2017 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI 2017)

November 16 - 18, 2017, Daegu, Korea

Explainable sleep quality evaluation model using machine learning

approach

Rock-Hyun Choi, Won-Seok Kang, Chang-Sik Son*

of sleeping medication, and daytime dysfunction [4]. It shows
Abstract— This research presents a scheme for explainable sleep good reliability and validity of results for patients who have
quality evaluation utilizing the heart rate based sleep index. In the psychiatric and sleep disorders, leading to its use in a significant
proposed model, the global covering rule induction of LERS body of research [5]. Psychometric evaluation supports the
(Learning from Examples based on Rough Sets) is used to generate consistency, reliability, and validity of results, but the scoring is
rules associated with sleep quality status, such as ‘Bad,’ ‘Normal,’ difficult to handle [6]. Recently, shortening of the PSQI survey
and ‘Good.’ These rules are used to interpret the three sleep has been conducted in order to reduce inaccuracies caused by
statuses. To show the applicability of the proposed scheme, we the burden of answering the lengthy survey [7].
construct a sleep quality evaluation model based on sleep intraday
time-series data collected from 280 factory and office workers with In addition, there are many objective measurement solutions.
Fitbit fitness trackers. An evaluation of the proposed model was Polysomnography (PSG) is the gold standard in sleep
provided through statistical cross validation experiments. evaluation. A PSG device includes a recorder which integrates
measures of cortical electroencephalographic activity, muscle
tone, and eye movements to characterize the structure of sleep
I. INTRODUCTION [8]. ‘The meaning of good sleep’ research team conducted a
Understanding sleep states is very important for human long term study of PSG and subjective sleep quality, and
health, as unbalanced sleep presents an immediate and pressing concluded that objective measures of sleep continuity were
threat [1]. Poor quality sleep and frequently altered sleeping closely reflected in perceived sleep quality and that sleep quality
patterns may cause mood imbalances [2]. For workers, research essentially means sleep continuity [9].
shows that lower quality sleep is linked to a higher risk of
Wearable wristband Smart phone Fitbit cloud server
accidents such as those involving motor vehicles. Sleep problems
have also been shown to have an impact on physical and mental
well-being. As an increasing number of workers have started
working in shifts, the amount of research into stressful workplace
experiences and sleep problems is increasing. Moreover, poor
sleep may also increase the likelihood of hypertension, diabetes,
muscle pain, headaches, and gastrointestinal problems. In the
case of mental health, sleep disorders are one of the main risk
factors of illnesses such as depression [3]. Therefore, an
appropriate solution for sleep quality measurement is important
for a healthy life.
Main server Data model analyzer Result reporting
Since 1989, the Pittsburgh Sleep Quality Index (PSQI) has
been widely used as a method to test sleep quality in clinical
populations. The PSQI is a subjective measure and consists of Figure 1. Data analytic system block diagram
nineteen individual items such as sleep quality, sleep latency,
sleep duration, habitual sleep efficiency, sleep disturbances, use However, the PSG test for continuous measurement requires
considerable resources. First, the device itself is expensive.
Second, the equipment requires well-trained technicians, and its
*
Corresponding author. usage is time consuming [8]. Thus, PSG is not employed for
This work was supported by the R&D program of the Ministry of Trade,
large-scale, long-term (i.e. weeks/months) sleep monitoring
Industry and Energy of Korea (Industrial Technology Innovation Program,
10063553, Program of Advanced Technology Development for Future Industry, research because of the associated complexity and cost [10].
10044323, Technology Innovation Program 10077260), and the DGIST R&D
Wearable devices are rapidly developing at present, and
Program of the Ministry of Education, Science and Technology of Korea
(17-IT-02). All collected data was used after receiving permission of device performance is improving. Thus, it is possible to collect
Institutional Review Board (IRB) approval (no. DGIST-160303-HR-012-01) personal health-related data every day, and wearable devices are
from DGIST cheaper and more comfortable. Bei et al. [11] researched
Rock-Hyun Choi, Won-Seok Kang, and Chang-Sik Son(co-authors) are with objective measurement of sleep using a wrist actigraph, and
the Convergence Research Center for Wellness, DGIST, 333 Techno obtained high correlations compared to PSG. The study tested
Jungang-daero, Hyeonpung-myeon, Dalseong-gun, Daegu, 42988, Republic of
Korea (e-mail: choimosi@dgist.ac.kr, wskang@dgist.ac.kr,
Fitbit Charge HR performance and found that it showed good
changsikson@dgist.ac.kr). performance compared to PSG and ECG during sleep [8]. The

978-1-5090-6064-1/17/$31.00 ©2017 IEEE 542

average HR was measured at 59.3±7.5 bpm (obtained via Fitbit intraday data has no sleep mode values, the corresponding data
Charge HR) and 60.2±7.6 bpm (obtained via ECG, p < 0.001) points are removed. After this process, the heart rate range is
throughout the night. The results are not significantly different classified [15].
and the difference can be negligible. For this reason, our
The proposed model provides a resting heart rate during
research selects the Fitbit Charge HR as its apparatus.
sleep; this rate is between 27 and 42 percent of the maximum
The machine learning approach is important as a data analytic heart rate [15]. The heart rate in a non-exercising state is less
strategy. Learning methods are a promising approach to various than 50 percent of the maximum heart rate, and the maximum
problems related to monitoring health conditions and diagnosing rate is about 220 minus user’s age [16]. In addition, the resting
pathologies. Bernardeschi et al. [12] used a supervised learning heart rate during sleep is usually about 8 percent lower than
approach in order to identify a reliable sleep quality data set from while awake [17].
objective physiological parameters and human subjective
The sleep ratio is calculated as the ratio of minutes of normal
judgment. Kaplan et al. [13] used machine learning analysis to
heart rate sleep to total sleep. A ‘good’ sleep ratio is more than
reduce bias and error.
90 percent of sleep with a resting heart rate. The ‘normal’ sleep
Preliminary researches have focused primarily on reducing ratio is between 0.6 and 0.9. A ‘bad’ sleep ratio is less than 0.6.
errors and determining sleep status, but it is also important to
interpret sleep quality status from a human point of view. For C. Sleep quality index based explainable model
qualitative interpretation and explanation of sleep quality status, As shown in Fig. 3, the five procedures of data processing
we use the global covering rule induction algorithm of LERS are the loading of data, discretization, rule induction,
(Learning from Examples based on Rough Sets). The rest of the classification, and performance evaluation. The proposed
paper is organized as follows: in section II, we present an explainable method can induce rules to estimate sleep quality
explainable sleep quality evaluation model. Section III shows status as ‘Bad,’ ‘Normal,’ or ‘Good.’
experimental results, i.e., rules associated with sleep quality
status such as ‘Bad,’ ‘Normal,’ and ‘Good,’ classification
performance. Finally, the conclusion is provided in section IV.

II. MATERIALS AND METHODS

A. Data acquisition
This research used the Fitbit Charge HR fitness tracker, a
wristband-type wearable device. The device consists of a 3-axis
accelerometer, an altimeter, and optical heart rate sensors. The
data provided includes steps, distance, calories, floors, heart rate Figure 3. Structure of explainable sleep quality evaluation model
(beats per minute) and sleep modes: asleep, awake, and really
awake, for every minute [14]. First, the ‘data load’ procedure generates a datum or data set
based on date information, e.g., YYYY/MM/DD, defined by
In the preliminary study, intraday time series data is stored in users. To construct an explainable sleep quality model, we
MongoDB, and Java 1.8 is used in order to collect Fitbit device collected data for one month. The attribute consists of 19
data from the Fitbit RESTful service, as shown in Figure 1. features, and the class label includes three sleep statuses. The
study population consists of 280 factory and office workers and
B. Heart Rate Based Sleep Quality Index covers a period from April 1st to April 30th, 2016. The total
number of collected intraday data sets is 2829. There are 256
female data sets and 2573 male, and the data of males is used
because of the data imbalance. Table 1 lists the data
characteristics used in the proposed model.
Second, we use a discretization technique, which converts
numerical (or quantitative) attributes into discrete (or
qualitative) ones, in order to generate candidate rules associated
with the three sleep statuses. The attribute domains are divided
into consecutive subintervals with a number of candidate
cut-points. There are many discretization methods which can be
applied before rule induction [18].
Figure 2. The estimation of sleep quality status [15] In this study, we used the equal size based discretization
(ESD) algorithm. This method divides the data points into p
Fig. 2 shows a flow chart of the heart rate-based sleep sub-groups which each sub-group contains an approximately
quality index. Initial data is transferred from the Fitbit cloud same number of attribute values. For the both methods, the best
server to our main server. The Fitbit charge HR denotes asleep way of determining p is by looking at the histogram and try
as “1,” awake as “2,” and really awake as “3.” If the sleep different intervals or sub-groups [19].

543
 During the third procedure, a global covering rule induction for each attribute a in A do
algorithm is used to search the set of all attribute values. This Q := P – {a}
algorithm, as a component of the data mining system LERS
(Learning from Examples using Rough Sets), is based on some Compute partition Q*
rough set concepts, i.e., lower and upper approximations [20]. If Q* ⊆ {d}* then P := Q
The global covering algorithm checks if the input data is
end for
consistent, i.e., if the data does not contain conflicting examples.
If the data is inconsistent, it computes lower and upper R := P
approximations of all concepts, e.g., ‘Good,’ ’Normal,’ and end if
‘Bad.’ The following shows the process to compute a single
global covering. end
During the fourth procedure, the classifier evaluates new
decisions of test instances from a rule set which was induced
Input: the set A of all attributes, partition {d}*, on U. from training instances. Given a rule set containing rules for
Output: a single global covering, R. each class, we use the best k rules for prediction of each test
instance, with the following procedure: (1) select all the rules
begin
whose conditions (or bodies) partially match the test instance;
Compute partition A* (2) select the best k rules from the candidate rules selected in
P := A step (1); and (3) choose the class of rule with the highest
matching degree of rule conditions as the predicted class.
R := Ø
We used multiple rules in prediction because the accuracy of
if A* ⊆ {d}* rules cannot be precisely estimated, and one cannot expect that
any single rule can perfectly predict the class label of every

TABLE I. CHARACTERISTIC OF COLLECTED DATA

Variables Description(Unit) Frequency( %) or Mean±SD

507(20’s)±1.53
1127(30’s)±2.66
Age Age (yrs)
772(40’s)±2.41
167(50’s)±1.41
BMI Body mass index (kg/m2) 25.51±0.30
Yes: 2252(90%)
Drinking Drinking or not (Yes or No)
No: 252(10%)
Yes: 805(32%)
Smoking Smoking or not (Yes or No)
No: 1699(68%)
Low_Avg_HR Average value of low heart rate during sleep 48.04±2.00

Normal_Avg_HR Average value of normal heart rate during sleep 62.10±5.63

High_Avg_HR Average value of high heart rate during sleep 82.55±8.13

Low_Sleep_Minutes Sleep time with low heart rate during sleep (minutes) 20.22±65.04

Normal_Sleep_Minutes Sleep time with normal heart rate during sleep (minutes) 333.95±146.01

High_Sleep_Minutes Sleep time with high heart rate during sleep (minutes) 37.34±81.10

Low_Asleep Asleep time with low heart rate during sleep (minutes) 19.21±61.62

Low_Awake Awake time with low heart rate during sleep (minutes) 0.95±3.92

Low_Really_Awake Really wake time with low heart rate during sleep (minutes) 0.07±0.41

Normal_Asleep Asleep time with normal heart rate during sleep (minutes) 309.48±136.67

Normal_Awake Awake time with normal heart rate during sleep (minutes) 22.11±15.58

Normal_Really_Awake Really wake time with normal heart rate during sleep (minutes) 2.36±2.93

High_Asleep Asleep time with high heart rate during sleep (minutes) 31.78±73.09

High_Awake Awake time with high heart rate during sleep (minutes) 4.26±9.12

High_Really_Awake Really awake time with high heart rate during sleep (minutes) 1.29±6.22

544
example satisfying its conditions [21]. TABLE III. CLASSIFICATION PERFORMANCE RESULTS (2-FOLD CROSS
VALIDATION).
During the last procedure, we perform a cross-validation
experiment for assessing how the results of the statistical Accuracy (%) 73.0
analysis will generalize to an independent data set. In general, it 1-fold No. of rules generated 263.0
is mainly used in settings where the goal is prediction, and one The highest frequency in the rule set 10
wants to estimate how accurately a prediction model will
perform in practice. Accuracy (%) 72.0
2-fold No. of rules generated 268
The highest frequency in the rule set 10
III. EXPERIMENT RESULTS
Average number of rules 265.5
All variables, as listed in Table I, are considered as input Average
Average accuracy (±SD*) 72.5±0.5
attributes to construct an explainable sleep quality evaluation
*SD, standard deviation.
model. For discretization, the size of sub-intervals was defined as
50 percent of the total number of instances. In the case of rule
induction, missing attribute values, labeled ‘Lost’, are ignored, as I) Sleep quality status ‘Bad’
in the previous study of [22]. Rule 18) IF 17.85 < BMI < 25.21 AND Smoking is Yes AND
61.81 < Normal_Avg_HR < 79.0 AND 0 < Normal_Awake <
We used 2-fold cross validation to provide an unbiased 19.0 AND 1.5 < Normal_Really_Awake < 24.0 AND 1.5 <
estimate of the generalization error. The full dataset was High_Asleep < 606.5 AND 1.5 < High_Awake < 155.0 AND
randomly divided to into 2 subsets: one subset was for training
0.5 < High_Really_Awake < 175.5 THEN Sleep Quality Status
(50%), and the other subset was used for testing (50%). The
is ‘Bad’ with support 8
process was then repeated twice.
Rule 25) IF 17.85 < BMI < 25.21 AND Smoking is No AND
The performance of the explainable sleep quality model was 0.0 < Normal_Asleep < 324.5 AND 0 < Normal_Awake < 19.0
evaluated using the accuracy evaluation criterion. A confusion AND 1.5 < Normal_Really_Awake < 24.0 AND 1.5 <
matrix contains the difference between actual and predicted High_Asleep < 606.5 AND 1.5 < High_Awake < 155.0 AND
outcomes assigned by the classification system [23]. 0.0 < High_Really_Awake < 0.5 THEN Sleep Quality Status is
Table II shows the confusion matrix for a binary ‘Bad’ with support 8
classification problem. The accuracy can be defined by using the Rule 9) IF 25.21 < BMI < 43.42 AND Smoking is Yes AND
elements of the confusion matrix: 83.61 < High_Avg_HR < 192.16 AND 0 < Normal_Awake <
19.0 AND 1.0 < Normal_Really_Awake < 0.5 AND 1.5 <
High_Awake < 155.0 AND 0.5 < High_Really_Awake < 175.5
Accuracy = (TP+TN) / (TP+FN+TN+FP) * 100 (2) THEN Sleep Quality Status is ‘Bad’ with support 7

TABLE II. CONFUSION MATRIX.

II) Sleep quality status ‘Normal’
Rule 56) IF 23.0 < Age < 36.5 AND Drinking is Yes AND
Actual output 61.81 < Normal_Avg_HR < 79.0 AND 70.73 < High_Avg_HR
Positive Negative < 83.61 AND 324.5 < Normal_Asleep < 840.5 AND 19.0 <
Predicted output Normal_Awake < 109.0 AND 0.0 < Normal_Really_Awake <
Positive True positive (TP) False Positive (FP) 1.5 AND 1.5 < High_Awake < 155.0 AND 0.0 <
Negative False Negative (FN) True Negative (TN) High_Really_Awake < 0.5 THEN Sleep Quality Status is
‘Normal’ with support 10.
Rule 11) IF 23.0 < Age < 36.5 AND 17.85 < BMI < 25.21
where TP and TN are the correctly classified positives and
AND Drinking is Yes AND Smoking is No AND 83.61 <
negatives, while FP and FN are the incorrectly classified
High_Avg_HR < 192.16 AND 0.0 < Normal_Asleep < 324.5
positive and negatives.
AND 1.5 < High_Asleep < 606.5 AND 1.5 < High_Awake <
155.0 AND 0.0 < High_Really_Awake < 0.5 THEN Sleep
All the algorithms used in this study were implemented in
Quality Status is ‘Normal’ with support 9.
the Python 3.6 environment using Numpy 1.11.3 and Pandas
Rule 73) IF 23.0 < Age < 36.5 AND 25.21 < BMI < 43.43
0.19.1. Table III shows the classification performance and the
AND Drinking is Yes AND Smoking is Yes AND 70.73 <
number of generated rules in each fold during the 2-fold cross
High_Avg_HR < 83.61 AND 324.5 < Normal_Asleep < 840.5
validation experiments. The average classification
AND 0.0 < Normal_Really_Awake < 1.5 AND 1.5 <
performance was 72.5% and the average number of rules,
High_Awake < 155.0 AND 0.5 < High_Really_Awake < 175.5
generated from each training datum, was 265.5. Among the
THEN Sleep Quality Status is ‘Normal’ with support 9.
generated rules associated with the three sleep quality statuses,
‘Bad,’ ‘Normal,’ and ‘Good,’ the top three rules with the
III) Sleep quality status ‘Good’
highest frequency are as follows:
Rule 75) IF 36.5 < Age < 55.0 AND 17.85 < BMI < 25.21

545
AND 0.0 < Normal_Asleep < 324.5 AND 0 < Normal_Awake in primary insomnia,” J. Psychosom. Res., vol. 53, no. 3, pp. 737–740,
Sep. 2002.
< 19.0 AND 1.5 < Normal_Really_Awake < 24.0 AND 1.5 <
[6] S. L. Beck, A. L. Schwartz, G. Towsley, W. Dudley, and A. Barsevick,
High_Asleep < 606.5 AND 1.5 < High_Awake < 155.0 AND “Psychometric evaluation of the Pittsburgh Sleep Quality Index in cancer
0.0 < High_Really_Awake < 0.5 THEN Sleep Quality Status is patients,” J. Pain Symptom Manage., vol. 27, no. 2, pp. 140–148, Feb.
‘Good’ with support 8. 2004.
[7] M. D. Olfert, I. Holaskova, M. L. Barr, Z. Wenjun, J. Morrell, and S. E.
Rule 14) IF 23.0 < Age < 36.5 AND 25.21 < BMI < 43.43
Colby, “Shortening Pittsburgh Sleep Quality Index Survey Using Factor
AND Smoking is Yes AND 324.5 < Normal_Asleep < 840.5 Analysis,” J. Nutr. Educ. Behav., vol. 48, no. 7, pp. S144–S145, Jul.
AND 0 < Normal_Really_Awake < 1.5 AND 1.5 < 2016.
High_Awake < 155.0 AND 0.0 < High_Really_Awake < 0.5 [8] M. de Zambotti et al., “Measures of sleep and cardiac functioning during
THEN Sleep Quality Status is ‘Good’ with support 7. sleep using a multi-sensory commercially-available wristband in
adolescents,” Physiol. Behav., vol. 158, pp. 143–149, May 2016.
Rule 19) IF Smoking is No AND 49.42 < Normal_Avg_HR < [9] T. Åkerstedt, K. Hume, D. Minors, and J. Waterhouse, “The meaning of
61.81 AND 0 < Normal_Awake < 0.5 AND 324.5 < good sleep: a longitudinal study of polysomnography and subjective
Normal_Asleep < 840.5 AND 19.0 < Normal_Awake < 109.0 sleep quality,” J. Sleep Res., vol. 3, no. 3, pp. 152–158, Sep. 1994.
AND 0 < Normal_Really_Awake < 1.5 AND 0.0 < [10] Z. Chen et al., “Unobtrusive sleep monitoring using smartphones,” in
2013 7th International Conference on Pervasive Computing
High_Really_Awake < 0.5 THEN Sleep Quality Status is Technologies for Healthcare and Workshops, 2013, pp. 145–152.
‘Good’ with support 7. [11] B. Bei, J. Milgrom, J. Ericksen, and J. Trinder, “Subjective Perception of
Sleep, but not its Objective Quality, is Associated with Immediate
The above rules are the most frequent or dominant rules used Postpartum Mood Disturbances in Healthy Women,” Sleep, vol. 33, no.
4, pp. 531–538, Apr. 2010.
to interpret the three sleep statuses. Although it may not be a [12] C. Bernardeschi, M. G. C. A. Cimino, A. Domenici, and G. Vaglini,
general logical representation, the differences in results between “Using Smartwatch Sensors to Support the Acquisition of Sleep Quality
‘Bad’ vs. ‘Normal’ and ‘Bad’ vs ‘Good’ can explain the Data for Supervised Machine Learning,” in Wireless Mobile
following consequences: when comparing the ‘Bad’ 18-th rule Communication and Healthcare, 2016, pp. 251–259.
[13] K. A. Kaplan, P. P. Hardas, S. Redline, and J. M. Zeitzer, “Correlates of
and the ‘Normal’ 56-th rule, smoking is a part of the ‘Bad’ status sleep quality in midlife and beyond: a machine learning analysis,” Sleep
and a high heart rate is frequently maintained when awake Med., vol. 34, pp. 162–167, Jun. 2017.
during sleep. A comparison of the ‘Normal’ and the ‘Good’ [14] “Fitbit.” [Online]. Available: https://dev.fitbit.com/kr. [Accessed:
statuses shows that the ‘high_really_awake’ state, which is in a 10-Aug-2017].
[15] Dongjoo Kim, Chang-Sik Son, Won-Seok Kang, “Development of Sleep
‘Normal’ status, has a longer time with a high heart rate Quality Index Using Heart Rate,” Int. Conf. Bioeng. Biomed. Eng. 2016,
compared to the ‘Good’ status. vol. 10, no. 7, Jul. 2016.
[16] “Target Heart Rates.” [Online]. Available:
IV. CONCLUSION http://www.heart.org/HEARTORG/HealthyLiving/PhysicalActivity/Fitn
%20essBasics/Target-Heart-Rates_UCM_434341_Article.jsp#.WYv2W
The sleep intraday time-series data is measured through the lEjEuU. [Accessed: 10-Aug-2017].
'Fitbit charge HR', which is not as accurate as the PSG, but still [17] K. Kräuchi and A. Wirz-Justice, “Circadian clues to sleep onset
mechanisms,” Neuropsychopharmacol. Off. Publ. Am. Coll.
provides a very good measurement result. Heart rate during the
Neuropsychopharmacol., vol. 25, no. 5 Suppl, pp. S92-96, Nov. 2001.
sleep duration time was measured, and in addition, the [18] J. W. Grzymala-Busse and J. Stefanowski, “Three discretization methods
demographic characteristics and historical information of users for rule induction,” Int. J. Intell. Syst., vol. 16, no. 1, pp. 29–38, Jan.
were used by the global covering rule induction algorithm to 2001.
generate rules related to sleep quality status. From the [19] Y. Yang and G. I. Webb, “Discretization for naive-Bayes learning:
managing discretization bias and variance,” Mach. Learn., vol. 74, no. 1,
experimental results, we found it important to interpret the pp. 39–74, Jan. 2009.
difference between the three sleep quality statuses, and [20] J. W. Grzymala-Busse, “Rule Induction,” in Data Mining and
confirmed that the proposed model can more easily understand Knowledge Discovery Handbook, Springer, Boston, MA, 2005, pp.
the sleep quality status by using objective measures rather than 277–294.
[21] X. Yin and J. Han, “CPAR: Classification based on Predictive
the subjective measurement of questionnaires. Association Rules,” in Proceedings of the 2003 SIAM International
Conference on Data Mining, Society for Industrial and Applied
REFERENCES Mathematics, 2003, pp. 331–335.
[22] J. W. Grzymala-Busse and W. J. Grzymala-Busse, “Handling Missing
[1] O. J. Walch, A. Cochran, and D. B. Forger, “A global quantification of Attribute Values,” in Data Mining and Knowledge Discovery Handbook,
‘normal’ sleep schedules using smartphone data,” Sci. Adv., vol. 2, no. 5, Springer, Boston, MA, 2009, pp. 33–51.
pp. e1501705–e1501705, May 2016. [23] C.-S. Son, Y.-N. Kim, H.-S. Kim, H.-S. Park, and M.-S. Kim,
[2] C. Asher, “Scientists identify molecular link between sleep and mood,” “Decision-making model for early diagnosis of congestive heart failure
Science, Feb. 2016. using rough set and decision tree approaches,” J. Biomed. Inform., vol.
[3] H. K. Knudsen, L. J. Ducharme, and P. M. Roman, “Job stress and poor 45, no. 5, pp. 999–1008, Oct. 2012.
sleep quality: Data from an American sample of full-time workers,” Soc.
Sci. Med., vol. 64, no. 10, pp. 1997–2007, May 2007.
[4] D. J. Buysse, C. F. Reynolds, T. H. Monk, S. R. Berman, and D. J.
Kupfer, “The Pittsburgh sleep quality index: A new instrument for
psychiatric practice and research,” Psychiatry Res., vol. 28, no. 2, pp.
193–213, May 1989.
[5] J. Backhaus, K. Junghanns, A. Broocks, D. Riemann, and F. Hohagen,
“Test–retest reliability and validity of the Pittsburgh Sleep Quality Index

546