J Sleep Res 2007;16:213–216.
3. Van Dongen, Hans PA, et al. The cumulative cost of additional wakefulness: Dose-response effects on neurobehavioral functions and sleep physiology from chronic sleep restriction and total sleep deprivation. Sleep 2003;26:117–129.
22. Lichstein KL, et al. Actigraphy validation with insomnia. Sleep 2006;29:232–239.
4. Mangold R, et al. The effects of sleep and lack of sleep on the cerebral circulation and metabolism of normal young men. J Clin Invest 1955;34: 1092–1100. 5. Takahashi Y, Kipnis DM, Daughaday WH. Growth hormone secretion during sleep. J Clin Invest 1968;47:2079–2090. 6. Corkum P, et al. Sleep problems in children with attention-deficit/hyperactivity disorder: Impact of subtype, comorbidity, and stimulant medication. J Am Acad Child Adolesc Psychiatry 1999;38:1285–1293. 7. Kushida CA, et al. Practice parameters for the indications for polysomnography and related procedures: An update for 2005. Sleep 2005;28:499–521. 8. Kushida CA, et al. Comparison of actigraphic, polysomnographic, and subjective assessment of sleep parameters in sleep-disordered patients. Sleep Med 2001;2:389–396. 9. Kripke DF, et al. Wrist actigraphic measures of sleep and rhythms. Electroencephalogr Clin Neurophysiol 1978;44:674–676. 10. Jean-Louis G, et al. Determination of sleep and wakefulness with the actigraph data analysis software (ADAS). Sleep 1996;19:739–743. 11. Jean-Louis G, et al. Sleep detection with an accelerometer actigraph: Comparisons with polysomnography. Physiol Behav 2001;72:21–28. 12. Sadeh A, Sharkey KM, Carskadon MA. Activity-based sleep–wake identification: An empirical test of methodological issues. Sleep 1994;17:201–207. 13. Cole RJ, et al. Automatic sleep/wake identification from wrist activity. Sleep 1992;15:461–469.
530 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
23. King MA, et al. The validation of a new actigraphy system for the measurement of periodic leg movements in sleep. Sleep Med 2005;6:507–513. 24. Cook KG, et al. An exploratory validation of actigraphic measures of insomnia [abstract 605.L]. Sleep 2004;27:A270. 25. So K, et al. Actigraphy correctly predicts sleep behavior in infants who are younger than six months, when compared with polysomnography. Pediatr Res 2005;58:761–765. 26. Wang D, Wong KK, Dungan GC 2nd, et al. The validity of wrist actimetry assessment of sleep with and without sleep apnea. J Clin Sleep Med 2008;4:450–455. 27. Birge´ L, Massart P. Gaussian model selection. J Eur Math Soc 2001;3:203–268.
Address correspondence to: Jaegeol Cho, PhD Digital Media and Communication R&D Center Samsung Electronics Co. Ltd. 129, Samsung-ro, Yeongtong-gu Suwon-si, Gyeonggi-do Republic of Korea E-mail: [email protected] Received: June 24, 2013 Revised: September 26, 2013 Accepted: September 26, 2013
This article has been cited by: 1. Barbara Galland, Kim Meredith-Jones, Philip Terrill, Rachael Taylor. 2014. Challenges and Emerging Technologies within the Field of Pediatric Actigraphy. Frontiers in Psychiatry 5. . [CrossRef] 2. Do Yoon Kim, Hangsik Shin. 2014. Movement Characteristic Analysis for Unconstrained Sleep Efficiency Analysis Based on the Smartphone. The Transactions of The Korean Institute of Electrical Engineers 63, 940-944. [CrossRef] 3. Hangsik Shin, Jaegeol Cho. 2014. Unconstrained snoring detection using a smartphone during ordinary sleep. BioMedical Engineering OnLine 13, 116. [CrossRef]
Hangsik Shin, PhD,1 Byunghun Choi, PhD,2 Doyoon Kim, PhD,2 and Jaegeol Cho, PhD2 1
Department of Biomedical Engineering, Chonnam National University, Yeosu-si, Jeollanam-do, Republic of Korea. 2 Digital Media and Communication R&D Center, Samsung Electronics Co. Ltd., Suwon-si, Gyeonggi-do, Republic of Korea.
Abstract Objective: The purpose of this study was to develop and validate a novel method for sleep quality quantification using personal handheld devices. Materials and Methods: The proposed method used 3- or 6-axes signals, including acceleration and angular velocity, obtained from built-in sensors in a smartphone and applied a realtime wavelet denoising technique to minimize the nonstationary noise. Sleep or wake status was decided on each axis, and the totals were finally summed to calculate sleep efficiency (SE), regarded as sleep quality in general. The sleep experiment was carried out for performance evaluation of the proposed method, and 14 subjects participated. An experimental protocol was designed for comparative analysis. The activity during sleep was recorded not only by the proposed method but also by well-known commercial applications simultaneously; moreover, activity was recorded on different mattresses and locations to verify the reliability in practical use. Every calculated SE was compared with the SE of a clinically certified medical device, the Philips (Amsterdam, The Netherlands) Actiwatch. Results: In these experiments, the proposed method proved its reliability in quantifying sleep quality. Compared with the Actiwatch, accuracy and average bias error of SE calculated by the proposed method were 96.50% and - 1.91%, respectively. Conclusions: The proposed method was vastly superior to other comparative applications with at least 11.41% in average accuracy and at least 6.10% in average bias; average accuracy and average absolute bias error of comparative applications were 76.33% and 17.52%, respectively. Key words: sleep monitoring, sleep efficiency, sleep quality, wavelet, healthcare application
Introduction
G
ood sleep is essential for human health and well-being. Sleep is a restorative activity of the brain; thus, sleep deprivation increases the risk of mental disorders such as depression and anxiety. Moreover, insufficient sleep
522 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
reduces the desire and the motivation for physical activity, contributing to weight gain, obesity, and associated disorders.1–4 Sleep is also related to memory and learning. During slow wave sleep, memory is consolidated, which means that the memory pieces of what we have learned during the day come together coherently so that the information can be accessed when needed. In particular, sleep disruption could delay development of children or infant because growth hormone is secreted during slow wave sleep.5 Furthermore, sleep disruption is associated with a higher incidence of behavioral problems in children such as attention deficit and hyperactivity disorder.6 In spite of the importance of good sleeping, millions of people do not get enough sleep, and many suffer from lack of sleep. For example, surveys (see www.apa.org/topics/sleep/ why.aspx) conducted by the National Sleep Foundation (1999–2004) reveal that at least 40 million Americans suffer from over 70 different sleep disorders, and 60% of adults report having sleep problems a few nights a week or more. Most of those with these problems go undiagnosed and untreated. In addition, more than 40% of adults experience daytime sleepiness severe enough to interfere with their daily activities at least a few days each month, with 20% reporting problem sleepiness a few days a week or more. Furthermore, 69% of children experience one or more sleep problems a few nights or more during a week. To improve sleep habit and quality, sleep monitoring techniques have undergone continuous development. Polysomnography (PSG), a representative sleep monitoring technique, measures multiple parameters during sleep, including the electroencephalogram, electromyogram, electrocardiogram, and respiration, and analyzes sleep in detail. For example, PSG reports parameters such as quality of sleep, muscle tensing, cardiac activity, respiratory disorder, sleep apnea test, oxygen saturation, and leg movement or body position during actual sleep.7 Although PSG is a reference technique in analyzing sleep, several limitations still remain in practical daily use. First of all, PSG requires numerous electrical leads, which disturb sleep, in order to measure physiological signal, and circumstances such as temperature, humidity, sound level, or light intensity must be controlled to perform PSG properly. Most of all, PSG is hard to conduct individually and daily because the PSG equipment is high-priced and complex to set up. For these reasons, PSG has just been conducted by specialized agencies, and consecutive measurement is almost impossible. As an alternative method of PSG, several studies have been carried out to validate the availability of actimetry for assessing sleep and waking. Actimetry is essentially designed for ambulatory activity, recording acceleration signals, and it is also used for sleep
DOI: 10.1089/tmj.2013.0216
SLEEP QUALITY MEASUREMENT TECHNIQUE
monitoring. Actigraphy is less expensive, noninvasive, and more conducive to repeated measures in comparison with PSG.8,9 In the last two decades, actimetry has found good agreement (97%) for total sleep time10 and overall agreement rates of 91–93%11,12 compared with PSG. Moreover, it is strongly correlated in normal sleepers and in sleepers with apnea, with correlation coefficients ranging from 0.89 to 0.98.9,13–17 Additionally, there is much evidence related to the clinical usefulness of actigraphy in many different fields of sleep study.18–26 Actimetry dramatically reduces the complexity of sleep monitoring; it is not only simple to use, but also reasonably priced. However, actimetry still necessarily requires PC-based software for device setup and checking up on the result; moreover, it is hard to use without additional training for operation. Also, consumers may not want to purchase the device because the functions of actimetry are quite limited to actigraph recording. Recently, various types of sleep management applications have been unveiled based on personal smart devices. These applications calculate the sleep efficiency (SE) using an acceleration signal measured by a device-embedded accelerometer and provide the result with an intuitive user interface. Some applications estimate the sleep stage and control the alarm to wake up at the proper time. The number of downloads of sleep management applications is increasing consistently because of the ease of approach by the application store, the low price, and the ease of use with an intuitive user interface. Likewise, the sleep monitoring technique has rapidly developed with personal handheld devices. However, we should emphasize that there are no reports to evaluate the clinical suitability of these methods, and most sleep monitoring applications have strict operation guidelines: the device should be located in a designated location such as a side of the head or should be placed in a specified direction. Functional examination of the sleep monitoring application should be designed to give practical assistance to the user. The purpose of this article can be summarized as follows: the first is to develop an improved method for calculating SE based on a smart device, the second is to evaluate the representative smart device–based sleep monitoring applications in practical situations, and the last is to verify the performance of each application compared with a reference device, the Actiwatch (Philips, Amsterdam, The Netherlands).
Materials and Methods
were validated for significance based on the paired t test. SE was calculated by the following procedure: (1) movement sensing with the 3-axes accelerometer and the 3-axes gyroscope (total 6-axes movement signal), (2) noise reduction with the discrete wavelet transform–based wavelet denoising technique, (3) sleep–wake decision with empirical threshold, and (4) SE calculation. This procedure is described in Figure 1.
MOVEMENT SIGNAL ACQUISITION We used two types of sensors in movement recording: the accelerometer and the gyroscope. These sensors are similar to recording motion, but they have different characteristics and specifications. The accelerometer measures the acceleration, including gravity. Thus the acceleration value keeps a steady constant, and it changes with movement and sensor position. In other words, the value of acceleration could be changed by phone direction and tilting without motion. The gyroscope is a sensor for measuring angular velocity. It has zero value basically and just is a constant when motion is occurring. The output value, of course, could change in sensor direction; it is not important in recognizing whether movement is occurring or not. We have measured both signals to distinguish wake status from sleep. Acceleration signal and angular velocity were recorded with built-in Samsung GT-I9300 sensors, and sampling frequency was 16.6667 Hz.
NOISE REDUCTION BASED ON REAL-TIME DISCRETE WAVELET TRANSFORM Acceleration and angular velocity measured from sensors included small variations from manufacturing characteristics, and such variation could be regarded as a noise in discriminating motion. Generally, this natural noise has no effect in detecting obvious movement because it has low amplitude near the baseline and has random characteristics. However, in some cases, the motion-induced signal is very small, so it is hard to discriminate movement signal from natural noise components. Therefore, a noise reduction algorithm was necessary to improve the signal-to-noise ratio of the measured signal. We use wavelet transform to reduce the occurrence of noise naturally. The wavelet is a useful mathematical function and is usually used in digital signal processing and image compression. Also, the wavelet could be used for spatial–temporal analysis of the signal based on mother function and scale functions. From these characteristics, the wavelet transform is very useful to analyze a nonstationary signal and becomes an alternative of short-term Fourier transform and Gabor transform. We have chosen the discrete Haar wavelet, which is the most general and simple, to discriminate movement signal and noise.
In our experiment, the Samsung (Gyeonggi-do, Republic of Korea) GT-I9300 (Galaxy S3) was used for signal acquisition and analysis. The signal processing algorithm was developed with MathWorks (Natick, MA) MATLAB 2011b and converted to Java programming language for loading to the Android (Google, Mountainview, CA) OS. The analysis application was designed to record and to analyze real-time, and it was installed on the GT-I9300 for the experiment. IBM (Armonk, NY) SPSS Statistics version 20 was used for statistical analysis, and data Fig. 1. Procedure of sleep efficiency calculation. DWT, discrete wavelet transform.
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 523
SHIN ET AL.
Discrete Haar wavelet transform is composed with two filters: lowpass filter and high-pass filter. The outputs are divided into the approximation coefficients of the low-pass filter (gn[n]) and teh detail coefficients of the high-pass filter (hn[n]). Each filter can be described in as Eqs. 1 and 2, respectively: 1 gk [n] = pffiffiffi (gj - 1 (2k) + gj - 1 (2k + 1)) 2
(1)
1 hk [n] = pffiffiffi (gj - 1 (2k) + gj - 1 (2k + 1)) 2
(2)
We decompose an input signal into six levels and denoise with threshold. The penalized threshold was used to set a noise rejection threshold. The penalized threshold is a level-wise threshold method in which the detail coefficients could be sorted in descending order. The threshold k could be computed with Eq. 327: " # t n 2 2 k = ar gmin + d + 2r t a + ln , t = 1, . . . , n (3) t k t k=1 where d is the detailed coefficient, r is the noise standard deviation, n is the signal length, and a > 1 is the sparsity parameter. Figure 2 shows the decomposition procedure with the six-level filter bank. In the case of a sleep monitoring system, the power consumption and short delay are very important issues because the handheld device is always turned on during monitoring; moreover, the resources of handheld devices are severely limited. Therefore, we have implemented a real-time wavelet filter for reducing the computational load. By sharing low- and high-pass filtering in every cycle, the computational load of the reporting cycle, which means every 32 cycles, could be decreased. Related contents are briefly shown in Figure 2.
SLEEP AND WAKE DECISION We have postulated that the sleep and wake status depends on movement during sleep. Thus, we have defined ‘‘sleep’’ as the ‘‘movement signal was not detected from all sensors’’ and ‘‘wake’’ as the ‘‘movement signal was detected from at least one sensor.’’ To determine if a movement has occurred or not, the empirical threshold was applied to the wavelet denoised signal on each axis. The threshold could be adjusted by changing the environment such as bed type and sensor position. In our experiment, we set *10–3 as a threshold value. Sleep and wake detection was performed with the following steps: (1) detect movement with the threshold from each x-, y-, z-axis acceleration signal and each x-, y-, z-axis angular velocity (a total of 6 axes), (2) summing up the detection results of 6 axes, and (3) decision sleep and wake with summing up the result.
SE CALCULATION SE, which is widely used to evaluate the quality of one’s sleep, is the ratio of time spent asleep (actual sleep time) to the amount of time spent in bed, and it could be calculated as the number of minutes of sleep divided by the number of minutes in bed. Generally, SE is approximately 85–90% or higher in normal adults, but the ‘‘laboratory effect’’ allows for a 75% efficiency in the sleep lab environment. Equation 4 describes the formula to calculate SE: Tsleep Twake SE(%) = · 100 (4) · 100 = 1 Trec Trec where Tsleep is the total sleep time, Twake is the total wake time, and Trec is the total recording time. In this study, Eq. 4 can be modified to Eq. 5 because sleep and wake status values were reported in every 1-min epoch: # epochsleep · 100 # epochrec # epochwake = 1· 100 # epochrec
SE(%) =
(5)
where #epochsleep is the number of total epochs of sleep, #epochwake is the number of total epochs of wake, and #epochrec is the number of total epochs of recording.
PERFORMANCE EVALUATION Performance of a sleep monitoring application depends on the accuracy of SE. We used mean absolute error (MAE) and bias to determine the accuracy. MAE is a quantity used to measure how close forecasts or predictions are to the eventual outcomes. The formula to calculate MAE is described in Eq. 6: Fig. 2. Schematic diagram of the wavelet decomposition process. This diagram schematically shows how the wavelet decomposition is processed. The hn[n] and gn[n] appearing in the diagram represent the detail coefficients and approximation coefficients, respectively.
524 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
MAE =
1 N + SEapp, k - SEref N k=1
(6)
SLEEP QUALITY MEASUREMENT TECHNIQUE
where SEapp is the SE measured by the handheld device application and SEref is the reference SE measured by the Actiwatch. Calculated SE could be biased by its detection method. Bias means an average tendency of estimation; if bias is positive, the monitoring algorithm overestimates the result, and if bias is negative, the result is underestimated. We calculated bias as an average of the difference between the reference value and the result of each application. Equation 7 shows the equation for calculating bias:
Table 2. Demographics for Subjects Who Participated in the Sleep Study GENDER
AGE (YEARS)
HEIGHT (CM)
WEIGHT (KG)
BMI (KG/M2)
1
F
27
163
52
19.6
2
M
29
183
76
22.7
3
M
28
173
60
20.0
4
M
27
172
65
22.0
5
F
26
155
48
19.9
6
M
25
170
66
22.8
Experimental Results
7
M
30
172
84
28.4
EXPERIMENTAL SETUP
8
M
28
186
82
23.7
9
F
29
165
56
20.6
10
M
26
181
77
23.5
11
M
31
170
79
27.3
12
M
30
178
72
22.7
13
M
30
177
57
18.2
14
M
34
164
60
22.3
Bias =
1 N + (SEapp, k - SEref ) N k=1
(7)
To evaluate the proposed method, we designed an experimental protocol and have compared our method with commercialized sleep monitoring applications. We implemented the proposed method and have installed it on the GT-I9300 to test in actual sleep conditions with various commercial applications simultaneously. We tested the most popular commercial applications, Sleep Cycle Alarm Clock and Motion X, with the proposed method and compared them with the result with the Actiwatch. The proposed algorithm is installed on the GT-I9300; however, Sleep Cycle Alarm Clock and Motion X are tested using the Apple (Cupertino, CA) iPhone 4s/iOS6 because these applications only support Apple iOS. The test protocol includes not only changes of mattress but also changes in measurement position. From self-investigation of existing sleep monitoring applications, we found that the major complaints are as follows: dysfunction in latex or Tempur-Pedic (Lexington, KY) mattress usage and discomfort and inaccuracy from the monitoring position. To guarantee more practical and universal usage, we performed the comparative test using spring and latex mattresses, as well as locating sensors beside both the head and the body. Test results were evaluated with the reference actimetry device, the Philips Actiwatch 2. Although its intersite reliability varies with activity type, the Actiwatch 2 has excellent interunit
SUBJECT NUMBER
No subject had a reported disease, including sleep disorder, and any kind of medication was restricted before sleep. BMI, body mass index; F, female; M, male.
Table 1. Experimental Conditions for Comparative Study of Sleep Monitoring Applications on Various Mattress Types and Measuring Locations CONDITION
VARIABLE
Reference
Phillips Actiwatch 2
Comparative sleep monitoring applications
Sleep Cycle Alarm Clock Motion X
Mattress type
Spring Latex
Measuring location
Head Body
Fig. 3. Experimental setting. Multiple devices including the Actiwatch are simultaneously used to record a movement signal of different locations using various sleep monitoring applications.
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 525
SHIN ET AL.
out of 10 times on body-side measurements. The proposed method showed minimum average error (3.50%) compared with commercial sleep monitoring applications. The proposed method also showed the best performance in the bias test. In the bias test, the bias of the proposed method is a mere - 1.91%, whereas those of the other applications are relatively high. The Sleep Cycle Alarm Clock Fig. 4. Performance comparison of calculated sleep efficiency in various applications: comparhas a very large error not only in the SE ison of (a) average error and (b) average bias between each application and the Actiwatch. MX, Motion X; SCAC, Sleep Cycle Alarm Clock. test but also in the bias test. Both absolute value of error and bias are over 30%. Average bias of the Sleep Cycle Alarm Clock is represented as reliability and good criterion validity in measuring SE. Therefore, - 32.43%. Here, the negative value means that the result is underin optimal application of sleep research, the Actiwatch’s results estimated compared with the reference. On the other hands, Motion X are regarded as valuable data. Experimental conditions are sumshows 8.01% bias except in detection failure cases. Motion X shows marized in Table 1. better performance than the Sleep Cycle Alarm Clock, but it is not up Twenty experiments were conducted to record sleep data, and, in to the proposed method’s performance (–1.91%). Figure 4b shows the total, 14 subjects participated. Every experiment was conducted in an comparison of bias test results. ordinary sleep environment without any control; most cases of experiments were conducted at home and in a private bed, but in some Comparative study. For more detailed analysis, we compared recases data were recorded in the laboratory with a common bed. sults in specified conditions: mattress type changes and various Subjects included 3 women and 11 men, with an average age of 28.57 – 2.38 years (range, 25–34 years) and body mass index of 22.41 – 2.84 kg/m2 (range, 18.19–28.39 kg/m2); no subject had any reported diseases, including sleep disorder, and any kind of medication was restricted. Subject information is presented in Table 2. Figure 3 shows an example of the experimental setup by sensor locations. In every experiment, subjects wore the Actiwatch on their left wrist, and the phone was situated on the bed near the subject. We set a default phone position as the vicinity of the head because all of sleep monitoring applications are based on the head-side recording. In some applications, from ithe operating manual, we located the phone under the pillow. For recording the signal of the head and body simultaneously, two or more devices were used with the Actiwatch.
SE Overall performance. The performance of the proposed method was determined by the accuracy of the SE calculation. We tried to design the experiment to be as real as possible because various events could occur in sleep conditions. For example, measurement conditions or the environment could be changed by phone position, phone direction, mattress type, sleep partner, and event off-bed conditions. Figure 4a represents the overall error of calculated SE, including the various sleep conditions, compared with the reference device, the Actiwatch. This result shows that the Sleep Cycle Alarm Clock has a very high error in calculating SE (32.43%). Although Motion X shows better results (14.91%) relatively, it failed 3
526 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
Fig. 5. Performance comparison of calculated sleep efficiency by changing the mattress type: (a) overall difference of measurement error between a spring mattress and a latex mattress, (b) difference of error in the Sleep Cycle Alarm Clock (SCAC), (c) difference of error in Motion X (MX), and (d) differences of error in the the proposed method.
SLEEP QUALITY MEASUREMENT TECHNIQUE
measuring locations. The comparative study consisted of two different focuses: one is the study related to mattress type, and another is the study related to the measurement location. First, in an experiment using various types of mattress, the error of SE was reduced in the latex bed compared with the spring bed except for the Sleep Cycle Alarm Clock. Figure 5 shows the result of comparative study by changing the mattress. Overally, performance of SE estimation was significantly ( p < 0.05) changed according to mattress type. Figure 5a shows that the detection error was decreased by 6.24% on the latex mattress compared with the spring mattress (68.91% of spring mattress error). With the latex mattress, errors of SE are dramatically reduced by 13.64% ( p = 0.098) and 1.25% ( p = 0.090) in Motion X (Fig. 5c) and the proposed method (Fig. 5d), respectively. That would amount to 62.8% and 30.2% of errors in spring mattress experiments. On the other hand, in the case of the Sleep Cycle Alarm Clock (Fig. 5b), errors are slightly decreased by 3.84% ( p = 0.358), and it would amount to 11.2% of the spring mattress error. However, significant differences were not found in each application, although the average detection errors were decreased according to mattress changes. The average differences of head and body measurement are represented as 0.85%, 15.23%, and 2.66% in the Sleep Cycle Alarm Clock, Motion X, and the proposed method, respectively. From these results, we could postulate that the Sleep Cycle Alarm Clock and the proposed method have less deviation to the measuring location
change compared with Motion X. These results are represented in Figure 6. Figure 6a represents the overall error changes according to changing measuring location; here, the overall difference is 3.90%, but it is not significant. Figure 6b and c show the error changes of the Sleep Cycle Alarm Clock and Motion X, respectively. From Figure 6b and c, we find that errors were varied by - 0.85% in the Sleep Cycle Alarm Clock and by 15.23% in Motion X by changing the measuring location. In both cases, there were no significant differences. However, in Figure 6d, a significant difference is found with the proposed method ( p < 0.05); however, it may not be considerable because the deviation is quite a bit smaller than the other application’s result. Detailed information is presented in Table 3. We find that the Sleep Cycle Alarm Clock underestimates SE in every case. On the other hand, Motion X always overestimates the SE with three times the failure rate. The proposed method showed a balanced estimation result relatively.
Discussion In this article, we evaluate the availability of sleep–wake detection using handheld devices. Handheld devices have many limitations in practical use compared with the existing wristband-type commercial sleep monitor. In other words, the handheld device can measure the human motion indirectly because it works on the bed, whereas the wrist-type sleep monitor is always in contact with the human body. There are many considerable factors of indirect measurement in sleep monitoring. For example, partner or pet movement can be measured as artifacts because indirect measurement measures the whole movement generated by anything on the mattress. In contrast, this type of sleep monitor can misdetect the off-bed condition as a deep-sleep condition because there is not even the least movement. Therefore, good care should be taken not to under- or overestimate the SE in using indirect sleep monitoring with a handheld device. Other important issues raised by commercial application users are the inaccuracy on a latex or Tempur-Pedic mattress and the uncomfortableness of measurement position. Considering that the most important advantages of the handheld device monitoring are the unconstrained measurement and usage of a familiar device, these inconveniences should be improved for practical use. From these points, we designed the experiment to validate the usefulness of a handheld device; specifically, we investigated the overall accuracy of SE calculation and effects of measuring location changes and mattress type changes.
ACCURACY OF SE
Fig. 6. Performance comparison of calculated sleep efficiency by changing the measuring position: (a) overall difference of measurement error between the head side and the body side, (b) difference of error in the Sleep Cycle Alarm Clock (SCAC), (c) difference of error in Motion X (MX), and (d) differences of error in the proposed method.
Overall performance of accuracy can be summarized in two noticeable results: one is the quite good performance of the wavelet-based sleep–wake detection method, and the other is the poor performance of existing commercialized sleep–wake detection applications. Every method has the same principle for detection of ‘‘awake.’’ They measure the movement signal from an accelerometer, filter the signal, decide awake with a specific threshold, and post-process the
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 527
SHIN ET AL.
Table 3. Calculated Sleep Efficiency and Error Compared with the Actiwatch in Different Bed Types and Measuring Locations BED TYPE, TEST NUMBER
RECORDED SE (%)
DIFFERENCE VERSUS REFERENCE (%)
POSITION
ACTI
SCAC
MX
PROPOSED
SCAC
MX
PROPOSED
Body
89.73
65
Fail
89
- 24.73
Fail
Head
89.73
66
91
92
- 23.73
Body
92.33
69
Fail
87
- 23.33
Head
92.33
68
99
94
- 24.33
Body
93.49
47
Fail
92
- 46.49
Head
93.49
46
96
97
- 47.49
2.51
3.51
Body
92.2
56
96
95
- 36.2
3.8
2.8
Head
92.2
57
97
92
- 35.2
4.8
- 0.2
Body
88.01
56
96
92
- 32.01
7.99
3.99
Head
88.01
56
91
93
- 32.01
2.99
4.99
Body
87.35
41
97
86
- 46.35
9.65
- 1.35
Head
87.35
35
98
75
- 52.35
10.65
- 12.35
Body
88.61
33
98
85
- 55.61
9.39
- 3.61
Head
88.61
34
94
76
- 54.61
5.39
- 12.61
Body
85.17
69
97
83
- 16.17
11.83
- 2.17
Head
85.17
67
97
77
- 18.17
11.83
- 8.17
Body
85.29
54
99
84
- 31.29
13.71
- 1.29
Head
85.29
56
99
79
- 29.29
13.71
- 6.29
Body
86.84
58
96
87
- 28.84
9.16
0.16
Head
86.84
58
96
79
- 28.84
9.16
- 7.84
Body
89.16
45
92
91
- 44.16
2.84
1.84
Head
89.16
54
90
86
- 35.16
0.84
- 3.16
Body
90.49
54
99
90
- 36.49
8.51
- 0.49
Head
90.49
41
93
92
- 49.49
2.51
1.51
Body
91.12
54
96
91
- 37.12
4.88
- 0.12
Head
91.12
51
95
92
- 40.12
3.88
0.88
Body
91.87
55
97
90
- 36.87
5.13
- 1.87
Head
91.87
70
97
92
- 21.87
5.13
0.13
Body
94.43
72
99
94
- 22.43
4.57
- 0.43
Head
94.43
67
95
95
- 27.43
0.57
0.57
Body
85.00
69
96
82
- 16.00
11.00
- 3.00
Head
85.00
55
99
77
- 30.00
14.00
- 8.00
Body
84.87
56
96
85
- 28.87
11.13
0.13
Head
84.87
57
96
79
- 27.87
11.13
- 5.87
Spring 1
2
3
4
5
6
7
8
9
10
1.27 Fail 6.67 Fail
- 0.73 2.27 - 5.33 1.67 - 1.49
Latex 11
12
13
14
15
16
17
continued "
528 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
SLEEP QUALITY MEASUREMENT TECHNIQUE
Table 3. Calculated Sleep Efficiency and Error Compared with the Actiwatch in Different Bed Types and Measuring Locations continued
BED TYPE, TEST NUMBER 18
19
20
RECORDED SE (%)
DIFFERENCE VERSUS REFERENCE (%)
POSITION
ACTI
SCAC
MX
PROPOSED
SCAC
Body
87.12
46
94
84
- 41.12
MX 6.88
PROPOSED - 3.12
Head
87.12
45
95
80
- 42.12
7.88
- 7.12
Body
76.15
47
97
83
- 29.15
20.85
6.85
Head
76.15
50
97
77
- 26.15
20.85
0.85
Body
86.93
80
96
84
- 6.93
9.07
- 2.93
Head
86.93
76
97
78
- 10.93
10.07
- 8.93
ACTI, Actiwatch; MX, Motion X; SCAC, Sleep Cycle Alarm Clock; SE, sleep efficiency.
result. In spite of these similar processes, each method shows a different performance. The most significant difference of the proposed method is the wavelet denoising. General acceleration or angular velocity signal contains natural variations, and they have a nonstationary characteristic. Because the motion-generated signal has a large amplitude and specified frequency range, most of the movement could be detected by simple bandpass filtering. However, in practical sleep, there are many unexpected movements such as turning over in bed, arm movement, head movement, leg movement, cough, and so on, and sometimes these movements yield just a small acceleration signal, which is hard to distinguish from natural noise. As we mentioned above, wavelet filtering is more powerful in suppressing nonstationary noise. Thus, we postulate that it is useful to remove the noise from weak movement and farfield noise such as leg movement, and these are proved by the experimental result. Without any post-processing, the proposed wavelet denoising–based method reduces the error of SE by 28.93% and 11.41% in the Sleep Cycle Alarm Clock and Motion X, respectively. As a result, the proposed method shows superior performance over other applications.
MATTRESS TYPE In tests with a mattress change, the result was contrary to our expectation. Most of the complaints about a sleep monitoring application involve the malfunction on the latex or Tempur-Pedic bed. However, in our experiment, the overall SE error is significantly less ( p < 0.05) with a latex mattress than with a spring mattress. We suppose that this result is closely related to the mattress cover. Because the latex mattress is generally used with its cover, the smartphone may sense the cover movement induced by human movement during sleep.
MEASURING LOCATION Existing commercial applications provide the measurement guideline for sleep monitoring, and all of them recommend that the phone should be placed at the side of the head. Some applications
even require that the phone should be located under the pillow. Moreover, most users express their discomfort because most of the applications need to change during sleep monitoring, and the device is located close to the pillow. Users are afraid of exposing their head to electromagnetic waves. Therefore, measuring at the unconstrained location is quite important to improve usability. Every monitoring algorithm has differences according to the measurement location. In the case of both mattresses, the difference in measured SE is less than 2% in the Sleep Cycle Alarm Clock and the proposed method; however, Motion X shows a relatively large difference with changes of measuring location. In particular, Motion X shows unstable measurement performance on body-side measurement with three times the failure rate. The above results imply that location dependency is present and that it is different in each application, and some applications have the limitation of measuring the location in spite of the good SE-measuring performance with the designated position.
Conclusions This research is focused on the validation of a novel sleep monitoring algorithm compared with existing commercialized applications. Specifically, in our experiment, we intensively investigated the sleep condition, including mattress changes and phone location changes. Therefore, we calculated overall performance including all of the SE calculation results from the different mattresses (spring/ latex) and different measuring locations. The result of this research could be summarized as follows: (1) the proposed wavelet-based SE calculation method is more accurate than commercial popular sleep monitoring applications; (2) the proposed SE calculation method has a significant difference by changing the measuring location (approximately 2.66%); however, it would not be a problem because absolute error is definitely smaller than with any other comparative applications; and (3) generally, SE measurement is more accurate on a latex mattress than on a spring mattress. These results provide a promising approach to measure sleep quality using general handheld devices instead of a pricey medical-specified device. This kind of
ª M A R Y A N N L I E B E R T , I N C . VOL. 20
NO. 6 JUNE 2014
TELEMEDICINE and e-HEALTH 529
SHIN ET AL.
approach is very important to accomplish daily healthcare because most people want to check their health condition 24/7 using ordinary devices without any additional add-ons. Our belief is that ordinary consumer electronics such as a smartphone or a television can play an important role as an innovative solution for practical and efficient daily healthcare or wellness.
Disclosure Statement
14. El Baz M, et al. Evaluation of actiwatch actimeter vs polysomnography in 29 patients with obstructive sleep apnea syndrome [abstract 149]. J Sleep Res 1998;7(Suppl 2):75. 15. Jean-Louis G, et al. The actigraph data analysis software: I. A novel approach to scoring and interpreting sleep-wake activity. Percept Motor Skills 1997;85:207–216. 16. Mullaney DJ, Kripke DF, Messin S. Wrist-actigraphic estimation of sleep time. Sleep 1980;3:83–92. 17. Sadeh A, et al. The role of actigraphy in the evaluation of sleep disorders. Sleep 1995;18:288–302.
B.C., D.K., and J.C. are employees of Samsung Electronics Co. Ltd. H.S. declares no competing financial interests exist.
18. King MA, et al. The validation of a new actigraphy system for the measurement of periodic leg movements in sleep. Sleep Med 2005;6:507–513.
REFERENCES
19. Morgenthaler T, et al. Practice parameters for the use of actigraphy in the assessment of sleep and sleep disorders: An update for 2007. Sleep 2007;30:519–529.
1. Colten HR, Altevogt BM, eds. Sleep disorders and sleep deprivation: An unmet public health problem. Washington, DC: National Academies Press, 2006.
20. Gironda RJ, et al. Preliminary evaluation of reliability and criterion validity of Actiwatch-score. J Rehabil Res Dev 2007;44:223–230.
2. Owens JA, Witmans M. Sleep problems. Curr Probl Pediatr Adolesc Health Care 2004;34:154–179.
21. Hyde M, et al. Validation of actigraphy for determining sleep and wake in children with sleep disordered breathing. J Sleep Res 2007;16:213–216.
3. Van Dongen, Hans PA, et al. The cumulative cost of additional wakefulness: Dose-response effects on neurobehavioral functions and sleep physiology from chronic sleep restriction and total sleep deprivation. Sleep 2003;26:117–129.
22. Lichstein KL, et al. Actigraphy validation with insomnia. Sleep 2006;29:232–239.
4. Mangold R, et al. The effects of sleep and lack of sleep on the cerebral circulation and metabolism of normal young men. J Clin Invest 1955;34: 1092–1100. 5. Takahashi Y, Kipnis DM, Daughaday WH. Growth hormone secretion during sleep. J Clin Invest 1968;47:2079–2090. 6. Corkum P, et al. Sleep problems in children with attention-deficit/hyperactivity disorder: Impact of subtype, comorbidity, and stimulant medication. J Am Acad Child Adolesc Psychiatry 1999;38:1285–1293. 7. Kushida CA, et al. Practice parameters for the indications for polysomnography and related procedures: An update for 2005. Sleep 2005;28:499–521. 8. Kushida CA, et al. Comparison of actigraphic, polysomnographic, and subjective assessment of sleep parameters in sleep-disordered patients. Sleep Med 2001;2:389–396. 9. Kripke DF, et al. Wrist actigraphic measures of sleep and rhythms. Electroencephalogr Clin Neurophysiol 1978;44:674–676. 10. Jean-Louis G, et al. Determination of sleep and wakefulness with the actigraph data analysis software (ADAS). Sleep 1996;19:739–743. 11. Jean-Louis G, et al. Sleep detection with an accelerometer actigraph: Comparisons with polysomnography. Physiol Behav 2001;72:21–28. 12. Sadeh A, Sharkey KM, Carskadon MA. Activity-based sleep–wake identification: An empirical test of methodological issues. Sleep 1994;17:201–207. 13. Cole RJ, et al. Automatic sleep/wake identification from wrist activity. Sleep 1992;15:461–469.
530 TELEMEDICINE and e-HEALTH J U N E 2 0 1 4
23. King MA, et al. The validation of a new actigraphy system for the measurement of periodic leg movements in sleep. Sleep Med 2005;6:507–513. 24. Cook KG, et al. An exploratory validation of actigraphic measures of insomnia [abstract 605.L]. Sleep 2004;27:A270. 25. So K, et al. Actigraphy correctly predicts sleep behavior in infants who are younger than six months, when compared with polysomnography. Pediatr Res 2005;58:761–765. 26. Wang D, Wong KK, Dungan GC 2nd, et al. The validity of wrist actimetry assessment of sleep with and without sleep apnea. J Clin Sleep Med 2008;4:450–455. 27. Birge´ L, Massart P. Gaussian model selection. J Eur Math Soc 2001;3:203–268.
Address correspondence to: Jaegeol Cho, PhD Digital Media and Communication R&D Center Samsung Electronics Co. Ltd. 129, Samsung-ro, Yeongtong-gu Suwon-si, Gyeonggi-do Republic of Korea E-mail: [email protected] Received: June 24, 2013 Revised: September 26, 2013 Accepted: September 26, 2013
This article has been cited by: 1. Barbara Galland, Kim Meredith-Jones, Philip Terrill, Rachael Taylor. 2014. Challenges and Emerging Technologies within the Field of Pediatric Actigraphy. Frontiers in Psychiatry 5. . [CrossRef] 2. Do Yoon Kim, Hangsik Shin. 2014. Movement Characteristic Analysis for Unconstrained Sleep Efficiency Analysis Based on the Smartphone. The Transactions of The Korean Institute of Electrical Engineers 63, 940-944. [CrossRef] 3. Hangsik Shin, Jaegeol Cho. 2014. Unconstrained snoring detection using a smartphone during ordinary sleep. BioMedical Engineering OnLine 13, 116. [CrossRef]
