Available via license: CC BY 4.0
Content may be subject to copyright.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier xxxxxxxxx.xx
A smart bed for non-obtrusive sleep
analysis in real world context
MARCO LAURINO1, LUCIA ARCARISI2, NICOLA CARBONARO2,3, ANGELO GEMIGNANI4,1,
DANILO MENICUCCI4AND ALESSANDRO TOGNETTI2,3 (Member, IEEE)
1Institute of Clinical Physiology, National Research Council, Pisa, Italy
2Department of Information Engineering, University of Pisa, Pisa, Italy
3Research Centre "E. Piaggio", University of Pisa, Pisa, Italy
4Department of Surgical, Medical and Molecular Pathology and Critical Care Medicine, University of Pisa, Pisa, Italy
Corresponding author: Marco Laurino (e-mail: laurino@ifc.cnr.it)
This research was funded by Regione Toscana through the L.A.I.D. project (POR FESR 2014-2020 - DD3383389/2014 - call 1)
ABSTRACT Sleep disorders are common health problems in industrialized societies and may be caused
by underlying health issues. Current methods to assess the quality of sleep are invasive and not suitable
for continuous monitoring in real world contexts. We have developed a smart sensing solution for non
invasive sleep monitoring specifically conceived for the early identification of pre-clinical sleep disorders
and insomnia in the general population. Our prototype, named the Smart-Bed, is a low-cost solution
that gathers and processes data on the movement and position of the subject, physiological signals, and
environmental parameters. Our tests on the prototype in controlled lab conditions highlighted that the
mattress can reliably detect subject’s position/motion, heart rate and breathing activity. It performs well
compared to polysomnography and correctly classifies four behavioural conditions (no bed occupancy,
wakefulness, non-REM sleep, and REM sleep), which are the basis for creating an objective sleep quality
index.
INDEX TERMS Accelerometers, piezoresistive devices, physiology, signal analysis, psychology, sleep
monitoring, real world data
I. INTRODUCTION
Sleep disorders including insomnia are among the most
common health problems in industrialized societies [1]. Poor
sleep quality increases the probability of incidents and ac-
cidents at work or during daily activities [2]. Furthermore,
insomnia correlates with high rates of absenteeism from work
[2]. Compared to people with good sleep quality, insomniacs
visit clinical structures more frequently and use drugs much
more. A large amount of data [3]–[7] indicates that chronic
insomnia increases vulnerability to mental disorders (de-
pression, anxiety, alcoholism), metabolic diseases (diabetes
and dyslipidemia), and cardiovascular diseases (myocardial
infarction and hypertension), as well as neuro-degenerative
disorders (i.e. mild cognitive impairment).
Systematic, preventive, personalized and non-invasive
methods for sleep quality assessment are thus of paramount
importance. Polysomnography is currently the gold-standard
for assessing sleep quality [8], [9], since it estimates the
macrostructure of sleep, i.e. the division of sleep into subse-
quent stereotypical stages [10]. This macrostructure carries
valuable objective information on the quality of sleep, and
in fact is corrupted in sleep disorders [11]. Unfortunately,
polysomnography is highly invasive. The lack of comfort
prevents the subject from sleeping naturally and it does not
provide fully reliable measurements of the quality of sleep in
real life.
To improve on this situation, within the research project
L.A.I.D. (Linking Automation to artificial Intelligence for
revealing sleep Dysfunctions) [12], we developed a smart
sensing solution for non-invasive sleep quality assessment.
In the project, we developed a smart mattress (hereinafter
called Smart-Bed) specifically conceived to assess the quality
of sleep to early identify pre-clinical signs of sleep disorders
and insomnia in the general population. Our Smart-Bed
collects and processes data on the motion and position of
the subject, physiological signals (heart rate and breathing
rate) and environmental parameters (sound intensity, relative
humidity, room temperature and luminosity). The Smart-
Bed is based on a single mattress made by Materassificio
Montalese [13] (the group leader of the L.A.I.D. project) in
VOLUME X, 20xx 1
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
which we integrated a pressure mapping system and a set of
tri-axial accelerometers.
Several non-invasive sensor technologies have been devel-
oped for in-bed monitoring of biomedical parameters, such
as sleeping posture/movements and physiological signals [9],
[14]–[24]. However, no previous work has focused on the
assessment of sleep quality by comparing it with the standard
(i.e. polysomnography).
The existing solutions for non invasive sleep monitoring
are based on pressure mapping systems that extract an image
of the pressure of the subject lying on the mattress. Con-
tact pressures are measured by multiple pressure sensors,
generally with high spatial resolution, using piezoresistive
[15], capacitive [18]–[20], optical [21] or piezoelectric [23]
technologies. The pressure maps are used to detect the sub-
ject’s presence, sleeping posture or movements or to identify
breathing or cardiac activity. Note that the main current
solutions are monomodal and are not able to simultaneously
detect the position/movement, breathing and heart rate. The
WhizPAD pressure mapping system developed by Yu-Wei
Liu et al. [15] detects user movements and breathing activity.
Chang et al. [19], [20] employed a capacitive matrix for
movement and breathing monitoring. The solution by Ko-
rtelainen et al. [23] records the ballistocardiographic signal
and breathing activity during sleep. All the other solutions
detect position and movement only. In addition, most current
solutions are for hospital locations [15], [18], [21], [25]–[27].
None of the previous works reported the classification of the
sleep macrostructure.
In this work, we developed a multimodal sensing system
by combining textile-based pressure mapping and tri-axial
accelerometers to detect position/movement, breathing ac-
tivity and heart rate in a non-obtrusive way. As previously
noted, the main current solutions are monomodal and, to the
best of our knowledge, our Smart-Bed is the only prototype
that can simultaneously detect position/movement, breathing
and heart rate. In addition, our Smart-Bed is specifically
designed as a consumer product for the general population.
In fact, we have designed a low cost solution that is still
able to detect key clinical parameters. More importantly, our
machine learning analysis identifies four behavioural condi-
tions, and differentiates between non-REM and REM phases.
We have thus demonstrated the possibility of identifying the
main sleep macrostructures. To the best of our knowledge,
our mattress is the first of its kind.
We developed the pressure mapping system by using
piezoresistive textile technology inspired by [28], with modi-
fications to reduce the cross-talk between pressure sensors.
We tested the prototype in controlled lab settings on sev-
eral groups of subjects. The results demonstrated that the
Smart-Bed gives a reliable detection of the subject’s po-
sition and movements as well as heart rate and breathing
activity. In addition, the preliminary assessment performed
very promisingly in comparison to polysomnography. In fact,
the mattress signals enabled us to classify four behavioural
conditions that represent the sleep macrostructure.
The results obtained are encouraging and highlight the
technical validity of the Smart-Bed. However, an extensive
validation phase on a high number of heterogeneous subjects
is needed. In fact, we are currently testing a high number
of subjects in order to validate the Smart-Bed as a tool for
reducing socio-economic costs due to sleep disorders and for
increasing individual well-being.
II. MATERIALS AND METHODS
A. OVERALL ARCHITECTURE
The architecture of the Smart-Bed prototype comprises the
following functional blocks (see Figure 1):
•Docking station (DS)
•Physiological data collector (PDC)
•Environmental data collector (EDC)
Both the PDC and EDC are wired to DS via USB serial
communication interface. The DS is a Microsoft Windows
10 based system with a touchscreen interface. The DS
is equipped with tailored software for: i) managing PDC
and EDC, ii) processing the signals and parameters from
PDC/EDC, iii) storing the collected data, and iv) extracting
the sleep and environmental quality indices. The PDC is a
FIGURE 1. Hardware of functional blocks of the Smart-Bed prototype.
Docking station (DS) is in the red frame, Physiological data collector (PDC) in
the yellow frame and Environmental data collector (EDC) in the green frame.
custom-designed acquisition unit with two different kinds
of sensors: pressure mapping system and three tri-axial ac-
celerometers. The architecture of the sensing components,
the PDC and the EDC were designed by IFC-CNR and Uni-
versity of Pisa. The mattress is made of memory foam and is
produced by Materassificio Montalese S.P.A. (Pistoia, Italy).
The PDC was developed by EB Neuro S.P.A. (Firenze, Italy),
and the EDC and DS by BP Engineering S.P.A. (Carrara,
Italy).
B. PRESSURE MAPPING SYSTEM
The PDC is equipped with a pressure mapping system based
on a piezoresistive textile applied onto the foam layer below
the mattress top cover. We designed the pressure mapping
system to detect the distribution of pressures when a subject
is lying on the bed. The pressure signals obtained can be used
to detect the subject’s position and movements and to extract
the subject’s breathing rate. On the basis of an analysis of the
2VOLUME X, 20xx
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
literature and prioritizing low complexity, low cost and good
tolerance to external disturbances, we developed a resistive
sensor matrix configuration, similar to the one reported in
[28]. In the design phase, as a compromise between reducing
the overall cost and complexity, we used a relative low num-
ber of sensing areas yet still maintained an acceptable quality
of signals. Considering a single mattress measuring 190x90
cm, we built a pressure sensing textile based on a resistive
matrix of 15x13 uniformly-spaced sensing areas that cover
a surface of 125x75 cm (head and feet are not considered).
Note that most solutions in the literature have a much higher
number of sensors (typically by a factor greater than 10).
Figure 2 shows our solution: the central layer is a pressure
sensing piezoresitive fabric, while the additional two layers
are fabrics with integrated row and column conductors. Row
and column conductors are perpendicular. Two analog mul-
FIGURE 2. Pressure mapping system based on a textile resistive matrix with
13 column conductors and 15 row conductors for a total of 195 sensing areas.
A single sensing area is highlighted in the inset.
tiplexers are used to scan rows (row mux) and columns (col
mux) in order to select all the sensing areas of the resistive
matrix. The row mux sequentially connects each row conduc-
tor to Vcc (3.3V) through a pull-up resistor R1 (2 KΩ). When
a row is selected (i.e. powered), the col mux sequentially
connects each column conductor to a voltage divider stage
(pull down resistor R2, 10KΩ). Each crossing between a row
and a column thus represents a sensing area whose electrical
resistance decreases as the applied pressure increases. For
the pressure-sensing layer, we used the piezoresistive fabric
CARBOTEX 03-82 manufactured by SEFAR AG (Heiden,
Switzerland). The top and bottom layers are made of a
PET fabric (from SEFAR AG) with integrated evenly-spaced
metallic stripes. In our design, the metallic stripes have a 2
cm width and are separated by 3 cm in the top layer (rows)
and 8 cm in the bottom (columns) layer. As described in
[29], this sensing architecture has parasitic resistivity in the
transversal directions due to the surface conductivity of the
pressure-sensing layer. To reduce the cross-talk due to the
parasitic resistivity, we built our prototype by cutting the
piezoresistive layer into strips parallel to the row direction
(around 3.5 cm width). The strips were then sewn onto the
top layer centered on the row conductors. Figure 3 shows
the pressure-sensing matrix prototype. The pressure mapping
FIGURE 3. Prototype of the pressure mapping system.
system provides a 800x600 image in which each pixel value
is related to the pressure applied on the specific point of the
mattress. The pressure image is constructed by a linear spatial
interpolation of the raw output values obtained by the 195
sensing areas to obtain the 800x600 image.
C. ACCELEROMETERS
The accelerometers of the PDC were used to extract the
ballistocardiograph (BCG) signal [30] for the unobtrusive
recording of cardiac activity (i.e. heart rate) of the subject
lying on the mattress. For the extraction of the BCG signals,
we selected the micropower digital accelerometer ADXL362
(Analog Device Inc, MA, US). The main specifications of
the ADXL362 are reported in Table 1. The Smart-Bed si-
multaneously collects the signals of the three accelerometers
(a1, a2 and a3) placed in different positions over the pressure
mapping system. As shown in Figure 4, an accelerometer (a1)
was placed in a central area, the others (a2 and a3) in lateral
and contro-lateral sites. We used multiple accelerometers for
two main reasons. Firstly, averaging the signals from dif-
ferent accelerometers reduces the noise and robustly detects
artifacts. Secondly, the multi-site configuration ensures that
at least one of the accelerometers lies below the subject, even
if the subject is not in a central position on the mattress.
D. ENVIRONMENTAL DATA COLLECTOR AND
ENVIRONMENT QUALITY INDEX
The EDC module is based on a Seeeduino V4.2 board, and is
equipped with four sensors that collect the following environ-
mental signals: i) sound intensity, ii) temperature, iii) relative
humidity, and iv) luminosity. All the signals are collected
with a sampling frequency of 1 Hz, except for the sound
intensity which is sampled at 20 Hz (due to the fast dynamics
of environmental noise and snoring events). The collection
TABLE 1. PDC-equipped accelerometer specifications
Specification Value
Number of axes 3
Range ±2g
Supply voltage 3.3 V
Sensitivity 1mg/LSB
Raw data noise level 175µg/√Hz (ultralow noise mode)
Bit resolution 12 bits
Sampling frequency 128 Hz
Shape and dimension Circular, diameter 2.2 cm
VOLUME X, 20xx 3
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 4. Position and topological configuration of the three accelerometers
(a1, a2 and a3) over the pressure mapping layer.
of all the ECD signals and parameters is synchronized with
respect to the data from the PDC.
We used the collected signals to estimate a binary in-
dex regarding the best environmental conditions for sleep
(environment quality index, EQI), based on data from the
literature. For each environmental parameter, we therefore
assigned optimal sleep value ranges in order to obtain four
environmental criteria:
•sound intensity: continuous noise level lower than 35
dB, no more than 45 dB for single noisy events [31],
[32];
•temperature: minimum temperature of 17◦Cand maxi-
mum temperature of 28◦C[33].
•relative humidity: between 40% and 60% [33];
•luminosity: less than 10 lux [34];
The EQI is evaluated for each night’s sleep and it is set to
1 if all four environmental criteria are satisfied, otherwise to
0 (not optimal environmental conditions).
In addition to the EQI extraction, we hypothesized that
information on sounds such as snoring, environmental noise
due to the subject’s activity, and voices could contain relevant
information on the subject’s state during sleep. We thus
used the recorded sound information as one of the inputs to
classify the behavioural conditions that will be described in
Section II-E. For each 30-second epoch, we therefore calcu-
lated the mean values and variances of the sound intensity
(SIe,S Iv).
E. SLEEP QUALITY ALGORITHM
The analysis of the PDC data acquired by the Smart-Bed
consists of three main post-processing steps: 1) analysis of
the PDC signals and extraction of physiological (heart rate
and breathing rate) and activity (movement and sleeping
posture) data; 2) automatic classification of the subject’s
behavioural conditions based on the physiological and activ-
ity data extracted including the sound information extracted
from the EDC as described in Section II-D; 3) exploitation
of the classified behavioural conditions to estimate standard
sleep evaluation parameters, which are then condensed into a
global Sleep Quality Index (SQI).
All the algorithms and analysis were performed in Matlab
(R2018b, Natick, Massachusetts: The MathWorks Inc.).
1) Physiological and activity data
To estimate the breathing rate, we employed a frequency
spectrum-based approach. Firstly, we derived the signal av-
eraged over the sensing areas of the pressure matrix (below
the lying subject). Then, we evaluated the average signal
spectrum (Welch periodogram) to detect the maximum peak
of the spectrum in the frequency range 0.1 Hz to 0.35 Hz. In
our hypothesis, the maximum peak is likely to correspond to
respiratory activity.
We estimated the heart rate from the squared modulus of
the grand average raw signal of the three accelerometers. To
remove possible components due to respiratory activity or
movements, the grand average signal was band-passed in the
frequency range of 0.3–20 Hz. We then extracted the heart
rate using a method based on an autocorrelation function
similar to the ones described in [30], [35], [36].
To evaluate the activity data consisting in the position and
motion of the subject on the mattress, we reduced the sensing
area density from 15x13 to 3x3 by topological averaging.
The position feature vector obtained (9 elements) is assigned
by an artificial neural network (ANN) [37] according to
six putative classes: i) not on bed, ii) supine position, iii)
lying on the left side, iv) lying on the right side, v) prone
position and vi) movement. We used a two-layer ANN, the
size of hidden layer was set to 10. For the training process,
we applied a backward propagation algorithm with scaled
conjugate gradient method [38].
Each estimated data sequence (heart rate, breathing rate,
position and movements) is temporally divided into 30-
second epochs in accordance with the clinical standard in
polysomnographic evaluation. For each 30-second epoch,
the mean values and variances of heart rate (HRe,HRv),
breathing rate (BRe,BRv), movements (MVe,M Vv), and
position (P Se,P Sv) were calculated.
2) Behavioural conditions classification
We classified the subjects’ behavioural conditions in 30-
second epochs using the input parameters described in Sec-
tion II-E1 and II-D: HRe,HRv,BRe,BRv,M Ve,M Vv,
P Se,P Sv,SIeand S Iv. We trained a decision tree algorithm
with bootstrap aggregation [39] to assign to each 30-second
epoch one of the following classes: no bed occupancy, wake-
fulness, non-REM sleep and REM sleep. We co-recorded
the Smart-Bed signals and standard polysomnography with
a clinical polysomnographic system in order to estimate
the real behavioural conditions and sleep staging follow-
ing the clinical criteria. The following signals were col-
lected using the standard polysomnographic recordings: elec-
troencephalography, electrocardiography, respiratory airflow,
snoring, electromyography, and oxygen saturation. Based on
polysomnographic data, each sleep recording is staged in 30-
second epochs according to standard clinical criteria [10],
then the sleep staging is used as a reference to train the
decision tree algorithm.
4VOLUME X, 20xx
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
3) Global Sleep Quality Index
The final step of the sleep quality algorithm was based on
the sleep macro-structure estimated previously. Firstly, the
sleep macro-structure was used to extrapolate the main sleep
time-domain parameters [9], [40], [41] related to each night’s
sleep, such as: sleep efficiency, sleep latency, REM latency,
total sleep time, and wake after sleep onset (WASO). The
SQI was then estimated based on the following partial criteria
(pc):
•pc1= 1 if sleep efficiency > 85%, otherwise it is set to 0
•pc2= 1 if sleep latency < 15 min, otherwise it is set to 0
•pc3= 1 if 60 min < REM latency < 120 min, otherwise
it is set to 0
•pc4= 1 if total sleep time > 7 hours, otherwise it is set
to 0
•pc5= 1 if 70 % < ratio between non-REM sleep and
total sleep time < 95%, otherwise it is set to 0
•pc6= 1 if WASO < 45 min, otherwise it is set to 0
Finally, the SQI is assessed for each night and defined as:
SQI =X
i
pci
SQI ranges from 0 (bad sleep) to 6 (good sleep), with an
ordinal scale of seven discrete levels.
F. EXPERIMENTAL PROCEDURES
To evaluate the estimation of position/movement, breathing
activity and heart rate, we tested the Smart-Bed prototypes
on a group of 15 volunteers (29 to 72 years old, mean=48,4;
8 female/7 male) during wakefulness (signal-testing group).
In these wakefulness tests, during experimental sessions the
subjects were asked to voluntarily modulate their respiratory
activity and change their body position, thus enabling us to
collect different respiratory and postural patterns.
The ANN for the classification of the activity data (Sec-
tion II-E1) was trained with a dataset of 30-second epochs
obtained from six subjects (29 to 59 years old, mean=47,3;
3 female/3 male) during wakefulness (ANN-training group,
with different subjects from the signal-testing group).
The decision tree used to classify the behavioural con-
ditions (Section II-E2) was trained with a dataset acquired
from an additional group (condition-training group) of eight
subjects (26 to 71 years old, mean=54,3;1 female/7 male)
sleeping on the Smart-Bed prototype whilst being recorded
by standard polysomnography (BE LTM, EB Neuro S.p.A.,
Florence Italy). We used the 75% of the epochs for training
and 25% for testing, epochs of training and testing subsets
are not overlapped.
To test the durability and robustness of the Smart-Bed
and to evaluate the SQI and EQI as a function of time, we
collected the Smart-Bed signals from a group (long-testing
group with different subjects from the signal-testing, ANN-
training and condition-training groups) of five subjects (27
to 45 years old, mean=35,2;2 female/3 male) over multiple
continuous days (from 12 to 231 days).
III. RESULTS
To date, we have developed seven Smart-Bed prototypes to
test the modules’ (DS, PDC and EDC) functionality, sensing
solutions and algorithms. Our estimated cost for our Smart-
Bed prototype is around 1000 euros of which approximately
150 euros is for the pressure-sensing matrix . This cost is low
for a research prototype and could be lowered even further
with a series production (e.g. to the best of our knowledge,
existing solutions like XSENSOR [25], SensorEdge [26] and
BodyTrak [27] are in the range of 5-10k euros).
Figure 5 shows four examples of sleeping postures de-
tected by the pressure mapping system when a subject is
lying on the Smart-Bed (prone, supine, left side, and right
side ). The chest, upper arms and legs are easily recognizable.
In static conditions, the pressure image detects the presence
of the subject on the mattress and his/her sleeping posture.
In dynamic conditions, when the subject moves while lying
on the mattress, the pressure image changes continuously
and this variation can be used to determine the subject’s
movement during sleep.
Figure 6 shows an example of breathing activity estimated
by the Smart-Bed and compared with a ground truth obtained
using a thermistor inserted in a nasal cannula. The reference
signal is based on measuring the temperature during expi-
ration and inspiration and of the air passing in the nasal
cannula. The subjects of the test (signal-testing group) vol-
untarily controlled and modulated their breathing in order
to verify the ability of the Smart-Bed to replay different
breathing rates without a significant delay. As shown in
Figure 6, the estimation of breathing rate extracted from
the Smart-Bed reproduces the reference data well with only
a small delay. The same accuracy was observed in all the
experimental sessions with different subjects and different
voluntarily-modulated respiratory rates.
Figure 7 shows three traces of heart rate estimation com-
pared with the heart rate obtained with reference ECG sig-
nals. The heart rate obtained by our mattress proved to be
very efficient and accurate. The evaluation of the mean heart
rate over the 30-second epoch was very stable and close
to real values. Considering the estimation over one-second
epochs , the estimated heart rate shows a low pass filtering be-
havior, with a loss of high frequency components. However,
the low-pass filtering showed no particular effects using the
30-second epochs for sleep staging and behavioural condition
classification. The estimation of heart rate via Smart-Bed-
BCG was accurate and robust when the subject was not
moving, as during sleep or in relaxed wakefulness before
sleep.
Figure 8 shows the results of the classification of the
position and motion of the subject on the mattress using
the ANN classifier on the ANN-training group. We obtained
an overall accuracy of about 91.8% and more specifically:
83.9% for not on bed, 83.8% for supine position, 96.4% for
lying on the left side, 94.5% for lying on the right side, 89.3%
for prone position, and 93.1% for movement.
We performed the training and testing of the automatic
VOLUME X, 20xx 5
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 5. Pressure maps measured in four typical sleeping postures: left side, supine, right side and prone.
FIGURE 6. Analysis of respiratory signals related to a subject in a supine
position: time course of breathing rate (Hz) obtained with Smart-Bed (blue
line) and with an estimation of the air-flow in the nasal cannula (reference
breath, red line). The respiratory frequency extracted from the Smart-Bed
accurately reflects the ground truth data.
classifier (decision tree algorithm) of behavioural conditions
on the condition-training group. From the eight nights of
recordings of the condition-training group, we collected a
total of 4761 epochs (i.e. about 40 hours). We removed 1280
artifactual epochs due to the poor EEG signal quality of
the polysomnography. We extracted 581 epochs of no bed
occupancy, 554 epochs of wakefulness, 2193 epochs of non-
REM sleep, and 153 epochs of REM sleep. The classification
performance obtained with the decision tree algorithm was
satisfactory in terms of aim f the classification. The overall
accuracy was about 86%, and specifically (see the confusion
matrix in Figure 9): 99% for "no bed occupancy", 83%
for "wakefulness", 83% for "non-REM sleep", and 79% for
"REM sleep".
Figure 10 reports the variation in estimated heart rate,
breathing rate, motion and behavioural conditions as a func-
tion of time (in epochs) for one subject in the condition-
training group. Within each epoch, the estimated physiologi-
cal and environmental information (see II-E2 section) is used
as input in the automatic classifier to evaluate the related
FIGURE 7. Analysis related to three different subjects: heart rate time course
(with different time lengths) detected by BGC with the Smart-Bed (blue lines)
and simultaneous heart rate signals collected by reference ECG system (red
lines). The estimation is performed with a 1 second time resolution.
behavioural condition. The behavioural condition pattern as-
sesses the sleep macrostructure (sleep staging) of the subject
and can reproduce the hypnogram.
Figure 11 reports the time course of the SQI and EQI for a
subject in the long-testing group, monitored with the Smart-
Bed for 60 days continuously. As regards the environmental
quality, the deviations in environmental parameters from the
optimal ranges are also reported. A comparison of the time
courses suggests that sub-optimal environmental conditions
have contributed to non- fully restorative sleep. However, it
seems that the subject adapted to the environmental condi-
6VOLUME X, 20xx
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 8. Confusion matrix related to the performance of the position and
motion classification (ANN algorithm) regarding the ANN-training group, with
six classes: i) not on bed, ii) supine position, iii) lying on the left side, iv) lying
on the right side, v) prone position and vi) movement. The positive predictive
values and false discovery rate are reported.
FIGURE 9. Confusion matrix related to the performance of the behavioural
conditions classification (decision tree algorithm) regarding the
condition-training group, with four classes: no bed occupancy, wakefulness,
non-REM sleep, and REM sleep. The positive predictive values and false
discovery rate are reported.
tions (from day 27), demonstrating an improvement in sleep
quality . During the durability tests, no functional issues, data
FIGURE 10. Physiological signals and behavioural conditions estimated from
the data collected by Smart-Bed for a subject (condition-training group) during
sleep. The estimated data are heart rate (beats per minute, bmp), breathing
rate (bpm) and motion (movement detected or not). The time unit is the epoch
(30 seconds).
losses or hardware faults were reported.
FIGURE 11. Single subject continuously monitored for 60 sleep nights. The
time course of the Sleep Quality Index (SQI, seven-level index) and the
Environmental Quality Index (EQI, dichotomic index) are shown in the upper
and lower panels, respectively. In the lower panel, for each parameter
contributing to EQI, up-pointing red triangles and down-pointing blue triangles
for each day indicate whether the corresponding parameter exceeded or losing
to the optimal range, respectively.
IV. DISCUSSION
The results of Section III show that our mattress is a valid
unobtrusive solution for detecting physical, physiological,
and environmental parameters.
To the best of the authors’ knowledge, our mattress is the
only in-bed solution that can simultaneously detect position
and motion as well as breathing and heart activity. The Whiz-
Pad [15] and the Kinotex [22] were tested for posture/motion
and breathing detection, while the solution developed by
Kortelainen [23] was assessed for breathing and cardiac ac-
VOLUME X, 20xx 7
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
tivity detection (i.e. no static postures due to the piezoelectric
technology). The other examples cited in Section I ( [18]–
[20], [24]) focus on the measurement of a single parameter
(position/movement or a single physiological parameter).
It should also be emphasized that some of the examples in
the literature show a much higher performance with regard
to the spatial resolution of the pressure mapping system.
For example, the XSENSOR pressure mapping system [25],
tested in [18] to prevent ulcers, has 6136 sensing areas (52
rows x 118 columns) that can be scanned at a rate of 1
Hertz. The Smart-Bed pressure mapping system has a smaller
number of sensors by a factor of about 30 (195 vs 6136).
If the objective was to monitor and prevent pressure ulcers,
the lower spatial resolution could be a limitation. However,
for a low-cost prototype suitable for future exploitation as a
consumer product that monitors sleep quality in the general
population, the lower spatial resolution is not a limitation.
In fact, we have demonstrated that the pressure mapping
system with a lower number of elements still leads to a good
evaluation of postural and motion parameters. Many of the
prototypes presented in the literature discussion in Section I
have a higher number of pressure sensing areas by a factor
greater than 10.
In addition, the Smart-Bed, unlike other solutions, is able
to monitor a set of basic environmental parameters (sound
intensity, relative humidity, temperature and luminosity) that
are known to be correlated with sleep quality.
From a technical point of view, we have demonstrated the
robustness of the Smart-Bed prototype which continuously
monitored five subjects for up to 231 days without significant
data losses (data collection and extraction of the SQI and EQI
indexes).
The most important result obtained with the Smart-Bed is
the classification of the main behavioural conditions charac-
terizing sleep, which is the basis for a robust extraction of an
index used to quantitatively assess sleep quality. To the best
of the authors’ knowledge, the existing smart bed solutions
have not been demonstrated as being capable of classifying
the different behavioural conditions that characterise sleep.
An approach to behavioural assessment based on machine
learning was developed and assessed by Wang et al. in [17]
with the WhizPad prototype [15]. However, in their work the
authors classified only two classes: sleep vs awake. In fact,
the ability to classify four conditions, in particular REM vs
non-REM, is a key feature of the design of our prototype. In
fact, despite the relative low number of sensors, the Smart-
Bed can detect relevant physiological and activity parameters
that are integrated through our machine-learning approach.
Monitoring environmental conditions together with physi-
ological parameters could lead to the development of specific
applications aimed at identifying causes and thus suggesting
solutions to the poor sleep quality. In addition, an integrated
environmental-subjective index of sleep quality is advisable.
V. CONCLUSIONS
In this work, we have described the architecture and prelim-
inary results of an innovative system for non-invasive sleep
monitoring in real world contexts. The proposed system,
developed within the L.A.I.D. research project, is a low-cost
smart mattress (Smart-Bed) capable of recording the sig-
nals related to the physiological (cardio-respiratory activity),
postural (body position and movements) and environmental
(temperature, noise, humidity and luminosity) aspects used
to identify the main stages of wakefulness and sleep and
to estimate a global sleep quality index and environmental
quality index.
Our preliminary tests with the Smart-Bed prototypes
verified the full functionality and robustness of the pro-
posed architecture, both considering the hardware and data-
processing solutions. However, the results presented herein
were obtained from a relatively small number of subjects.
In fact, we are in the process of performing additional in-
depth tests on a higher number of subjects to validate and
improve the prototypes. This will lead to the final version
of the Smart-Bed for the easy and low-cost monitoring of
sleep quality for large populations. Sleep investigations based
on a validated Smart-Bed solution would enable the sleep
macrostructure to be studied in real life. We are also de-
signing a second version of the prototype to improve various
engineering aspects. This includes a wireless connection
between the sensors and a remote data collection hub in
which the data are transmitted when the Smart-Bed detects
that the subject is not present.
ACKNOWLEDGMENT
Authors would like to thank the L.A.I.D. project industrial
partners: Materassificio Montalese S.P.A, EB Neuro S.P.A
and BP Engineering S.P.A.
REFERENCES
[1] L. A. Chilcott and C. M. Shapiro, “The socioeconomic impact of insom-
nia,” Pharmacoeconomics, vol. 10, no. 1, pp. 1–14, 1996.
[2] M. Daley, C. M. Morin, M. LeBlanc, J.-P. Grégoire, J. Savard, and
L. Baillargeon, “Insomnia and its relationship to health-care utilization,
work absenteeism, productivity and accidents,” Sleep medicine, vol. 10,
no. 4, pp. 427–438, 2009.
[3] D. Léger, C. Guilleminault, G. Bader, E. Lévy, and M. Paillard, “Medical
and socio-professional impact of insomnia,” Sleep, vol. 25, no. 6, pp. 621–
625, 2002.
[4] G. Medic, M. Wille, and M. E. Hemels, “Short-and long-term health
consequences of sleep disruption,” Nature and science of sleep, vol. 9, p.
151, 2017.
[5] J.-A. Palma, E. Urrestarazu, and J. Iriarte, “Sleep loss as risk factor for
neurologic disorders: a review,” Sleep medicine, vol. 14, no. 3, pp. 229–
236, 2013.
[6] M. G. Terzano, L. Parrino, F. Cirignotta, L. Ferini-Strambi, G. Gigli,
G. Rudelli, and S. Sommacal, “Studio morfeo: insomnia in primary care, a
survey conducted on the italian population,” Sleep medicine, vol. 5, no. 1,
pp. 67–75, 2004.
[7] M. Michal, J. Wiltink, Y. Kirschner, A. Schneider, P. S. Wild, T. Münzel,
M. Blettner, A. Schulz, K. Lackner, N. Pfeiffer et al., “Complaints of sleep
disturbances are associated with cardiovascular disease: results from the
gutenberg health study,” PloS one, vol. 9, no. 8, p. e104324, 2014.
[8] N. J. Douglas, S. Thomas, and M. A. Jan, “Clinical value of polysomnog-
raphy,” The Lancet, vol. 339, no. 8789, pp. 347–350, 1992.
8VOLUME X, 20xx
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
[9] A. Crivello, P. Barsocchi, M. Girolami, and F. Palumbo, “The meaning of
sleep quality: a survey of available technologies,” IEEE Access, 2019.
[10] C. Iber, “The aasm manual for the scoring of sleep and associated events:
Rules,” Terminology and Technical Specification, 2007.
[11] J. M. Kortelainen, M. O. Mendez, A. M. Bianchi, M. Matteucci, and
S. Cerutti, “Sleep staging based on signals acquired through bed sensor,”
IEEE Transactions on Information Technology in Biomedicine, vol. 14,
no. 3, pp. 776–785, 2010.
[12] “Laid project,” https://media- perdormire-com.s3.amazonaws.com/com/
LAID_SMART_BED.pdf, accessed: 2019-11-03.
[13] “Materassificio montalese, pistoia, italy.” [Online]. Available: \url{https:
//www.materassificiomontalese.com}
[14] A. Schwarz-Pfeiffer, M. Hoerr, and V. Mecnika, “Textiles with integrated
sleep-monitoring sensors,” in Advances in Smart Medical Textiles. Else-
vier, 2016, pp. 197–214.
[15] Y.-W. Liu, Y.-L. Hsu, and W.-Y. Chang, “Development of a bed-centered
telehealth system based on a motion-sensing mattress,” Journal of Clinical
Gerontology and Geriatrics, vol. 6, no. 1, pp. 1–8, 2015.
[16] C. Lu, J. Huang, Z. Lan, and Q. Wang, “Bed exiting monitoring system
with fall detection for the elderly living alone,” in 2016 International
Conference on Advanced Robotics and Mechatronics (ICARM). IEEE,
2016, pp. 59–64.
[17] C. Wang, T.-Y. F. Chiang, S.-H. Fang, C.-J. Li, and Y.-L. Hsu, “Machine
learning based sleep-status discrimination using a motion sensing mat-
tress,” in 2019 IEEE International Conference on Artificial Intelligence
Circuits and Systems (AICAS). IEEE, 2019, pp. 160–162.
[18] H. Wong, J. Kaufman, B. Baylis, J. M. Conly, D. B. Hogan, H. T. Stelfox,
D. A. Southern, W. A. Ghali, and C. H. Ho, “Efficacy of a pressure-sensing
mattress cover system for reducing interface pressure: study protocol for a
randomized controlled trial,” Trials, vol. 16, no. 1, p. 434, 2015.
[19] W.-Y. Chang, C.-C. Huang, C.-C. Chen, C.-C. Chang, and C.-L. Yang,
“Design of a novel flexible capacitive sensing mattress for monitoring
sleeping respiratory,” Sensors, vol. 14, no. 11, pp. 22 021–22 038, 2014.
[20] W.-Y. Chang, C.-C. Chen, C.-C. Chang, and C.-L. Yang, “An enhanced
sensing application based on a flexible projected capacitive-sensing mat-
tress,” Sensors, vol. 14, no. 4, pp. 6922–6937, 2014.
[21] K. Sakai, G. Nakagami, N. Matsui, H. Sanada, A. Kitagawa, E. Tadaka,
and J. Sugama, “Validation and determination of the sensing area of the
kinotex sensor as part of development of a new mattress with an interface
pressure-sensing system,” Bioscience trends, vol. 2, no. 1, pp. 36–43, 2008.
[22] S. S. Gilakjani, M. Bouchard, R. A. Goubran, and F. Knoefel, “Long-term
sleep assessment by unobtrusive pressure sensor arrays,” in Proceedings
of the 2018 3rd International Conference on Biomedical Imaging, Signal
Processing. ACM, 2018, pp. 23–35.
[23] J. M. Kortelainen and J. Virkkala, “Fft averaging of multichannel bcg sig-
nals from bed mattress sensor to improve estimation of heart beat interval,”
in 2007 29th annual international conference of the IEEE engineering in
medicine and biology society. IEEE, 2007, pp. 6685–6688.
[24] R. Grimm, S. Bauer, J. Sukkau, J. Hornegger, and G. Greiner, “Markerless
estimation of patient orientation, posture and pose using range and pres-
sure imaging,” International journal of computer assisted radiology and
surgery, vol. 7, no. 6, pp. 921–929, 2012.
[25] “Xsensor technology,” https://xsensor.com, accessed: 2019-11-01.
[26] “Sensoredge,” http://sensoredge.com/flexible_sensors.html, accessed:
2019-11-02.
[27] “Boditrak,” https://www.boditrak.com/products/medical.php, accessed:
2019-11-02.
[28] J. Cheng, M. Sundholm, B. Zhou, M. Hirsch, and P. Lukowicz, “Smart-
surface: Large scale textile pressure sensors arrays for activity recogni-
tion,” Pervasive and Mobile Computing, vol. 30, pp. 97–112, 2016.
[29] B. Zhou and P. Lukowicz, “Textile pressure force mapping,” in Smart
Textiles. Springer, 2017, pp. 31–47.
[30] M. Nakano, T. Konishi, S. Izumi, H. Kawaguchi, and M. Yoshimoto,
“Instantaneous heart rate detection using short-time autocorrelation for
wearable healthcare systems,” in 2012 Annual International Conference
of the IEEE Engineering in Medicine and Biology Society. IEEE, 2012,
pp. 6703–6706.
[31] E. Öhrström and A. Skånberg, “Sleep disturbances from road traffic and
ventilation noise—laboratory and field experiments,” Journal of Sound and
Vibration, vol. 271, no. 1-2, pp. 279–296, 2004.
[32] B. Berglund, T. Lindvall et al., Community noise. Center for Sensory
Research, Stockholm University and Karolinska Institute ..., 1995.
[33] Z. A. Caddick, K. Gregory, L. Arsintescu, and E. E. Flynn-Evans, “A
review of the environmental parameters necessary for an optimal sleep
environment,” Building and environment, vol. 132, pp. 11–20, 2018.
[34] C.-H. Cho, H.-K. Yoon, S.-G. Kang, L. Kim, E.-I. Lee, and H.-J. Lee, “Im-
pact of exposure to dim light at night on sleep in female and comparison
with male subjects,” Psychiatry investigation, vol. 15, no. 5, p. 520, 2018.
[35] C. Brüser, S. Winter, and S. Leonhardt, “How speech processing can help
with beat-to-beat heart rate estimation in ballistocardiograms,” in 2013
IEEE International Symposium on Medical Measurements and Applica-
tions (MeMeA). IEEE, 2013, pp. 12–16.
[36] I. Sadek and J. Biswas, “Nonintrusive heart rate measurement using
ballistocardiogram signals: a comparative study,” Signal, Image and Video
Processing, vol. 13, no. 3, pp. 475–482, 2019.
[37] A. Landi, P. Piaggi, M. Laurino, and D. Menicucci, “Artificial neural
networks for nonlinear regression and classification,” in 2010 10th Interna-
tional Conference on Intelligent Systems Design and Applications. IEEE,
2010, pp. 115–120.
[38] M. F. Møller, “A scaled conjugate gradient algorithm for fast supervised
learning,” Neural networks, vol. 6, no. 4, pp. 525–533, 1993.
[39] T. G. Dietterich, “An experimental comparison of three methods for
constructing ensembles of decision trees: Bagging, boosting, and random-
ization,” Machine learning, vol. 40, no. 2, pp. 139–157, 2000.
[40] M. M. Ohayon, M. A. Carskadon, C. Guilleminault, and M. V. Vitiello,
“Meta-analysis of quantitative sleep parameters from childhood to old
age in healthy individuals: developing normative sleep values across the
human lifespan,” Sleep, vol. 27, no. 7, pp. 1255–1273, 2004.
[41] A. Gemignani, A. Piarulli, D. Menicucci, M. Laurino, G. Rota, F. Mastorci,
V. Gushin, O. Shevchenko, E. Garbella, A. Pingitore et al., “How stressful
are 105 days of isolation? sleep eeg patterns and tonic cortisol in healthy
volunteers simulating manned flight to mars,” International Journal of
Psychophysiology, vol. 93, no. 2, pp. 211–219, 2014.
MARCO LAURINO (PhD) obtained his specialist
degree in biomedical engineering at the University
of Pisa in 2009. In 2014, he received the PhD in
Neurosciences and Endocrine-Metabolic Sciences
at the University of Pisa. He is a researcher at
the Institute of Clinical Physiology of the Ital-
ian National Research Council. Currently, his re-
search interests include: i) processing of biomed-
ical signals (e.g. High-density-EEG, ECG, skin
conductance, actigraphy); ii) artificial intelligence
algorithms applied to medicine, iii) study of the cognitive and emotional
aspects in humans through functional neuroimaging techniques; iv) neuro-
stimulation techniques (transcranial magnetic stimulation and transcranial
current stimulation) applied to psychopathological conditions; v) modelling
of biomedical systems and control algorithms applied to biomedical systems;
vi) development of biomedical technologies in underwater and hyperbaric
environments; vii) telemedicine and Connected Health.
LUCIA ARCARISI is Research fellow at the In-
formation Engineering Department of the Univer-
sity of Pisa. She received the bachelor degree in
Electronic Engineering at University of Palermo
and the Master degree in Biomedical Engineer-
ing at University of Pisa in 2018. She has been
involved in the design of medical devices and
wearable electronic solutions for monitoring and
rehabilitation. She has good skills in rapid proto-
typing, CAD design and Arduino programming.
VOLUME X, 20xx 9
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2976194, IEEE Access
Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
NICOLA CARBONARO (PhD) is assistant Pro-
fessor at the Information Engineering Department
and the Research Center "E. Piaggio" of the Uni-
versity of Pisa. He graduated in Electronic Engi-
neering at the University of Pisa in 2004. In 2010,
he earned a PhD in Information Engineering from
University of Pisa working on the development
of wearable system for human activity classifica-
tion. In 2009, he spent six months as a visiting
researcher at the “Neural Control of Movement”
Laboratory of Arizona State University. His research is mainly focused on
hardware and software development for wearable sensing technology for
physiological and behavioural human monitoring for biomedical applica-
tions. Since 2014, he is chair of the curse of Biosensors of the Biomedical
Engineering Degree of the University of Cagliari. Dr. Carbonaro has collabo-
rated on different research projects both at National and European level and
he has published several papers, contributions to international conferences
and books’ chapters.
ANGELO GEMIGNANI (MD, PhD, M.Sc.) is
medical doctor, psychiatrist and doctor in psychol-
ogy. As full professor of neuroscience at Univer-
sity of Pisa, he is a) Director of the Department
of Surgical, Medical and Molecular Pathology and
Critical Care Medicine; b) Director of the Master
in Neuroscience, Mindfulness and Contemplative
Practices; and c) Director of Clinical Psychology
branch of the Pisa University Hospital. His re-
search activity is mainly devoted to the study of
psychobiological mechanisms of a) sleep functions, b) consciousness and
its related non-ordinary states (i.e. induced by meditation), c) negative
emotions. As P.I. of international research projects on adaptive mental
and somatic reactions to huge stressful conditions, he studied sleep and
consciousness changes in different human experimental models from elite
apnea divers to healthy volunteers simulating the human flight to Mars in
the context of MARS 500 and SIRIUS frameworks projects.
DANILO MENICUCCI (PhD) received the M.S.
degree in Applied Physics and the Ph.D. degree in
Basic Neuroscience from the University of Pisa,
Pisa, Italy. He is currently a Researcher in Psy-
chophysiology with the Department of Surgical,
Medical and Molecular Pathology and Critical
Care Medicine, Pisa. His current research inter-
ests include sleep psychophysiology, cognitive and
emotional modulation of brain, cognitive neuro-
science, and experimental psychology and neuro-
physiology.
ALESSANDRO TOGNETTI (PhD) is Associate
Professor of Bioengineering at the Department of
Information Engineering and the "E.Piaggio" Re-
search Center of the University of Pisa. He gradu-
ated in Electronic Engineering and obtained a PhD
in Bioengineering at the University of Pisa. He is
vice-president of the degree course in Biomedical
Engineering of the University of Pisa where he
holds the courses in Biosensors and Bioelectric
Phenomena. His research activity is focused on
the design and development of wearable multisensory technologies for
measurement and analysis - outside the laboratory and during daily activities
- of posture and human movement and physiological signals. The main
scientific skills range from the development of innovative biomedical sensors
- starting from the physical principle and enabling technology - to system
integration, from signal processing and information to the development of
ICT applications for health monitoring, rehabilitation, robotics and human-
machine interaction. The main scientific interests concern the development
of soft sensors (flexible, e-textiles) for monitoring and analyzing human
movement (upper and lower limbs with particular attention to both phys-
ical and neurological rehabilitation applications), non-invasive biomedical
devices (physiological signals, physical activity, innovative interfaces for
prostheses) and sensory fusion (fusion between physiological signals and
inertial sensors for an accurate classification of human activity, fusion of soft
and inertial sensors to improve movement reconstruction performance by
systems wearable). Much of the research was carried out in the context of na-
tional and international research projects (more than 20 projects of which 12
European) and through numerous scientific collaborations with Universities,
Research Institutes and companies both nationally and internationally. His
research activity has generated more than 100 international contributions.
10 VOLUME X, 20xx
