2.2. Participants
To score the sleep quality of the participants, EEG and accelerometer data were recorded for 10 male adults who agreed to participate in the experiments. Their ages were 23.9 ± 2.74 and ranges from 21 to 28. Gender was controlled as male following a study that gender could affect sleep quality [
40]. The manufactured mattress and air cells were more fit for the body size of the average Korean male, so males were recruited priory in this study. When recruiting participants, they were interviewed to determine whether they met the inclusion and exclusion criteria. The inclusion criterion was that the height is within the range of 170 to 175 cm to fit the size of the mattress. The exclusion criterion was that the participant may be at risk of leaving the mattress due to a bad sleeping habit. Participants with poor sleeping habits were excluded through screening interviews, as Tang et al. did [
41]. The average of their BMI scores was 23.3 ± 1.60, and ranges from 20.76 to 25.95. Their Pittsburgh Sleep Quality Index (PSQI) score, which is a self-report questionnaire that assesses sleep quality over a month with a score from 0 to 21, was 5.6 ± 2.22 and the range of their PSQI score was 2 to 10 [
42]. Among the participants, there were 5 poor sleepers with a PSQI score higher than 5. To ensure that the participants’ sleep quality was not affected by other factors, the participants were forbidden from consuming caffeine and alcohol for 48 h before the experiment [
43,
44]. In addition, participants who violate the above policy were excluded from the experiment on the same day through a screening interview before the experiment as Tang et al. did [
45]. Second, participants finished their dinners before 7 p.m., had a shower between 9 and 10 p.m., and went to bed at 11 p.m. Lastly, each participant had a gap of 48 h between experiments to ensure that sleep under one mattress condition did not affect sleep under another mattress condition.
2.4. Procedure
In this study, sleep scoring was not conducted through polysomnography (PSG), but was conducted through the estimation using a single channel EEG according to [
47,
48,
49]. As shown in
Figure 3, four Ag/Ag-Cl electrodes were attached to positions FP1, FP2, A2, and A1 on the participants’ head. The electrodes at positions FP1 and FP2 record the EEG signals; the electrode at position A2 serves as a reference; and the electrode at position A1 is grounded. The two devices used to acquire EEG were BIOPAC MP36 (BIOPAC Systems Inc, Goleta, CA, USA), whose sampling rate was set to 500 Hz, and g.USBamp amplifier (g.tec medical engineering GmbH, Schiedlberg, Austria), whose sampling rate was 600 Hz. In addition, the participants wore wGT3X-BT (ActiGraph LLC, Pensacola, FL, USA) on their wrists such that their movement could be monitored; the sampling rate of the device was 100 Hz.
To avoid the first night effect [
50], all participants slept at the laboratory in the same environment with attaching all electrodes before the actual experiment. The experiment was conducted on three days, and one of three different mattress conditions was considered randomly for each day.
Figure 4 shows the procedure of the experiment. On the first day, all the participants answered the PSQI questionnaire before going to bed. Before the electrodes were attached to their respective positions, their exfoliation was removed such that the maximum impedance was 5 k
[
51]. As shown in
Figure 4, the participant had a 10 min adaption time lying on the mattress [
52,
53]. The participants went to bed at 11 p.m. and slept for 7 h in a room where the temperature and humidity were maintained at 24
C and 50% RH, respectively. After waking up at 6 am, the participants responded to a questionnaire designed to help subjectively evaluate their sleep quality. The questionnaire comprised two questions regarding ‘Sleep Length’ and ‘Sleep Depth’, which are the question number 5 and 8 in the “sleep diary” presented by Åkerstedt et al. [
54]. For more detailed analysis than the original questionnaire, the selection range was widened from 5 to 7, i.e., a range of 1–7.
2.5. Sleep Scoring Algorithm
Subjective sleep assessment was conducted through questionnaires. The length information of each sleep stage was obtained from a hypnogram, as an objective sleep evaluation, where the duration of REM and Deep stages (slow wave sleep, SWS) are the length information of the stages, conventionally used to measure the sleep quality [
55]. REM stage maintains the necessary levels of central nervous system activities, promotes a recovery with providing periodic stimulation to the brain [
56] and preserves emotional memory sources selectively [
57]. SWS induces an endocrine environment that could strongly support the initiation of an adaptive immune response and cleans metabolites [
58]. Additionally, RNR (REM to Non-REM ratio) and SSI (Stage Shift index) were calculated as sleep indices corresponding to the sleep quality. Mendonca et al. reported that the higher RNR and the lower SSI result in the improvement of sleep quality [
55]. RNR is an index representing the ratio of the duration of the REM stage to that of the non-REM stage, as shown in Equation (
1). SSI is an index obtained by dividing the number of times the sleep stage shift during whole sleep by the total sleep time (TST), as shown in Equation (
2).
To estimate the lengths of the sleep stages, an automatic sleep scoring algorithm is employed, which was introduced in a study by Onton et al. [
49,
59]. Onton et al. acquired EEG signals from the FP1-A2 and FP2-A2 channels. The signals of one channel were then classified, in terms of 30-s epoch units, into five sleep stages. This was realized by employing the hidden Markov model (HMM) algorithm along with Viterbi and expectation-maximization (EM) algorithms, in an unsupervised manner [
60]. The hypnograms of the sleep stages were then estimated. In this study, N1 and N2 were merged to form a Light sleep stage and N3 was considered to be a Deep sleep stage, reducing the number of whole stages from five to four.
A single-channel EEG signal was segmented into 30-s epochs, which served as input to the HMM algorithm [
61]; further, the frequency band powers of these epochs were calculated after being filtered using wavelet transform [
49,
62,
63] (Wake: 35–50 Hz, REM: 20–30 Hz, Light: 10.15–15.75 Hz and Deep: 1–3 Hz). There are three parameters for the HMM algorithm, which are an initial probability (
), a transition matrix (
Q), and an emission matrix (
R). Since an initial sleep stage always starts with the participant being awake, referring to the paper by Lo et al. [
64], the initial probability (
) was set as Equation (
3) and the initial transition matrix (
) of the HMM algorithm was initialized as Equation (
4). In Equation (
4), the element
in the matrix represents the probability of transition from the stage
i to the stage
j on the next time step. The emission matrix (
R) comprises the mean (
) and standard deviation matrices (
). The mean of the initial emission matrix (
) was set as Equation (
5). Please note that the probability of one particular frequency band corresponding to a sleep stage (see the diagonal terms in Equation (
5)) is higher than those of the others, which was initial expected based on a study by Onton et al. [
49]. The standard deviation of the initial emission matrix (
) was set as Equation (
6). The parameters (
) were updated using the EM algorithm [
65], after which the Viterbi algorithm provides estimations of the sleep stages based on the maximum posteriori [
66].
To validate the HMM algorithm as an automatic sleep scoring method, we obtained a public data set, Sleep-EDF Database Expanded (Sleep-EDFx), containing two EEG channels (Fpz-Cz and Pz-Oz), an electrooculogram (EOG), and an electromyogram (EMG) recorded at a sampling rate of 100 Hz [
67,
68]. According to the Rechtschaffen and Kales manuals, well-trained technicians manually scored the sleep stages, and thus, the Sleep-EDFx includes all the sleep stages, including wake, REM, S1, S2, S3, and S4. To calculate sleep quality for this data set, S1 and S2 are combined into the light sleep stage, S3 and S4 into Deep stage, reducing the six stages to four: wake, REM, Light, and Deep stages.
The accelerometer data were analyzed using the Sadeh algorithm [
69], implemented in ActiLife6 (ActiGraph LLC, Pensacola, FL, USA) software, which estimates sleep and wake states. Sleep efficiency (SE) could be calculated using the formula,
, based on the estimation of the sleep and wake states. In addition, sleep onset latency (SOL) which is time taken to sleep, wake after sleep onset (WASO) which is total waking time after sleep, and total sleep time (TST) which is total sleep time during time in bed could also be obtained from the actigraphy information [
70].
2.6. Analyses
To investigate how the proposed mattress affects the quality of sleep, a subjective sleep assessment and objective sleep evaluation were conducted. The subjective sleep assessment was conducted with a questionnaire written about the depth and length of sleep experienced by the participants. To analyze the condition of each mattress, comparative analysis among three different mattress conditions was conducted on the subjective sleep assessment under each mattress condition.
Objective sleep evaluation was conducted with various sleep parameters which obtained from actigraphy and EEG analysis. There are two methods for determining the sleep and wake states, which is a binary decision, using actigraphy analysis and EEG analysis as shown in
Figure 5a,b. We can get two indices each that can be obtained with wake information such as SE, TST, SOL, and WASO. In this study, comparative analysis is conducted to see how the above indices are affected by the mattress conditions. A hypnogram was estimated using the HMM algorithm, as shown in
Figure 5b, and sleep parameters such as the length of REM and Deep, RNR, and SSI were obtained from the estimated hypnogram. A comparative analysis was conducted to investigate these sleep parameters depending on the three mattress conditions.
Since the number of participants in our study is 10, which is less than 30 samples, statistical analysis is performed for validation by nonparametric analysis rather than parametric analysis. Since the results of those sleep parameters in each mattress condition, indicate the conclusion of this study, a nonparametric wilcoxon signed-rank test was conducted between the base condition, A condition, and the condition with the highest score among S and SH condition [
71].
There were five poor sleepers with a PSQI score higher than 5, so 10 subjects were divided into 5 poor sleepers and five good sleepers. A comparative analysis of each sleep index is conducted between the two groups to compare and analyze how the customized mattress affects the good sleeper and the poor sleeper.