Study on the objective assessment of sleep disturbance due to

environmental noise by wearable devices

Makoto Morinaga 1 Kanagawa University 3-27-1 Rokkakubashi, Kanagawa Ward, Yokohama City, Kanagawa 221-8686, Japan Chikashi Takara Nakamura Clinic 4-2-1 Iso, Urasoe City, Okinawa 901-2132, Japan Yosiaki Sasazawa University of Ryukyus 4-2-1 Iso, Urasoe City, Okinawa 901-2132, Japan Hiroshi Nakamura Nakamura Clinic 4-2-1 Iso, Urasoe City, Okinawa 901-2132, Japan

ABSTRACT Introducing noise standards or limits at night is necessary to prevent sleep disturbance due to environmental noise, such as transportation noise. In general, noise standards should be based on robust exposure–response relationships derived from a large amount of human response data, and subjective evaluation obtained from large-scale socio-acoustic surveys can provide reliable scientific evidence. However, it is also necessary to investigate an easy method for collecting objective data on sleep effects with little cost and effort. In the present study, we conducted a laboratory study to investigate the accuracy of wearable devices for estimating sleep stages. Participants wore a smartwatch capable of photoplethysmography, and were also monitored by electrocardiography and polysomnography simultaneously. Agreement of sleep stages between the estimations by photoplethysmography and polysomnography was moderate (κ = 0.49). Furthermore, we estimated the sleep stages by applying a random-forest algorithm to the electrocardiography data, and those estimations agreed well with those from polysomnography (κ = 0.79). The results suggest that sleep stages such as waking, rapid eye movement (REM), and non- REM are easily estimated with a certain level of accuracy by these wearable devices.

1. INTRODUCTION

Many studies have reported that physiological stress responses due to environmental noise increase the risk of adverse health effects [1]. The World Health Organization Regional Office for Europe has provided environmental noise guidelines [2] and recommended noise levels to prevent high annoyance and sleep disturbance. Meanwhile, it has been several decades since the environmental

1 m-morinaga@kanagawa-u.ac.jp

Sia inter noice 21-24 AUGUST SCOTTISH EVENT CAMPUS GLASGOW

noise standards for traffic noise in Japan were established, and they need to be reviewed. In general, noise standards should be based on robust exposure–response relationships derived from a large amount of human response data, and subjective evaluation obtained from large-scale socio-acoustic surveys can provide reliable scientific evidence. The exposure–response relationship for noise in Japan has been studied for a long time [3], but most of the data are from surveys conducted around 2000, and the lack of recent data is a significant issue, especially for aircraft noise. Also, those social surveys were conducted by means of self-reported questionnaires, but for sleep disturbance, an objective method of assessing noise effects is also required. Polysomnography (PSG) is the gold standard for evaluating sleep stages, but it requires expensive medical devices and the judgment of an expert who is familiar with using such equipment. Also, the considerable burden on study participants of wearing many electrodes makes it difficult to collect the large amount of data required for studying environmental standards. Basner et al. [4] developed an original algorithm for estimating sleep stages from heartbeats using portable electrocardiography (ECG), and they used the algorithm to conduct a field study in the United States [5]. Furthermore, in recent years, some studies have used machine learning to estimate sleep stages based on heart rate [6, 7].

To investigate a method for collecting sleep data objectively at low cost and with little effort, in the present study we conducted a laboratory study to investigate how accurately wearable devices can estimate sleep stages. We examined methods for evaluating the effects of environmental noise on sleep using (i) a smartwatch capable of photoplethysmography (PPG) and (ii) a portable ECG device. 2. EXPERIMENTAL METHOD

A laboratory experiment was conducted using PSG, PPG, and ECG simultaneously to investigate how the sleep stages estimated by the wearable devices matched those estimated by PSG. The experimental method was reviewed and approved by the Ethical Review Committee for Research Involving Human Subjects at Kanagawa University.

2.1. Experimental room The experiment was conducted between September and November 2021 on a hospital bed in the Nakamura Clinic in Urasoe City, Okinawa Prefecture, Japan. The hospital bedroom had a general interior so that the experimental participants could sleep as they usually would, and the windows were double-glazed and adequately sound-insulated.

2.2. Equipment for estimating sleep stages Each participant slept for 1 night while wearing the devices for PSG, PPG (Fitbit Sense; Fitbit, USA), and ECG (Faros 180; Bittium, Finland) simultaneously.

The PPG used in this experiment was provided by a smartwatch with a unique algorithm for estimating four sleep stages, namely, Wake, REM, Light, and Deep, with 30 s as one epoch. The results were compared with those from PSG, but because PSG evaluates five sleep stages, namely, Wake, REM, N1, N2, and N3, Light in PPG was analyzed as being equivalent to N1+N2 in PSG.

The portable ECG device recorded heartbeats from time to time, and heart rate variability (HRV) analysis values were extracted for a 5-min period centered on the same epoch as PSG and PPG. In total, 46 indices were extracted in the time domain, frequency domain, and nonlinear parameters using HRV analysis software (Kubios HRV Premium). Furthermore, machine learning was performed using those analyzed values as input parameters, and the results were compared with those from PSG. A random-forest algorithm was used for the machine learning, with the PSG results used as the correct labels; 10-fold cross-validation was performed to validate the prediction accuracy on the test data that was not used for the training.

2.3. Participants Patients undergoing snoring treatment were invited to participate in the experiment. A request form explaining the experiment and its purpose was prepared in advance, and all participants signed a consent form for participation in the experiment. Seventeen participants consented to the study, and the experiment was carried out. However, for some of the participants, data could not be measured

correctly because of loosening of the equipment; therefore, excluding the data from those participants, the data of 9 men and 1 woman were used for the PPG analysis, and those from 13 men and 1 woman were used for the EEG analysis. The sex and age distributions of the participants are given in Tables 1 and 2. The participants were between 21 and 55 years old and were paid an honorarium of 5000 JPY after the experiment.

Tabl e 1: Sex and age distributions of participants in the PPG ana lysis.

20–29 30–39 40–49 50–59

Female 1 0 0 0

Male 3 1 0 5

Tabl e 2: Sex and age distributions of participants in the ECG ana lysis.

20–29 30–39 40–49 50–59

Female 1 0 0 0

Male 3 2 1 7

3. RESULTS AND DISCUSSION

3.1. Comparison between PPG and PSG

Table 3 compares the sleep stages estimated by PPG and PSG; the metrics therein are defined as follows:

!"#!$

Accuracy =

!"#%$#%"#!$ , (1)

!$

Recall =

%"#!$ , (2)

!$

Precision =

%$#!$ , (3)

&($()*+,+-.×0)*122)

F– measure =

$()*+,+-.#0)*122 . (4)

Here, TN , TP , FN , and FP indicate true negative, true positive, false negative, and false positive, respectively. These results are based on data from 10 participants, with κ ≈ 0.5, indicating that on average the PSG and PPG estimates show moderate agreement.

Table 4 gives the agreement of three sleep stages between PPG and PSG, and for comparison Table 5 gives the results of a previous study that estimated three sleep stages by ECG [7]. For the present study, it is found that the accuracy and the mean of the F-measure are higher than those of the previous study, and the κ value is similar.

3.2. Comparison between ECG and PSG Indices obtained from the HRV software include (i) time-domain analysis values such as the mean and standard deviation of the R-R interval, (ii) frequency-domain analysis values including LF/HF, (iii) values calculated by Poincaré plot [8], and (iv) Kubios HRV indices [9] which can be introduced by combining the preceding values. The values of these indices calculated every epoch (30 s) were used for the random-forest machine learning, with the PSG result as the correct answer. The results of the 10-fold cross-validation are given in Tables 6 and 7.

Table 3 : Agreement of estimation of sleep stages by PSG and PPG (four categories).

Estimated by PPG

Total Recall Wake REM Light Deep

Wake 498 168 505 7 1178 0.42

REM 38 1094 236 68 1436 0.76

PSG

N1+N2 382 628 4290 412 5712 0.75

N3 17 54 596 799 1466 0.55

Total 935 1944 5627 1286 9792 -

Precision 0.53 0.56 0.76 0.62 - -

F-measure 0.47 0.65 0.76 0.58 - -

Accuracy 0.68

κ 0.49

Table 4: Agreement of estimation of sleep stages by PSG and PPG (three categories).

Estimated by PPG

Total Recall Wake REM Non-REM

Wake 498 168 512 1178 0.42

PSG

REM 38 1094 304 1436 0.76

Non-REM 399 682 6097 7178 0.85

Total 935 1944 6913 9792 -

Precision 0.53 0.56 0.88 - -

F-measure 0.47 0.65 0.87 - -

Accuracy 0.79

κ 0.47

Table 5: Agre ement in previous study [7]: three categories estim ated by ECG.

Accuracy 0.72–0.74

Average of F-measure 0.45–0.61

κ 0.24–0.48

Table 6: Agreement of estimation of sleep stages by PSG and ECG (five categories).

Estimated by ECG (HRV analysis)

Total Recall Wake REM N1 N2 N3

Wake 109 3 19 6 0 137 0.80

REM 4 182 4 10 0 200 0.91

PSG

N1 18 3 33 14 0 68 0.49

N2 22 18 52 628 35 755 0.83

N3 0 0 0 18 140 158 0.89

Total 153 206 108 676 175 1318 -

Precision 0.71 0.88 0.31 0.93 0.80 - -

F-measure 0.75 0.90 0.38 0.88 0.84 - -

Accuracy 0.83

κ 0.79

Table 7: Agreement of estimation of sleep stages by PSG and ECG (three categories).

Estimated by ECG (HRV analysis)

Total Recall Wake REM Non-REM

Wake 109 3 25 137 0.80

PSG

REM 4 182 14 200 0.91

Non-REM 40 21 920 981 0.94

Total 153 206 959 1318 -

Precision 0.71 0.88 0.96 - -

F-measure 0.75 0.90 0.95 - -

Accuracy 0.92

κ 0.78

5. CONCLUDING REMARKS

To investigate a simple and objective method for evaluating sleep stages, we conducted a laboratory study to examine methods for evaluating the effects of environmental noise on sleep using (i) a smartwatch capable of PPG and (ii) a portable ECG device. Estimations with high or moderate accuracy were obtained with both methods, and it is suggested that they can be used in field studies to compare the effects on sleep in noisy and control areas. By conducting field studies over a period of days to weeks using devices that can record changes in sleep stages and simultaneously measuring noise events and level, it could be possible to establish an exposure-response relationship for changes in noise levels and sleep stages.

A limitation of the present study was that the PPG method was less accurate for older participants and those who experienced many sleep-apnea events. Also, although the estimation accuracy of applying the random-forest algorithm to the ECG data was extremely high, the portable ECG device is cumbersome to wear compared to a smartwatch and may not be suitable for long- term field studies. Finally, it is also necessary to test whether the estimation accuracy remains for new participants from a broader population. 6. ACKNOWLEDGMENTS

We gratefully acknowledge Mr. Junki Furuta for his assistance in carrying out this experiment. This work was supported by JSPS KAKENHI Grant Number JP21K20467. 7. REFERENCES

1. World Health Organization Regional Office for Europe. Night Noise Guidelines for Europe,

Copenhagen, Denmark 2009. 2. World Health Organization Regional Office for Europe. Environmental Noise Guidelines for

the European Region, Copenhagen, Denmark 2018. 3. Yokoshima, S., Morinaga, M., Tsujimura, S., Shimoyama, K., and Morihara, T., Representative

Exposure–Annoyance Relationships Due to Transportation Noises in Japan. Int. J. Environ. Res. Public Health , 18 , 10935, doi.org/10.3390/ijerph182010935 (2021). 4. Basner, M., Brink, M., and Elmenhorst, E. M., Critical appraisal of methods for the assessment

of noise effects on sleep. Noise & Health . 14(61) , 321-329 (2012). 5. Basner, M., Witte, M., and McGuire, S. Aircraft Noise Effects on Sleep-Results of a Pilot Study

Near Philadelphia International Airport. Int. J. Environ. Res. Public Health , 16 , 3178, doi.org/10.3390/ijerph16173178 (2019). 6. Perslev M., Darkner S., Kempfner L., Nikolic M., Jennum, P. J., and Igel, C. U-Sleep: resilient

high-frequency sleep staging, npj Digital Medicine , 72 , doi.org/10.1038/s41746-021-00440-5 (2021). 7. Takeda, T., Watanabe, T., Yoshida, K., and Mizuno, O., Time-dependent Sleep Stage Transitin

Model Based on Heart Rate Variability, DBSJ Japanese Journal , 14-J(16) (2016). (in Japanese) 8. Brennan, M., Palaniswami, M., and Kamen P., Do existing measures of Poincaré plot geometry

reflect nonlinear features of heart rate variability, IEEE Trans Biomed Eng , 48(11) , 1342–1347 (2001). 9. Sahoo, T. K., Mahapatra, A., and Ruban, N., Stress Index Calculation and Analysis based on

Heart Rate Variability of ECG Signal with Arrhythmia, Proceedings of 3rd International Conference on Innovations in Power and Advanced Computing Technologies (i-PACT) , (2019).