SLEEPIC

Developments for a Wearable On-line Sleep and Wake Discrimination System

Walter Karlen

Electrical and Computer Engineering in Medicine Group, Department of Electrical and Computer Engineering

The University of British Columbia (UBC), 2332 Main Mall, V6T 1Z4, Vancouver, BC, Canada

Dario Floreano

Laboratory of Intelligent Systems, Institute of Micro-engineering

Ecole Polytechnique F

erale de Lausanne (EPFL), CH-1015 Lausanne, Switzerland

Keywords:

Wearable, Context awareness, Sleep and wake discrimination, On-line classiﬁcation, Embedded intelligence.

Abstract:

The design of wearable systems comes with constraints in computational and power resources. We describe

the development of customized hardware for the wearable discrimination of human sleep and wake based on

cardio-respiratory signals. The device was designed for efﬁcient and low-power computation of Fast Fourier

Transforms and artiﬁcial neural networks required for the on-line classiﬁcation. We discuss methods for

reducing computational load and consequently power requirements. The SleePic prototype was tested for

autonomy and comfort on eight healthy subjects. SleePic showed an energetic autonomy of more than 36

hours. The SleePic device will require further integration for increased comfort and improved user interaction.

1 INTRODUCTION

Monitoring sleep and wake behavior of subjects at

home allows the early detection of sleep troubles and

disorders, and can reduce health care costs (Colten

and Altevogt, 2006). Home monitoring of sleep/wake

behavior imposes particular challenges, namely that

necessary sensors, electronics and intelligent signal

processing algorithms require integration into a com-

fortable, wearable device.

The most common physiological signal used for

sleep discrimination in clinical sleep monitoring is the

recording of brain activity with an electroencephalo-

gram (EEG) (Ogilvie, 2001). Unfortunately, EEG

cannot be easily recorded with a wearable system. Al-

ternatively, actigraphs are often used for long term

sleep studies (Sadeh, 2002). Actigraphy is a pas-

sive measure of sleep/wake behavior. The actigraphy

wristbands are small, lightweight, low-power, and

therefore easy to wear over several days. However,

actigraphy algorithms often incorrectly classify low

activity tasks (e.g. reading or watching television) as

sleep because the measured behavioral quiescence is

not unique to sleep (Sadeh, 2002; de Souza et al.,

2003). In previous studies we showed that sleep /

wake classiﬁcation is possible with frequency-domain

features of cardio-respiratory signals using a single-

layer, feed-forward artiﬁcial neural network (ANN)

(Karlen et al., 2008; Karlen et al., 2009).

Comfort and wearability imposes constraints on

the design of devices. For instance, micro-controllers

are preferred to the high-performance 32-bit ap-

plication processors from the standpoint of power

consumption (i.e. compared <0.1 W for micro-

controllers to 1-2 W for application processors used

in smart phones). However, micro-controllers are

limited by the computational resources they provide,

as well as being less ﬂexible by not running operat-

ing systems and rich data formats. These constraints

therefore impose particular challenges in the design

of accompanying digital signal processing (DSP) al-

gorithms to have low computational complexity.

We have developed a wearable hardware sys-

tem called SleePic (derived from Sleep discrimination

with a Programmable interface controller).

The SleePic implements the preprocessing and

ANN classiﬁcation principles previously suggested

in (Karlen et al., 2009). We describe the research

and development of the wearable hardware design of

SleePic and present methods to minimize the power

132

Karlen W. and Floreano D..

SLEEPIC - Developments for a Wearable On-line Sleep and Wake Discrimination System.

DOI: 10.5220/0003131701320137

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2011), pages 132-137

ISBN: 978-989-8425-35-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

consumption of the classiﬁcation algorithm to be em-

bedded in wearable systems. We also describe valida-

tion experiments with data obtained from eight users

wearing SleePic for 36 hours.

2 HARDWARE REQUIREMENTS

For the development of the wearable sleep / wake

discrimination system, we formulated a series of de-

sign criteria that were based on general recommenda-

tions for a wearable bio-medical device (Martin et al.,

2000; Schefﬂer and Hirt, 2004) and the classiﬁcation

algorithm described in (Karlen et al., 2009). The sys-

tem was expected to:

• Be wearable. From sensors to user feedback, we

wanted to integrate all elements of the device into

a wearable system. This implies low size and low

weight. It is desirable that the wearable system

is comfortable and does not display any cables or

wires.

• Be autonomous: The system is expected to have

power autonomy for at least 24 hours. After us-

age, easy recharge is a plus.

• Record cardio-respiratory data. The algorithm in

(Karlen et al., 2009) processes electrocardiogram

(ECG) and respiratory (RSP) signals for classiﬁ-

cation.

• Be able to execute the classiﬁcation algorithm on-

line. The processing resources on the system need

to cope with the signal processing and classiﬁ-

cation in minimal time without excessive energy

consumption.

• Give feedback to the user. The system needs an

interface to communicate with the wearer. The

feedback method has to adhere to the energy re-

strictions.

To our knowledge, there exist no systems on

the market that comply with all these design crite-

ria. Therefore, to embed the classiﬁcation algorithm

developed in (Karlen et al., 2009) into a wearable

demonstrator, we built our own prototype. To keep

development cost low, we relied on commercially

available sensors and components for prototyping.

Missing functionality, such as user interface and com-

putational resources for on-line classiﬁcation, were

added by designing custom extension modules.

3 SLEEPIC DESCRIPTION

The SleePic sleep / wake discrimination system is

composed of three modules (Figure 1):

Figure 1: A subject wearing the complete SleePic sleep

/ wake detection system. 1) Equivital

Sensor Belt; 2)

SleePic Core extension module; and 3) SleePic Watch for

user interaction.

1. The SleePic Sensor module is worn around the

chest.

2. The SleePic Core processing module is responsi-

ble for all signal preprocessing and classiﬁcation.

The module connects directly to the SleePic Sen-

sor module.

3. The SleePic Watch is worn on the wrist. The

SleePic Watch is responsible for remote sensing

and user feedback. It communicates wirelessly

with the SleePic Core.

3.1 Wearable Sensor Module

Sensors integrated into textiles appeared on the mar-

ket only recently. The technology is not yet estab-

lished and expensive. The Equivital

physiological

monitoring system (Hidalgo Ltd, UK) has been cho-

sen as sensor module for this system development. It

offers a good compromise between cost, wearability

and real time sensor access. It offers high integration

of sensors and textile electrodes in a single belt worn

across the upper chest area, which facilitates the cor-

rect wearing and use of the system by inexperienced

users. Equivital

can operate continuously for up to

48 hours with the integrated, rechargeable Lithium-

Ion battery (760 mAh). However, there is a lack of

user feedback possibilities, and options for comput-

ing the classiﬁcation algorithm on-line are also miss-

ing. Equivital

is composed of two units:

1. A washable sensor belt which integrates 3 textile

dry electrodes for 2-lead ECG recordings and a

piezo-resistive strain gauge to measure RSP. The

sensor belt weights 94 grams (size M).

SLEEPIC - Developments for a Wearable On-line Sleep and Wake Discrimination System

133

2. A Sensor Electronics Module (SEM) that can be

attached to the sensor belt with ﬁve connective

clips. The SEM measures 3-axis acceleration

(ACC), temperature and features a real time clock

(RTC). The SEM also embeds a vibrating actu-

ator. The SEM weights 75 grams including the

battery.

3.2 SleePic Core Processing Module

The SleePic Core (Figure 2) is a printed circuit board

(PCB) that plugs to the SEM. It was designed for re-

ceiving the raw sensor data from the Wearable Sensor

module and the SleePic Watch , the processing of the

sensor signals, executing the classiﬁcation task, and

the coordination of the tasks between all three SleePic

modules.

The SleePic Core is composed of a process-

ing unit, a power unit, storage unit and a com-

munication unit. The processing unit is the cen-

tral element on the SleePic Core and contains the

40 MIPS dsPIC33256GP710 micro-controller (Mi-

crochip Technology Inc., USA). This programmable

interface controller has a 16-bit architecture, 32 kB

RAM, 256 kB program memory and DSP function-

ality. It has been chosen because of the presence of

multiply-and-accumulate (MAC) and a barrel shifter

functionality which simpliﬁes the Fast Fourier Trans-

formation (FFT) and ANN calculation. At time

of development, it was the largest available micro-

controller from the dsPIC33 series. The relatively

high amount of RAM was required for the FFT

preprocessing. The communication unit contains a

UART port for the communication with the SEM, an-

other UART port to communicate with a PC over a

USB converter and an SPI port for the NRF02L wire-

less chip (Nordic Semiconductor ASA, Norway) to

communicate with the SleePic Watch module. For

saving energy, each of the units can be disabled when

not needed by the micro-controller, including itself.

The dimensions of the SleePic Core module are 65

mm×35 mm×8 mm, and weights 8 grams without

battery.

3.3 SleePic Watch Module

The SleePic Watch module is used for the SleePic

user interaction (Figure 2). The module is placed on

the wrist to be highly visible to the user. The user in-

teraction unit is composed of 5 LED’s in the colors of

green, orange, red. A button provides the user possi-

bilities for feedback. The sensor unit contains a 3-axis

accelerometer (Freescale Semiconductor, USA) and a

photo-diode to measure ambient light. The data from

Antenna

Power

Switch

Micro-

controller

Wireless

Chip

MicroUSB

Connector

Button

LED’s

Photodiode

On/o

Switch

Figure 2: Bottom view of the SleepPic Core processing

module (left). Top view of the SleePic Watch module pro-

totype (right).

these sensors were not evaluated in this study. The

communication unit is composed of a USB-UART

converter which allows data exchange and recharg-

ing via microUSB The diameter of the SleePic Watch

module is 41 mm and the height 10 mm. A fully as-

sembled SleePic Watch weights 28 grams.

4 EXPERIMENTS

A series of recordings were conducted to test the

SleePic device and to obtain sleep / wake data in real-

life situations. Following informed consent, eight

(two female and six male) volunteers in the age be-

tween 24 and 30 years wore the SleePic system. The

subjects were in good health and reported no cardio-

respiratory disease or any sleep disorders. The sub-

jects came to the laboratory in the evening and were

instructed about the experiment procedure and how to

wear the device. The subjects wore the SleePic device

for a minimum of 36 hours that included two nights.

They were allowed to remove the belt during heavy

sport or when showering. During the experiment,

the subjects performed a randomly scheduled reaction

task using the button on the SleePic Watch. The sub-

jects were asked to sleep at home. After the recording,

the subjects returned the SleePic recording system to

the laboratory, were debriefed, and ﬁlled out a ques-

tionnaire about the usability and comfort of the sys-

tem. Because of the ambulatory nature of the experi-

ment, the subjects were expected to move freely and

perform normal daily activities. Therefore, we did not

consider the possibility of recording EEG signals for

reference. Instead, the subjects had to maintain a log-

book by indicating the system-off times, their sleep

times, and particular events related to the system that

may happen during the experiment. Additionally, a

technician installed an infra-red video camera in the

bedroom to record the sleep behavior during bedtime.

After the study, the technician analyzed the logbook

entries and video recordings and labeled the wake /

sleep periods of the subjects.

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

134

5 RESULTS & DISCUSSION

Using the SleePic system, 250 hours of valid record-

ings were obtained and 37 % of which were labeled by

the technician as sleep. Classiﬁcation and user adap-

tation experiments involved the cardio-respiratory

and activity data obtained. These experiments were

designed to assess the sleep/wake prediction perfor-

mance and are not described in details in this paper.

5.1 User Acceptance

From the post-study questionnaire, we were able to

evaluate the subjects’ acceptance of the SleePic proto-

type. Interestingly, the female subjects felt the wear-

able device comfortable enough, whereas the male

subjects either could not decide (50 %) or were not

comfortable with it (50 %). The male participants ar-

gued that they could not get used to the belt because

of it was either too tight or too big. The SleePic sys-

tem was most disturbing when the subjects were lying

face down. An integration of the SleePic Core ex-

tension module into the Sensor module would reduce

the overall size of the system worn on the chest and

decrease disturbance when lying on it. All subjects

found the light vibration that accompanied the reac-

tion task helpful and not as disturbing. The subjects

also agreed with the frequency of the reaction time

tasks. Wearing the device was inﬂuencing the daily

activities of the subjects only slightly (60 %) or not

at all (40 %). Compared to the perceived decrease in

comfort from an actigraphy device, the users did not

have a visible advantage from the device in this study.

Additional experiments with a device displaying on-

line sleep/wake schedules are required. For this, a

low-power liquid crystal display might be added to

the SleePic Watch. We also plan to demonstrate im-

proved comfort by embedding the SleePic Core mod-

ule into the body of the Sensor module.

5.2 Reducing Computational Load of

Preprocessing

Although the dsPic micro-controller with the highest

amount of RAM memory available was embedded in

the SleePic Core, the RAM was not sufﬁcient to cal-

culate the FFT of ECG and RSP at the Sensor Module

sampling rates. As the high frequency features were

not needed for the classiﬁcation, it was reasonable to

reduce the sampling rate accordingly. We therefore

sub-sampled the ECG from 256 Hz to 51.2 Hz and

maintained the RSP at the recording sampling rate of

the Equivital

(25.6 Hz). For an efﬁcient calculation

of the FFT on the dsPic micro-controller, the sliding

5 10 20 40

Test Accuracy [%]

a) FFT Window Size [s]

original pruned

b) ANN Input

Figure 3: a) Performance analysis for different sliding win-

dow sizes (5, 10, 20 and 40 seconds) for the FFT prepro-

cessing. b) Performance comparison when the entire fea-

ture set (original) or the reduced feature set (pruned) is used

as input of the ANN. The bullets represent the mean over

all trials and subjects, the bars cover the standard deviation.

window size had to contain power of two sampling

points for each window. The use of a window size

of 40 seconds was suggested (Karlen et al., 2009).

For our sampling rate, this window size resulted in

1024 (RSP) and 2048 (ECG) samples. The window

ﬁtted the available RAM, but left only little space for

other variables. Therefore, the processing load was

further reduced by decreasing the window size for the

FFT preprocessing. However, by the nature of the

FFT, with an increase in time resolution, the resolu-

tion of the frequency output decreases. The reduction

of the window from 40 to the next possible size of 20

seconds resulted in a frequency resolution shift from

0.0122 Hz to 0.025 Hz. This change could result in

a change in classiﬁcation accuracy which had to be

investigated.

We conducted an off-line experiment using data

from (Karlen et al., 2008) and compared the classi-

ﬁcation accuracies depending of the size of the win-

dows. Figure 3a shows the test results when the FFT

was calculated on a window size of 5, 10, 20 and 40

seconds respectively. Smaller sliding windows than 5

seconds were not meaningful because all useful fre-

quencies contained in the original physiological sig-

nal would lie in one frequency band. Larger win-

dows could not be computed with the dsPic micro-

controller and were not considered. Although the ac-

curacies for the different window sizes were not sig-

niﬁcantly different (Figure 3a), we observed a posi-

tive trend toward the 20-second segments. This was

expressed with the highest mean performance, mini-

mal standard deviation and the highest total accuracy.

It was therefore reasonable to use a 20-second win-

dow instead of the 40-seconds suggested in (Karlen

et al., 2009) for the FFT computation on the SleePic

device.

5.3 Network Inputs

We analyzed the input network weights of the sin-

gle layer ANNs obtained from the previous experi-

SLEEPIC - Developments for a Wearable On-line Sleep and Wake Discrimination System

135

ment where the entire frequency spectrum served as

input. We observed that the weights of the neural in-

puts with the higher frequencies of the ECG and RSP

spectrum did show a very weak activation and varia-

tion compared to the weights of the low frequencies.

Based on physiological reasons and because the sim-

ple structure of the ANN, we hypothesized that the

low activation and variation in the domain above 3

Hz is linked to a reduced importance of these features

for the classiﬁcation. Therefore we designed a differ-

ent network topology that only used the most relevant

input weights. In our particular case (single-layered

network), all input features i ∈ 1, ..., N were consid-

ered as relevant when the mean weight over all train-

ing runs 1 to M was larger than the median standard

deviation of all layer weights of all runs, as follows

S(w

) =







1 mean(~w

)

> median(std(~w

1,...,N

))

0 otherwise,

(1)

where ~w

= (w

,. .. ,w

) and S is the selection func-

tion. The resulting pruned network input numbers and

the pruning frequencies for each signal are shown in

Table 1. The input size of the pruned network was

reduced to 8.3 % of its original size.

We compared the performance of the reduced net-

works with the networks using all frequency inputs.

Figure 3b shows the accuracy for the original and

pruned network inputs when the experiments from

(Karlen et al., 2008) were repeated. We observed

that the accuracies showed no statistical difference,

but revealed a tendency for a higher median and bet-

ter worst performance for the pruned input. The pos-

itive effects in accuracy of the pruned ANNs can be

attributed to the reduction of the search space and

therefore to the reduction of over-ﬁtting due to the

increased noise on the higher-frequency inputs. A

smaller input feature set did accelerate the training

and the network computation, and also reduced the

required amount of training input vectors (curse of di-

mensionality). It was therefore reasonable to use the

pruned network for the SleePic implementation.

Using DSP functionality, 202’742 instruction cy-

cles were necessary to compute a single classiﬁcation.

If the device is running at the maximum speed of 40

MIPS, the required processing time is about 10 ms.

5.4 Energy Consumption

For a wearable device it is important to keep energy

consumption minimal. In the initially designed con-

ﬁguration, both energetically independent systems on

chest and wrist achieved an energetic autonomy of 36

hours as required by the design criteria. Both systems

could easily be recharged by connecting them to a PC.

Table 1: Properties of the input feature space for the ANN

topology. The original features are all the frequency bins

available from the FFT preprocessing, as used in (Karlen

et al., 2008). The pruned features correspond to the features

obtained by applying Equation 1.

RSP ECG Total

Frequencies [Hz]

original 0-12.8 0-25.6

pruned 0-1.4 0-2.25

ANN Input Size [number]

original 257 513 770

pruned 28 46 74

5.4.1 SleePic Core and Sensor Module

The total average power consumption of 18.06 mW

(6.02 mA @ 3 V) measured at room temperature lead

to an autonomy of 36 hours. We observe that the ma-

jor energy resources went to the dsPic and the wire-

less chip (NRF) in rx (receiving) mode. Compared to

a standard micro-controller, the dsPic was relatively

power hungry, because of the high clocking and the

large RAM size. The power consumption was re-

duced by putting the dsPic into a very efﬁcient idle

mode (0.1 mA) at 93 % of the time.

In a future design, a signiﬁcant decrease in en-

ergy consumption (50 %) could be achieved when the

Recording Mode would be substituted by a single re-

quest for the 20-second sensor data packet instead of

the continuous recording of the data stream. This was

not possible with the Equivital

because it only al-

lowed a single continuous output stream for the sen-

sor values. The total consumption could further be

decreased by 20 % by reducing the duty time of the

wireless chip. This could be achieved by implement-

ing an updated communication protocol that would

not require a continuous powering of the wireless chip

while waiting for a message from the SleePic Watch.

A more radical, but surely very energy efﬁcient

and power saving option would be the replacement

of the dsPic micro-controller with a dedicated silicon

chip for the recording, preprocessing and ANN cal-

culation. However, these types of chips are very task

dependent and their development expensive. There-

fore we preferred the more ﬂexible micro-controller

solution for the prototyping.

5.4.2 SleePic Watch

The total average power consumption of 10.36 mW

(2.8 mA @ 3.7 V) lead to an autonomy of 51 hours

when using the 145 mAh Lithium-Polymer battery.

The energy autonomy of the SleePic Watch could be

increased in a next step by selecting a smaller and

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

136

more energy efﬁcient micro-controller that also of-

fers a lower stand-by power consumption. The func-

tion of the SleePic Watch in our experiments was to

attract attention from the users during the reaction

task (LEDs) and provide an interface for respond-

ing (button). The peripheral location of the SleePic

Watch required wireless transmission of data and an

independent energy storage. The frequent wireless

communication increases the power consumption of

the SleePic Watch and Core. In future versions it is

imaginable to replace the custom SleePic Watch with

another, already existing pervasive device such as a

smart phone or digital wristwatch.

6 CONCLUSIONS

We developed a SleePic prototype based on com-

mercially available sensors and components. Miss-

ing functionality, such as user interface and com-

putational resources for on-line classiﬁcation, were

added by custom designed extension modules. A DSP

micro-controller was used to efﬁciently compute FFT

for extracting the frequency features of the signals and

to compute the ANN for classiﬁcation. FFT and ANN

were optimized for speed and lower power consump-

tion. We used no liquid crystal display. To render the

feedback information more meaningful to the user, it

might be advantageous to integrate a liquid crystal

display in a next step into the SleePic Watch.

Although cardio-respiratory signals were mea-

sured non-invasively with a wearable belt, it was more

cumbersome to use than a wrist worn actigraphy de-

vice. It is therefore application-dependent whether

a single actigraph device or a combined system like

SleePic should be adopted. If, for example, the appli-

cation is oriented toward sleep disorder diagnosis, the

system with additional cardio-respiratory functional-

ity offer a clear advantage, because current standards

in sleep disorder diagnosis require the recording of

these signals (Patel and Davidson, 2007).

The SleePic prototype demonstrated wearable

sleep/wake classiﬁcation that is less obtrusive than the

EEG monitors currently used in sleep centers. We

also showed how existing classiﬁcation algorithms

can be modiﬁed in order to comply with the narrow

constraints given by wearable systems.

The SleePic prototype can be considered as a

proof-of-concept for a new generation of health and

wellness devices. It shows the feasibility of a wear-

able, on-line sleep/wake classiﬁer using low-cost

components. The minimal processing power require-

ments of the presented system opens new doors for

the use of high-level algorithms required for context-

awareness or automated diagnostics in point-of-care

health care.

ACKNOWLEDGEMENTS

We thank all subjects for participating in the SleePic

validation study. Dr. Steffen Wischmann and Michael

Chiang provided valuable comments on previous ver-

sions of this manuscript. Adam Klaptocz and James

Roberts contributed to the PCB design. Andr

Badertscher and Peter Br

uhlmeier helped with the

manufacturing and assembling of SleePic. We would

like to thank Dr. med Werner Karrer, Dr. med Thomas

Rote and Isabelle Arnold of the Luzerner H

ohenklinik

Montana, Switzerland, for providing expert knowl-

edge on sleep analysis.

REFERENCES

Colten, H. and Altevogt, B. (2006). Sleep disorders and

sleep deprivation: an unmet public health problem.

National Academies Press, Washington, DC.

de Souza, L., Benedito-Silva, A., Pires, M., Poyares, D.,

Tuﬁk, S., and Calil, H. (2003). Further validation of

actigraphy for sleep studies. Sleep, 26(1):81–5.

Karlen, W., Mattiussi, C., and Floreano, D. (2008). Improv-

ing actigraph sleep/wake classiﬁcation with cardio-

respiratory signals. In Annual International Confer-

ence of the IEEE Engineering in Medicine and Biol-

ogy Society, pages 5262–5.

Karlen, W., Mattiussi, C., and Floreano, D. (2009). Sleep

and Wake Classiﬁcation With ECG and Respiratory

Effort Signals. IEEE Transactions on Biomedical Cir-

cuits and Systems, 3(2):71–8.

Martin, T., Jovanov, E., and Raskovic, D. (2000). Issues

in Wearable Computing for Medical Monitoring Ap-

plications: A Case Study of a Wearable ECG Moni-

toring Device. International Symposium on Wearable

Computers ISWC 2000, Atlanta, October.

Ogilvie, R. (2001). The process of falling asleep. Sleep

Medicine Reviews, 5(3):247–70.

Patel, M. R. and Davidson, T. M. (2007). Home sleep test-

ing in the diagnosis and treatment of sleep disordered

breathing. Otolaryngologic Clinics of North America,

40(4):761–784.

Sadeh, A. (2002). The role of actigraphy in sleep medicine.

Sleep Medicine Reviews, 6(2):113–124.

Schefﬂer, M. and Hirt, E. (2004). Wearable devices for

emerging healthcare applications. In Proc. 26th An-

nual International Conference of the IEEE Engineer-

ing in Medicine and Biology Society IEMBS ’04, vol-

ume 2, pages 3301–3304.

SLEEPIC - Developments for a Wearable On-line Sleep and Wake Discrimination System

137