Autonomic cardiorespiratory activity changes across sleep stages. However, it is unknown to what extent it is affected by between- and within-subject variability during sleep. As it is hypothesized that the variability is caused by differences in subject demographics (age, gender, and body mass index), time, and physiology, we quantified these effects and investigated how they limit reliable cardiorespiratory-based sleep staging. Six representative parameters obtained from 165 overnight heartbeat and respiration recordings were analyzed. Multilevel models were used to evaluate the effects evoked by differences in sleep stages, demographics, time, and physiology between and within subjects. Results show that the between- and within-subject effects were found to be significant for each parameter. When adjusted by sleep stages, the effects in physiology between and within subjects explained more than 80% of total variance but the time and demographic effects explained less. If these effects are corrected, profound improvements in sleep staging can be observed. These results indicate that the differences in subject demographics, time, and physiology present significant effects on cardiorespiratory activity during sleep. The primary effects come from the physiological variability between and within subjects, markedly limiting the sleep staging performance. Efforts to diminish these effects will be the main challenge.
Polysomnography (PSG) is the gold standard and common practice for the objective analyses of overnight sleep architecture (displayed by a so-called hypnogram) and sleep-related disorders such as insomnia/parasomnia, sleep-disordered breathing, and rapid-eye-movement (REM) sleep behavior disorder [
Cardiorespiratory activity has been proven to associate with the autonomic sympathetic and parasympathetic (or vagal) nervous systems in humans, which relates to sleep stages [
In addition to sleep stages, the cardiorespiratory activity can be influenced by between-subject variability with respect to
In regard to automatic sleep staging with autonomic cardiorespiratory activity, parameters are usually derived from cardiac and respiratory signals on a 30 s epoch basis [
The aim of this fundamental study was to quantitatively investigate the effects of between- and within-subject variability on cardiorespiratory activity during sleep and to evaluate how they are limited to reliable cardiorespiratory-based sleep staging results.
A total of 165 healthy subjects participating in the SIESTA project [
Subject demographics and sleep statistics (
Mean ± SD | Range | |
---|---|---|
Gender (77 men and 88 women) | ||
Age, y | 51.8 ± 19.4 | 20–95 |
BMI, kg⋅m−2 | 24.6 ± 3.5 | 17.0–35.3 |
Total recording time, h | 7.8 ± 0.5 | 6.0–9.3 |
Wake, % | 22.7 ± 13.2 | 1.2–78.6 |
REM sleep, % | 13.6 ± 5.3 | 0–26.5 |
Light sleep, % | 52.3 ± 10.0 | 15.6–72.1 |
Deep sleep, % | 11.4 ± 6.6 | 0–28.5 |
For each subject, single-night full PSG recordings were obtained. Each recording consists of at least 16 channels including EEG (C3-M2, C4-M1, O1-M2, O2-M1, Fp1-M2, and Fp2-M1), EMG (chin and leg), EOG (2 leads), electrocardiogram (ECG, single-channel, modified V1 lead), nasal airflow, respiratory effort (abdominal and chest wall with respiratory inductance plethysmography), snoring (microphone), and blood oxygen saturation [
Each PSG recording was visually annotated in 30 s epochs as nighttime wake, REM sleep, and one of the NREM sleep stages S1–S4 by two independent raters according to the R&K rules. In case of disagreement, the consensus annotations between the two raters were obtained. For the analysis in this study, we considered four stages: wake, REM sleep, light sleep (merging S1 and S2), and deep sleep or slow wave sleep (merging S3 and S4). Table
The ECG and respiratory effort signals of all subjects were preprocessed before computing the parameters used for analyses. The baseline wander of the ECG signal was removed with a linear phase high-pass filter using an 1.106 s Kaiser window with a 0.8 Hz cutoff frequency and a 30 dB side-lobe attenuation [
The respiratory effort signal was first low-pass-filtered using a 10th order Butterworth filter with a cut-off frequency of 0.6 Hz to eliminate high-frequency noise. Afterwards, the signal baseline was removed by subtracting the median peak-to-trough amplitude estimated over the entire signal. The respiratory peaks and troughs were detected by locating the signal turning points based on sign changes of signal slopes. Finally, we excluded incorrectly detected peaks and troughs
We analyzed six cardiorespiratory (two respiratory and four cardiac) parameters. The respiratory parameters were the mean breathing rate or respiratory frequency (BR) and the standard deviation of breathing rates (SDBR). For cardiac activity, the time-domain parameters included the mean heart rate (HR) and the standard deviation of heartbeat intervals (SDNN). The spectral-domain parameters included the spectral power of heartbeat intervals in the low-frequency band (LF) and the spectral power in the high-frequency band (HF). The LF and HF were normalized by dividing them by the total spectral power minus the power in the very-low-frequency (VLF, 0.003–0.04 Hz) band [
All the six parameters have been widely used for the task of cardiorespiratory-based sleep staging [
Values of the cardiorespiratory parameters (mean ± SD) measured from subjects with different demographics (gender, age, and BMI) and time of night are presented. We considered different cohort sets including three age groups: young (20–39 y), middle (40–69 y), and elderly (>69 y), and three BMI groups: underweight (<18.5 kg·m−2), normal weight (18.5–25 kg·m−2), and overweight (>25 kg·m−2). In addition, total sleep time was divided into four periods: 0–90 min, 90–180 min, 180–270 min, and >270 min. Significance of difference between groups was tested with the analysis of variance (ANOVA)
Traditional statistical methods such as repeated measures ANOVA (rANOVA), repeated measures multivariate ANOVA (rMANOVA), and multiple regression analysis (MRA) are often used to analyze longitudinal data. However, they might not be appropriate since they expect uncorrelated and independent observations or they cannot model variables in different levels [
Due to the presence of its advantages, multilevel analysis has been widely deployed in many areas such as psychophysiology [
On the one hand the between-subject variability effects of cardiorespiratory activity can be linked to physiology and subject demographics (age, gender, and BMI). On the other hand, cardiorespiratory activity may change depending on the time of night within subjects [
To evaluate the between- and within-subject effects, we constructed a multilevel model with two levels (level two: subject; level one: time or epoch) for a given cardiorespiratory parameter Model #1:
in which
Intuitively, the mean value of a specific cardiorespiratory parameter over the entire night may differ from subject to subject, which might be due to the physiological variability between subjects at the general mean level. Cronbach [ Model #2:
where
The multilevel modeling was implemented using the MLwiN software (Centre for Multilevel Modeling, the University of Bristol, UK), where an iterated generalized least square (IGLS) algorithm is issued for the model estimation, that is, the estimates of regression coefficients and their variances [
A Wald
The models described in (
It is of particular interest in interpreting how much the model variance is explained by different variables or effects. As described in Table
Description of the seven explanatory effects (with exclusion of sleep stage effects) on cardiorespiratory activity considered in this study.
Effect | Description |
---|---|
Overall between-subject effect | |
Demographic effect | Fixed effect, variability in age, gender, and BMI between subjects |
Centering (physiological) effect | Fixed effect, variability in overnight mean level between subjects |
Between-subject time effect | Random effect, variability in time of night between subjects |
Between-subject physiological effect | Random effect, variability in physiology between subjects |
Overall within-subject effect | |
Within-subject time effect | Fixed effect, variability in time of night within subjects |
Within-subject physiological effect | Random effect, variability in physiology within subjects |
Cross-interaction effect | |
Demographic-related time effect | Fixed effect, demographic-related variability in time of night |
Linear discriminant (LD) has been shown to be an appropriate algorithm in classifying overnight sleep stages based on cardiorespiratory activity in many studies [
The objective was to examine how much the between- and within-subject effects on the cardiorespiratory activity would restrict the performance in classifying sleep stages (wake, REM sleep, light sleep, and deep sleep) and then estimating the sleep statistics. For comparison, we analyzed three different “correction” schemes (CS) based on the optimized Model #2 with estimated model coefficients to correct (or predict) the values for each parameter. The corrected values were then used to perform sleep staging. The sleep staging using the original measured values without any corrections served as the baseline scheme (BS). The first correction scheme (CS1) predicts the parameter values with subtraction of all the fixed effects independent of sleep stages, such that CS1: The second correction scheme (CS2) corrects the parameter values by subtracting all the (sleep stage independent) fixed effects and all the between-subject random effects, such that CS2: The third correction scheme (CS3) excludes all the (sleep stage independent) fixed effects and the within-subject effect to correct the parameter values, such that CS3:
Note that, again, the exclusive aim of analyzing these correction schemes in the present study was to evaluate in what aspect and how far the cardiorespiratory parameters can be improved for sleep staging instead of really performing sleep staging. In other words, we intended to answer the question, what sleep staging performance can be achieved if we can eliminate the effects caused by the between- or within-subjects variability? Investigating methods of estimating the fixed coefficients and random variances without knowing sleep stages was not addressed in this study.
Figure
Values (mean ± SD) of the six cardiorespiratory parameters in different cohort sets.
Cohort set | Respiratory parameters | Cardiac parameters | ||||
---|---|---|---|---|---|---|
( |
BR, ln-Hz | SDBR, ln-Hz | HR, ln-bpm | SDNN, ln-ms | LF, nu | HF, nu |
Gender | ||||||
Man | −1.20 ± 0.24 | −3.67 ± 0.75 | 4.13 ± 0.15 | 3.74 ± 0.77 | 0.42 ± 0.23 | 0.47 ± 0.23 |
Woman | −1.22 ± 0.23 | −3.81 ± 0.76 | 4.16 ± 0.16 | 3.49 ± 0.71 | 0.39 ± 0.22 | 0.50 ± 0.23 |
|
||||||
Age | ||||||
Young | −1.24 ± 0.24 | −3.85 ± 0.74 | 4.11 ± 0.16 | 3.94 ± 0.63 | 0.36 ± 0.20 | 0.56 ± 0.22 |
Middle | −1.20 ± 0.24 | −3.71 ± 0.78 | 4.15 ± 0.16 | 3.52 ± 0.69 | 0.45 ± 0.23 | 0.45 ± 0.23 |
Elderly | −1.18 ± 0.20 | −3.70 ± 0.71 | 4.17 ± 0.13 | 3.39 ± 0.81 | 0.38 ± 0.24 | 0.45 ± 0.22 |
|
||||||
BMI | ||||||
Underweight | −1.24 ± 0.14 | −4.00 ± 0.66 | 4.11 ± 0.12 | 4.01 ± 0.53 | 0.36 ± 0.18 | 0.56 ± 0.19 |
Normal | −1.23 ± 0.23 | −3.77 ± 0.74 | 4.14 ± 0.16 | 3.72 ± 0.73 | 0.41 ± 0.22 | 0.48 ± 0.23 |
Overweight | −1.18 ± 0.24 | −3.70 ± 0.77 | 4.15 ± 0.15 | 3.46 ± 0.75 | 0.39 ± 0.23 | 0.48 ± 0.23 |
|
||||||
Time of night | ||||||
0–90 min | −1.22 ± 0.22 | −3.81 ± 0.80 | 4.16 ± 0.15 | 3.52 ± 0.73 | 0.39 ± 0.22 | 0.50 ± 0.23 |
90–180 min | −1.21 ± 0.22 | −3.85 ± 0.75 | 4.17 ± 0.15 | 3.58 ± 0.74 | 0.42 ± 0.23 | 0.46 ± 0.23 |
180–270 min | −1.20 ± 0.23 | −3.77 ± 0.77 | 4.15 ± 0.16 | 3.61 ± 0.77 | 0.41 ± 0.23 | 0.48 ± 0.23 |
>270 min | −1.21 ± 0.24 | −3.66 ± 0.72 | 4.12 ± 0.15 | 3.67 ± 0.75 | 0.40 ± 0.22 | 0.49 ± 0.23 |
|
||||||
Sleep stage | ||||||
Wake | −1.16 ± 0.23 | −3.25 ± 0.62 | 4.19 ± 0.15 | 3.61 ± 0.78 | 0.42 ± 0.24 | 0.44 ± 0.23 |
REM sleep | −1.18 ± 0.22 | −3.44 ± 0.52 | 4.15 ± 0.16 | 3.64 ± 0.76 | 0.45 ± 0.23 | 0.42 ± 0.22 |
Light sleep | −1.23 ± 0.23 | −3.89 ± 0.73 | 4.13 ± 0.15 | 3.64 ± 0.73 | 0.40 ± 0.22 | 0.49 ± 0.23 |
Deep sleep | −1.24 ± 0.23 | −4.29 ± 0.71 | 4.14 ± 0.15 | 3.45 ± 0.72 | 0.33 ± 0.21 | 0.57 ± 0.21 |
Note: ln, natural logarithm; nu, normalized unit; young, 20–39 y; middle, 40–69 y; elderly, >69 y; underweight, <18.5 kg⋅m−2; normal weight, 18.5–25 kg⋅m−2; overweight, >25 kg⋅m−2; light sleep, S1 and S2 stages; deep sleep, S3 and S4 stages. For all the parameters, values between each cohort group were significantly different (
Skewness comparison of cardiorespiratory parameters with and without natural logarithm transformation, indicating that BR, SDBR, HR, and SDNN should be log-transformed.
In comparison with the
Coefficients and their standard errors (SE) of the optimized multilevel model without the between-subject centering effect (Model #1) for the six cardiorespiratory parameters analyzed in this study.
Model coef. | Respiratory parameters | Cardiac parameters | ||||
---|---|---|---|---|---|---|
BR, ln-Hz | SDBR, ln-Hz | HR, ln-bpm | SDNN, ln-ms | LF, nu | HF, nu | |
Fixed | Coefficient (SE) | |||||
|
−1.458 (0.087) | −3.320 (0.032) | 4.221 (0.016) | 4.823 (0.255) | 0.464 (0.014) | 0.535 (0.027) |
|
Baseline | Baseline | Baseline | Baseline | Baseline | Baseline |
|
0.002 (0.008)NS | −0.205 (0.026) | −0.028 (0.001) | −0.104 (0.027) | 0.030 (0.007) | −0.037 (0.007) |
|
−0.035 (0.008) | −0.611 (0.026) | −0.061 (0.001) | −0.052 (0.021) | −0.027 (0.006) | 0.039 (0.006) |
|
−0.044 (0.010) | −0.997 (0.033) | −0.055 (0.001) | −0.249 (0.026) | −0.096 (0.008) | 0.106 (0.008) |
|
−0.009 (0.002) | −0.002 (0.001) | ||||
|
0.042 (0.021) | −0.247 (0.069) | −0.045 (0.018) | 0.052 (0.017) | ||
|
0.011 (0.004) | −0.025 (0.011) | ||||
|
0.001 (0.0004) | −0.0001 (0.2 |
0.001 (0.0002) | 0.0004 (0.0001) | −0.0004 (0.0001) | |
|
−1.0 |
|||||
|
||||||
|
−2.8 |
−1.7 |
1.7 | |||
|
||||||
Random | Coefficient (SE) | |||||
Ω0 | ||||||
Ωwake | 0.030 (0.003) | 0.159 (0.018) | 0.018 (0.002) | 0.224 (0.025) | 0.018 (0.002) | 0.014 (0.002) |
ΩREM | 0.029 (0.003) | 0.171 (0.020) | 0.019 (0.002) | 0.280 (0.031) | 0.022 (0.002) | 0.018 (0.002) |
Ωlight | 0.030 (0.003) | 0.219 (0.024) | 0.020 (0.002) | 0.256 (0.028) | 0.019 (0.002) | 0.017 (0.002) |
Ωdeep | 0.031 (0.003) | 0.257 (0.029) | 0.020 (0.002) | 0.324 (0.036) | 0.020 (0.002) | 0.017 (0.002) |
|
1.2 |
6.6 |
3.5 |
7.2 |
5.0 |
4.6 |
|
||||||
Residual | ||||||
|
0.019 (0.0001) | 0.290 (0.001) | 0.003 (0.00001) | 0.230 (0.001) | 0.033 (0.0001) | 0.033 (0.0001) |
|
||||||
Deviance | −150487 | 217253 | −398075 | 186380 | −75029 | −74306 |
Note: ln, natural logarithm; nu, normalized unit; NS, not significant. The statistically significant effects (Wald
Most of the analyzed parameters were found to be time-variant (i.e., they were modulated by time of night) with an exception of breathing rate (Table
Anexample of multilevel regressions of the six cardiorespiratory parameters for a man (age: 24 y, BMI: 21.3 kg⋅m−2) and a woman (age: 70 y, BMI: 28.6 kg⋅m−2) using coefficients estimated through Model #1 excluding the random coefficients and residual term. The regression variables included age, gender, BMI, time, and time × age, time × gender, time × BMI, and sleep stages: wake, REM sleep, light sleep, and deep sleep.
With the addition of the centering variable to Model #1, we have Model #2 and the estimated regression coefficients after model optimization (Wald
Coefficients and their standard errors (SE) of the optimized multilevel model with the additional between-subject centering effect (Model #2) for the six cardiorespiratory parameters analyzed in this study.
Model coef. | Respiratory parameters | Cardiac parameters | ||||
---|---|---|---|---|---|---|
BR, ln-Hz | SDBR, ln-Hz | HR, ln-bpm | SDNN, ln-ms | LF, nu | HF, nu | |
Fixed | Coefficient (SE) | |||||
|
−0.098 (0.079)NS | −0.012 (0.017)NS | 0.104 (0.028) | −0.060 (0.047)NS | −0.018 (0.034)NS | 0.131 (0.030) |
|
0.973 (0.011) | 0.884 (0.020) | 0.993 (0.007) | 0.979 (0.011) | 0.936 (0.012) | 0.923 (0.011) |
|
Baseline | Baseline | Baseline | Baseline | Baseline | Baseline |
|
0.002 (0.008)NS | −0.199 (0.025) | −0.027 (0.004) | −0.104 (0.027) | 0.030 (0.007) | −0.037 (0.007) |
|
−0.035 (0.008) | −0.606 (0.026) | −0.062 (0.004) | −0.052 (0.020) | −0.027 (0.005) | 0.039 (0.006) |
|
−0.044 (0.010) | −0.992 (0.033) | −0.054 (0.004) | −0.248 (0.026) | −0.096 (0.008) | 0.105 (0.008) |
|
−0.002 (0.001) | −0.0001 (0.5 |
0.0004 (0.0001) | 0.0002 (0.0001) | ||
|
−0.024 (0.012) | |||||
|
0.005 (0.001) | −0.004 (0.001) | ||||
|
0.0003 (0.0001) | −0.0001 (0.2 |
0.001 (0.0001) | 0.0004 (0.0001) | −0.0004 (0.0001) | |
|
−1.0 |
|||||
|
0.0001 (0.5 |
|||||
|
−1.8 |
1.7 | ||||
|
||||||
Random | Coefficient (SE) | |||||
Ω0 | ||||||
Ωwake | 0.012 (0.001) | 0.093 (0.011) | 0.004 (0.0004) | 0.094 (0.011) | 0.006 (0.001) | 0.005 (0.001) |
ΩREM | 0.014 (0.002) | 0.099 (0.011) | 0.003 (0.0003) | 0.095 (0.011) | 0.007 (0.001) | 0.006 (0.001) |
Ωlight | 0.006 (0.001) | 0.061 (0.007) | 0.002 (0.0003) | 0.044 (0.005) | 0.004 (0.001) | 0.003 (0.0004) |
Ωdeep | 0.010 (0.001) | 0.131 (0.015) | 0.003 (0.0003) | 0.087 (0.010) | 0.006 (0.001) | 0.006 (0.001) |
|
1.1 |
6.7 |
3.5 |
7.1 |
4.8 |
4.3 |
|
||||||
Residual | ||||||
|
0.019 (0.0001) | 0.290 (0.001) | 0.003 (0.00001) | 0.230 (0.001) | 0.033 (0.0001) | 0.033 (0.0001) |
|
||||||
Deviance | −151084 | 216873 | −398866 | 185774 | −75617 | −74903 |
Note: ln, natural logarithm; nu, normalized unit; NS, not significant. The statistically significant effects (Wald
Normality of the variances was tested and suggested using the Q-Q plot method for all models. For example, the Q-Q plots of the residual variances
Q-Q plots of residual variance
To discover which effects explained the variance and how much each constituted we computed for each cardiorespiratory parameter the PVE for each effect by analyzing the estimated variances of random intercept and residual in a sequence of models (Models A–G in the Appendix). The variance changes in the models with the inclusion of different effects in a specific order are shown in Table
Variances of a sequence of models (Models A–G in the appendix) with different effects for computing their PVE for the six cardiorespiratory parameters analyzed in this study.
Models A–G with different effects |
Respiratory parameters | Cardiac parameters | ||||
---|---|---|---|---|---|---|
BR, ln-Hz | SDBR, ln-Hz | HR, ln-bpm | SDNN, ln-ms | LF, nu | HF, nu | |
Model A: |
||||||
|
0.0229 | 0.3306 | 0.0043 | 0.2626 | 0.0354 | 0.0356 |
|
0.0328 | 0.1389 | 0.0192 | 0.2997 | 0.0151 | 0.0156 |
Dev | −125045 | 232926 | −348717 | 202249 | −66487 | −65952 |
|
||||||
Model B: Model A + within-subject time effect (fixed) | ||||||
|
0.0228 | 0.3284 | 0.0040 | 0.2600 | 0.0353 | 0.0355 |
|
0.0328 | 0.1393 | 0.0191 | 0.2999 | 0.0150 | 0.0155 |
Dev | −125109 | 232056 | −357783 | 200926 | −66724 | −66131 |
|
||||||
Model C: Model B + demographic effect (fixed) | ||||||
|
0.0228 | 0.3284 | 0.0040 | 0.2600 | 0.0353 | 0.0355 |
|
0.0308 | 0.1329 | 0.0183 | 0.2230 | 0.0147 | 0.0136 |
Dev | −125120 | 232048 | −357790 | 200877 | −66730 | −66152 |
|
||||||
Model D: Model C + centering effect (fixed) | ||||||
|
0.0228 | 0.3284 | 0.0040 | 0.2600 | 0.0353 | 0.0355 |
|
0.0001 | 0.0098 | 0.0001 | 0.0033 | 0.0003 | 0.0002 |
Dev | −126064 | 231624 | −358718 | 200200 | −67367 | −66850 |
|
||||||
Model E: Model D + demographic-related time effect (fixed) | ||||||
|
0.0227 | 0.3284 | 0.0040 | 0.2597 | 0.0352 | 0.0354 |
|
0.0001 | 0.0098 | 0.0001 | 0.0033 | 0.0003 | 0.0002 |
Dev | −126393 | 231624Ne | −358718Ne | 200027 | −67718 | −67206 |
|
||||||
Model F: Model E + between-subject time effect (random) | ||||||
|
0.0210 | 0.3157 | 0.0034 | 0.2476 | 0.0343 | 0.0346 |
|
0.0003 | 0.0097 | 0.0001 | 0.0041 | 0.0003 | 0.0002 |
|
1.1 |
7.3 |
3.6 |
7.1 |
5.2 |
4.5 |
Dev | −136185 | 226933 | −380964 | 194316 | −70913 | −69899 |
|
||||||
Model G: Model F + between-subject physiological effect (random) | ||||||
|
0.0186 | 0.2896 | 0.0029 | 0.2298 | 0.0328 | 0.033 |
|
0 | 0 | 0 | 0 | 0 | 0 |
|
1.0 |
6.7 |
3.5 |
7.1 |
4.8 |
4.3 |
Dev | −151084 | 216874 | −398866 | 185774 | −75617 | −74903 |
Note: ln, natural logarithm; nu, normalized unit; Dev, model deviance; Ne, no effect. All the models include fixed (
Proportion of variance explained (PVE, %) accounted for by different effects for the six cardiorespiratory parameters analyzed in this study.
Effect | Respiratory parameters | Cardiac parameters | ||||
---|---|---|---|---|---|---|
BR | SDBR | HR | SDNN | LF | HF | |
Overall between-subject effect | ||||||
Demographic effect | 3.55% | 1.37% | 3.36% | 13.69% | 0.63% | 3.70% |
Centering (physiological) effect | 55.26% | 26.23% | 77.95% | 39.06% | 28.63% | 26.41% |
Between-subject time effect | 2.74% | 2.72% | 2.67% | 2.00% | 1.87% | 1.58% |
Between-subject physiological effect | 5.03% | 7.62% | 2.27% | 3.91% | 3.49% | 3.44% |
Overall within-subject effect | ||||||
Within-subject time effect | 0.01% | 0.37% | 1.32% | 0.42% | 0.16% | 0.14% |
Within-subject physiological effect | 33.39% | 61.69% | 12.43% | 40.87% | 65.04% | 64.54% |
Cross-interaction effect | ||||||
Demographic-related time effect | 0.02% | Ne | Ne | 0.06% | 0.18% | 0.19% |
Note: ln, natural logarithm; Ne, no effect. For each cardiorespiratory parameter, the sum of PVEs from all the effects is 100%, representing the total variance for that parameter. The centering effect reflected some between-subject physiological variability (at the overnight mean level) that was assumed to be independent of sleep stage composition over the entire night.
Specifically, a relative larger percentage (13.7%) of the demographic effect can be found on SDNN compared with the other parameters. The PVE of between-subject physiological variability (in the random part) ranged from 2.27% to 7.62% depending on the parameters. For the time effect, the PVE in the fixed part (0.01–1.32%) reflecting the linear changes of parameters over time within subjects was smaller than in the random part (1.58–2.74%) with the indication of different changes over time between subjects. In general, the time effect accounted for much less of the total variance than most other effects. Finally, although the cross-interactions existed between time and demographics for BR, SDNN, LF, and HF, the proportion of variance they explained was very small (<0.20%).
The results of sleep staging are presented in Table
Comparison of sleep staging results (wake/REM sleep/light sleep/deep sleep) using different schemes in correcting the cardiorespiratory parameters.
PSG | BS | CS1 | CS2 | CS3 | |
---|---|---|---|---|---|
Overall performance | |||||
Accuracy, % | — | 55.8 ± 9.8 | 60.4 ± 8.8 | 62.9 ± 7.8 | 83.5 ± 14.4 |
Kappa coefficient | — | 0.19 ± 0.10 | 0.29 ± 0.11 | 0.35 ± 0.09 | 0.72 ± 0.23 |
Sleep stage composition (percentage) | |||||
Wake, % | 19.8 ± 12.5 | 19.9 ± 14.4 | 18.4 ± 4.9 | 20.6 ± 6.4 | 19.7 ± 10.7 |
REM sleep, % | 14.0 ± 5.6 | 0.7 ± 1.0 | 2.4 ± 2.0 | 3.0 ± 1.7 | 10.5 ± 7.8 |
Light sleep, % | 53.4 ± 10.7 | 74.7 ± 15.1 | 73.5 ± 8.1 | 71.0 ± 8.2 | 59.9 ± 12.0 |
Deep sleep, % | 12.8 ± 7.2 | 4.7 ± 5.6 | 5.7 ± 5.2 | 5.4 ± 4.0 | 9.9 ± 7.6 |
Note: BS, baseline with original parameter values without correction; CS1, with correction by fixed effects; CS2, with correction by fixed effects and between-subject random effects; CS3, with correction by fixed effects and within-subject random effect (model residual). For CS2 and CS3, results were obtained when assuming the sleep stages were known, which was usually not the case in practice. For accuracy and Kappa coefficient, significance of difference between using each correction scheme and BS was confirmed with a paired (two-sided) Wilcoxon signed-rank test, all at
The results of demographic and time of night effects found in this study are consistent with the findings reported in previous work [
It should be noted that the model used to facilitate the interpretation of the demographic effects (Model #1) should not include the (between-subject) centering variable. This is because the demographic differences usually correspond to the autonomic changes at the overnight mean level. Due to the inclusion of the centering effect in Model #2, it came as a surprise that some demographic variables still had significant effects (see Table
It is important to note that since some effects were correlated with each other, the order in the procedure of constructing the sequence of models (see the Appendix) must be specifically determined. This aimed at precisely quantifying the proportion of variance explained by each effect. The procedure should follow the way that the model with fixed effects (e.g., demographic effects) that are explainable by other effects should be first addressed and the model with random effects should be included later [
In Tables
Regarding the quantified between- and within-subject effects, they were found to be statistically significant and they explained a relatively large portion of the total variance as we expected. In fact, several factors in addition to internal physiology may also explain some of the total variance within subjects in cardiorespiratory activity such as body movements, body position, sleep environment, conscious breathing control, and even daytime activity. However, we did not answer which of these effects takes place in this work and this should be studied in the future.
When evaluating the performance of sleep staging using the cardiorespiratory parameters, Model #2 should be regarded as the preference. For each parameter, although the estimate of its overnight mean value for each subject was not completely accurate (due to the difference of sleep stage composition between subjects), correcting it can still result in a reduction of the physiological variability between subjects to a great extent. As a consequence, the sleep staging results can be improved. Table
Table
The sequence of models constructed to compute the PVE values for different effects is described in as follows.
(i) The first model is the model with solely the constant and random intercepts as well as the fixed sleep stage dependent variables. This baseline model can be written as Model A:
where
(ii) Let us then consider the model with fixed time effect at the first level Model B:
For the variance analysis of the time variable, instead of using the original time stamps mentioned before (i.e.,
Now we consider the subject-level fixed effects.
(iii) The model including demographic variables is as follows: Model C:
Similarly, the PVE explained by the between-subject demographic variables can be computed by
(iv) Further, Model D is the model with the inclusion of between-subject centering effect (expressing the physiological difference between subjects at the overnight mean level), given by Model D:
from which the corresponding PVE is computed such that
(v) For the inclusion with cross-interactions that express the demographic-related time effects, the model is as follows: Model E:
and the proportion of cross-interaction variance is
In addition to the fixed part, we consider the random part of some effects.
(vi) The model with additional random time effect is as follows: Model F:
The computation of the PVE accounted for by the random time effect can be accordingly obtained by
(vii) Afterwards, the model with random effects for different sleep stages (expressing the between-subject physiological variability associated with each sleep stage in random part) is then expressed as Model G:
In this model, the random variance
(viii) Finally, the remaining residual variance is assumed to only associate with the physiological variability within subjects and its proportion can be obtained such that
Note that all these models are optimized by only keeping the variables that do not statistically equal zero.
No conflict of interests, financial or otherwise, is declared by the authors.
The authors gratefully thank Tine Smits and Dr. Sam Jelfs from Philips Research for their insightful comments and proofreading of the paper. The SIESTA database used in the present study was supported by the European Commission, DG XII (Project no. Biomed-2 BMH4-CT97-2040) between Sep. 1997 and Aug. 2000.