Instrumentation & Measurement Magazine 24-2 - 34

Table 1 - Sleep Statistics for Each Night (units are in 30 second epoch counts)
TIB
Night 1

TST

936

541

TWT

Stage 1

Stage 2

Stage 3

REM

SE%

395

184

228

100

29

57.8

119

SOL

WASO
208

Night 2

884

542

342

143

208

132

59

61.3

156

330

Night 3

931

665

266

121

328

191

25

71.4

68

213

Night 4

726

600

126

146

247

203

4

82.6

60

161

Night 5

795

534

261

107

190

161

76

67.2

102

158

Night 6

859

614

245

111

115

212

176

71.5

86

297

Average

855.17

582.67

272.5

135.33

219.33

166.5

61.5

68.63

98.5

227.83

74.65

47.99

26.27

63.93

56.30

8.00

32.42

64.78

STD

83.89

sampling rate every five minutes and would be interpolated to
fill each of the 30 second epochs. The sound and BCG devices
both output at a one second sampling rate that would be averaged into 30 second epochs. The IMU-based actigraphy had a
sampling frequency of 100 Hz, which was averaged into the 30
second epochs. As it lacked a time stamp feature, the synchronization was performed manually by clapping near the sound
sensor to line up the acceleration and sound signals.

Experimental Demonstration Results and
Discussion
Initially, the PSG signals were scored manually using the RemlogicTM software for the Embletta X100 PSG system to score
the sleep stages as defined by the AASM guidelines [3]. All
the data processing and software development was done in
Matlab® R2020a (Version 9.8). Table 1 shows the sleep statistics for each of the six nights in counting the 30 second epochs
scored from the PSG. TIB is the total time in bed and accounts
for when the participant lays in bed for the first time to sleep to
the final moment they get up to be awake. TST and TWT are the
total sleep and wake times, respectively, showing how much
sleep and wake was obtained during the night. SE% is the sleep
efficiency percentage by dividing the total sleep time by the total time in bed to observe how well they slept overall. Stages
1-3 are each of the non-REM sleep stages, with stages 1 and
2 being the changeover from wakefulness to light sleep with
reduced HR, respiration, eye and muscle activity before transitioning to deeper sleep; and stage 3 being the deepest sleep
body activity reduced to its lowest levels and the most restful sleep [20]. The REM stage is the rapid eye movement sleep
stage where a lot of eye activity is detected with irregular respiration and HR and most dreaming occurs [20]. SOL is the sleep
onset latency and refers to how long it took to fall asleep upon
first laying in bed. WASO is wake after sleep onset which refers
to the first major awakening after sleep. The resulting sleep

40.09

statistics from the classification algorithm are shown in Table
2 for both the sleep/wake and light/deep sleep with the sleep
stage epochs combined into instances of light and deep sleep.
The Xsens IMU output acceleration data was calculated
with the zero-crossing method to count how many times the
accelerations changed sign from positive to negative or vice
versa to evaluate how much acceleration was happening in the
30 second epoch [21]. The BCG sensor was setup based on the
setting and recommendations from the user manual from Murata and resulting documentation from the clinical trials [15],
[16]. The BCG sensor outputs the HRV through the following
equation, by having LF and HF reversed and included respiration depth (Rdepth) so that large respiratory activities that
affect cardiac activity would not immediately be classified as
an active state [15]:
HRV 

	

HFHRV
	(3)
LFHRV  Rdepth

Therefore, a higher ratio reflects a more restful state and lower
ratio reflects a more active state.
Next, all the sensor output was combined and trained in a
long short-term memory (LSTM) network to sequence deep
learning classification model with Matlab®. The LSTM recurrent neural network (RNN) technique was chosen for the
sleep stage classification because of the memory units that
consider the past output with current inputs with long-term
internal memory [22]. The parallel stacks of a LSTM network
take into account future and past epochs to classify sleep in
appropriately ordered time steps. As all the data gathered are
time series, maintaining order of the sleep cycle is important
as to not distort the natural sleep cycle [22]. The training data
comprised the first five nights and was tested on the final
night to estimate the sleep stages with the added environmental readings. Table 3 summarizes the initial features used
for the model training. Acceleration and actigraphy features

Table 2 - Sleep statistics for classified sleep stages
TIB

TST

TWT

Light and deep

895

545

314

Sleep and wake

895

545

314

34	

Stage 1

Stage 2

Stage 3

98
N/A

REM

447
N/A

N/A

N/A

IEEE Instrumentation & Measurement Magazine	

SE%

SOL

WASO

0.63

131

198

0.63

131

198
April 2021
Instrumentation & Measurement Magazine 24-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-2
