Performance of an Electroencephalography-Measuring Headband or Actigraphy Compared with Polysomnography in Older Adults with Sleep Disturbances

Abstract

Study Objectives:

We assessed the performance of an electroencephalography-measuring headband (HB) or actigraphy (ACT) compared with polysomnography (PSG) in older adults with sleep disturbances.

Methods:

Sixty-one older adults reporting insomnia and/or daytime sleepiness wore the HB for up to seven nights, actigraphy for seven days and nights, and completed an in-home PSG. We compared total sleep time (TST), wake after sleep onset (WASO), sleep onset latency (SOL), and sleep efficiency (SE) from all devices on the PSG night. For HB-PSG, we compared time in light, deep, and rapid eye movement sleep. For all comparisons, we calculated absolute differences and intraclass correlation coefficients (ICCs). We also evaluated the performance of the HB among the poorest sleepers (e.g., severe sleep apnea or insomnia).

Results:

Average age was 72.6[SD=6.5] years, 62.3% were female, and 77.1% were non-Hispanic White. For HB-PSG, we found good agreement for TST, WASO, and SE (ICCs ranging 0.82–0.91), while SOL and sleep stages were lower (ICCs 0.44–0.66). For ACT-PSG, we found moderate agreement for TST (ICC 0.73) and poor agreement for WASO, SOL, and SE (ICCs <0.50). For the poorest sleepers, HB-PSG showed good to excellent agreement for TST and WASO (ICCs 0.56–0.91), while ACT-PSG showed lower levels of agreement (ICCs 0.55–0.80 for TST and <0.59 for WASO). On average, participants wore the HB for 6.5 [0.8] nights and usability was rated highly.

Conclusions:

The HB outperformed ACT, including among the poorest sleepers. Devices like the HB are accurate, feasible, and could advance sleep health in older adults.

INTRODUCTION

Sleep disturbances occur in nearly half of older adults.^1,2 These disturbances are associated with cognitive decline, depression, disability, institutionalization, and high healthcare costs,^3–7 calling for their accurate identification and treatment. Unfortunately, self-reported sleep may not accurately reflect habitual sleep, leading to missed treatment opportunities or overtreatment with high-risk medications.^1,2,8–11 Polysomnography (PSG), which includes electroencephalography (EEG), is the gold standard for evaluating sleep,¹² but it is costly, burdensome, and may not reflect habitual sleep patterns.^13–15 Actigraphy (ACT) is the reference standard for objectively evaluating sleep over multiple nights, but it relies on activity counts to estimate sleep and wake time indirectly. As a result, ACT overestimates sleep time, underestimates time awake, and is less accurate in older adults with sleep disturbances or multiple chronic conditions.^10,11,16 Thus, tools that feasibly and accurately characterize habitual sleep in older adults are needed to improve the targeting and monitoring of interventions to improve sleep health and prevent adverse outcomes.

Recently, new wearable devices with EEG capability have been introduced. One such device is a multi-sensor EEG headband (HB). The HB was validated against PSG in small samples of healthy middle-aged adults (N=25 adults; average age 35 years)¹⁷ and older adults with Parkinson’s disease (N=10; average age 70 years),¹⁸ where it showed over 80% accuracy in both cases for sleep-staging compared to PSG. However, these studies excluded persons using hypnotic medications or with known insomnia or sleep apnea, resulting in diminished generalizability.^17,18 A recent study examined the performance of the HB vs. PSG in 62 older adults (average age 70 years), including 12 adults living with dementia, and found the HB to have acceptable performance for sleep-wake classification and more moderate performance for five-stage sleep classification.¹⁹ Average scores on validated sleep questionnaires were in the normal range in this study, suggesting most did not have disrupted sleep. The validity and usability of the HB in a more generalizable population of older adults with sleep disruption is needed, as this group has the highest prevalence of sleep disturbances and is also likely to experience disagreement between gold-standard and actigraphy measures of sleep.¹⁶

In the current study, we compared the agreement of HB-PSG and ACT-PSG in a sample of older adults with self-reported sleep disturbances. Participants wore the HB for up to 7 nights, ACT for 7 days and nights, and completed an in-home PSG. We hypothesized that sleep measures from the HB, measured via EEG rather than activity counts, would have higher agreement with PSG than ACT. We also assessed the feasibility of using the HB in this population. Due to its ergonomic design, we hypothesized that participants would have high rates of adherence to the 7-night protocol and report high rates of usability.

METHODS

Study Population and Protocol

This is a cross-sectional study of community-living older adults in the greater New Haven area of Connecticut. Participants were ≥60 years and reported sleep disturbances, including insomnia symptoms (i.e., difficulties with sleep onset, sleep maintenance, or early morning awakenings) and/or daytime sleepiness (i.e., frequent dozing, needing to nap, or feeling like one could fall asleep easily during the day) at least once a week. Participants were recruited via flyers at Yale-New Haven Hospital-affiliated primary care clinics and local senior centers. We excluded participants with severe cognitive impairment (defined as ≥4 errors on the 6-item Callahan screener),²⁰ and those who could not converse in English, refused or were unable to consent, lived in an assisted living or extended care facility, did not live within driving distance, or had an unplanned hospitalization in the past month. Importantly, we did not exclude otherwise eligible persons who were taking hypnotic medications or had a known diagnosis of sleep apnea, insomnia, or restless legs syndrome. All recruitment and data collection took place between January 2022 and December 2023. Of the 78 persons screened, 71 ultimately enrolled in the study (5 dropped before the baseline interview; 2 were not eligible [1 had an unplanned overnight hospitalization in the past month; 1 did not meet age requirements]). Among the 71 enrolled in the study, 8 did not complete the study protocol (6 dropped out of the study; 1 had no PSG data; 1 had no ACT data). Two additional participants were excluded due to HB data that could not be analyzed on the PSG night (1 had the HB off head for more than half the night; 1 had a HB recording that started in the middle of the PSG recording). Hence, the final analytical sample included 61 participants (85.9% of persons enrolled). The Yale University Human Investigation Committee approved the study protocol, and all participants provided written informed consent.

We conducted a week-long protocol with collection of demographic and clinical information, and subjective and objective sleep measures. During the initial visit (day 1), we collected clinical information (described below) and introduced participants to the sleep devices (HB, ACT, PSG) to be used during the protocol. On the night of the participant’s choosing (but after several nights of accommodation to the HB), we visited participant’s homes to set up an overnight PSG. This night, the three sleep devices were worn simultaneously. To minimize participant burden and prevent data loss, our team started HB recordings on the PSG night before leaving participant homes, while PSGs had an automatic start time selected to occur one hour before usual sleep onset for the participant. The start time for the PSG was later than for the HB to ensure the battery power would be sufficient to capture the entire sleep period on PSG. The following morning, participants removed the PSG equipment and continued to wear the HB (at night) and ACT (all day) for the remainder of the protocol. Participants were encouraged, but not required, to wear the headband for a total of 7 nights. Participants completed daily sleep diaries for 7 days and completed a usability questionnaire for the HB at the end of the protocol, when all study devices and diaries were collected.

Demographic and Clinical Characteristics

Demographic characteristics included age, sex, race/ethnicity (non-Hispanic White vs. other), education (college graduate vs. other), and occupational status (retired vs. other). Clinical characteristics included body mass index (BMI), cognitive impairment (Montreal Cognitive Assessment [MoCA] short form score < 12),²¹ self-reported physician diagnoses of diabetes, coronary artery disease, heart failure, or chronic obstructive pulmonary disease, physician diagnosis of a psychiatric condition (including depression, anxiety, bipolar disorder, post-traumatic stress disorder, or substance use disorder), and physician diagnosis of a sleep disorder (including sleep apnea, insomnia, or restless legs syndrome).

Participants were asked about their use in the previous two weeks of prescription and over-the-counter medications. Medication information was ascertained through a review of medication bottles, medication lists, or electronic medical records. For the total number of medications used, we included prescription medications and over-the-counter use of aspirin, pain medication (e.g., acetaminophen, ibuprofen, naproxen), allergy/cold medicines (e.g., diphenhydramine, cetirizine), or antacids (e.g., tums, famotidine, proton pump inhibitors). The count did not include bowel regimens, eye drops, topical lotions, nasal sprays, multivitamins or probiotics. We also inquired specifically about sleep medications by asking, “In the past two weeks, have you taken a sleep aide to help you sleep during the night? These can include a prescription or over-the-counter medication or an herbal or nutritional supplement.”

Measures of Self-Reported and Objective Sleep

Self-reported sleep measures included the Pittsburgh Sleep Quality Index (PSQI),²² the Epworth Sleepiness Scale (ESS),²³ and the Insomnia Severity Index (ISI).²⁴ The PSQI assesses sleep quality over the past month (score range=0–21), with higher scores indicating worse sleep quality and scores >5 denoting poor sleep quality.²² The ESS measures the likelihood of dozing or sleeping in eight hypothetical situations during the day (score range= 0–24), with higher scores indicating more severe daytime sleepiness.²³ The ISI captures symptoms and daytime consequences of insomnia based on the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition, Text Revision (DSM-5-TR).²⁴ ISI scores range from 0–28, with higher scores indicating more severe insomnia.²⁴ An ISI score of ≥8 establishes sub-threshold insomnia, while scores ≥15 indicate moderate to severe insomnia.²⁴ Self-reported daily napping was assessed via respondents answering “7” to the question “over the course of a week, how many days do you take a nap or doze?”.

Objective sleep measures included the HB, ACT, and PSG. The Dreem 2 HB (formerly from Dreem, Paris, France; now from Beacon Biosignals, Boston, USA) is a wireless, self-administered, non-invasive, FDA class II medical device. The HB measures physiological signals via five dry EEG electrodes (described below), a 3-dimensional accelerometer (to gauge movements, position, and breathing), and an infrared pulse oximeter.¹⁷ It records and stores multiple nights of data. The HB connects via Bluetooth to a mobile device application on a smartphone or tablet and transfers data via Wi-Fi to the sponsor’s servers. The HB electrodes are located over the anterior forehead and the back of the head, where flexible silicone protrusions enable signal acquisition from the scalp. Signal acquisition is at 250 Hz with a 0.4–35 Hz bandpass filter. The HB has five electrodes: F7 and F8 (located over the left and right frontal areas); FpZ (ground sensor located on the frontal band); and O1 and O2 (located in the left and right occipital areas). The five electrodes yield seven derivations (FpZ-O1, FpZ-O2, FpZ-F7, F8-F7, F7-O1, F8-O2, FpZ-F8). For this study, participants were not required to use a mobile device to manage the HB. During the baseline visit, participants were instructed on the placement and operation of the HB, including how to start and stop recordings manually, and how to use the power cord to charge the HB. While the battery life accommodates multiple nights of recording on a full charge, we instructed participants to charge the HB in the morning after every use. Data from all nights of recording were stored by the HB until devices were retrieved at the end of the week-long protocol. For the current study, we measured TST (total time of overnight sleep between sleep onset and final awakening), WASO (amount of time awake between sleep onset and final awakening), SOL (time to initiate sleep), SE (ratio of TST to time in bed), time in light sleep (stages N1 and N2), time in deep sleep (stage N3), and time in REM sleep. Sleep variables from the HB were generated using the HB’s proprietary autoscoring algorithm. Data quality metrics for the HB recordings were automatically generated at the time of autoscoring and included the proportion (%) of good quality EEG over the recording for a given channel and the off-head time (i.e., % of time where the HB was not detected as on the head). Quality scores are generated by a proprietary algorithm based on human expert interpretation of signal quality that would be sufficient for scoring sleep. Additional details for this algorithm are provided in Supplemental Table S1.

We used the Actiwatch Spectrum Plus (ACT; Phillips Respironics, Murrysville, PA, USA), a standard wrist-worn actigraph. ACT estimates sleep objectively and non-invasively by integrating the occurrence and degree of limb movements.²⁵ For the current study, participants were instructed to wear the ACT for 7 days and nights on the non-dominant wrist, averaging 7.2 (0.5) days worn. ACT data were analyzed using Phillips Respironics software with proportional integration mode. The software algorithm and sleep diaries were used to edit the raw data and generate different sleep metrics, including TST, WASO, SOL, and SE.

We performed a one-night, home-based PSG with the NOX-Self applied system (Nox Medical, Reykjavik, Iceland).²⁶ Trained assistants set up the entirety of the PSG system for participants around their habitual sleep time. The NOX-Self Applied system provides full-montage PSG, including channels for EEG, electrooculography, electromyography, airflow sensors, chest and abdominal respiratory effort bands, EKG, and oxygen saturation. For the current study, the HB was placed over the head after PSG equipment was placed to ensure that the sensors would not interfere with each other. PSGs were scored by a trained sleep technologist at Sleep Strategies, Inc. Sleep technologists used sleep diary data from the night of PSG to set lights-off and light-on times. The Apnea-Hypopnea Index (AHI) was the average number of apneas or hypopneas with at least 3% oxygen desaturation per hour during sleep. The Periodic limbic movement index (PLMI) was defined as the average number of occurrences of leg movements per hour of sleep.

Adherence and Usability of the HB

Adherence for the HB was assessed by calculating: 1) the mean number of nights the HB was used; 2) the percentage of nights participants used the HB for at least 4 hours (on nights used); and 3) the percentage of participants who used the HB for all seven nights of the protocol. Usability of the HB was assessed by eleven items from the WEarable Acceptability Range (WEAR) scale, which includes items on the aesthetics, ergonomics, and social desirability of wearable technologies.^27,28 Responses were recorded on a Likert scale that ranged from 1 (strongly disagree) to 5 (strongly agree).^27,28 Participants completed the WEAR scale at the end of the 8-day protocol. Five items with a negative valence (i.e., items 2–4, 8, 10; shown in Table 4) were reverse coded, meaning that higher scores indicated higher levels of disagreement for statements with a negative valence.

Table 4.

Adherence and usability data for the headband

Adherence	Mean (SD) or n (%)
Total Nights Worn	6.5 (0.8)
4+ Hours/Night Adherence	57 (93.4)
7 Nights Adherencea	38 (62.3)
Usabilityb	Median (IQR)
1. The sleep headband seems to be useful and easy to use.	4 (3, 5)
2. The sleep headband seems like “too much” technology.c	5 (4, 5)
3. The sleep headband restricts movement or physically gets in the way.c	4 (3, 5)
4. I needed the support of a technical person to be able to use the headband.c	5 (4, 5)
5. The sleep headband seems comfortable, not bulky.	4 (3, 5)
6. The sleep headband is sleek, not clunky.	4 (3, 4)
7. I would imagine that most people would learn to use the sleep headband very quickly.	4 (3, 5)
8. I found the sleep headband very cumbersome to use.c	4 (3, 5)
9. I felt very confident using the sleep headband.	4 (3, 5)
10. I needed to learn a lot of things before I could get going with the sleep headband.c	5 (4, 5)
11. I think the sleep headband could help people.	4 (4, 5)

Abbreviations: SD=standard deviation

^a60 (98.4%) participants wore the headband for at least 4 nights.

^bFrom the WEarable Acceptability Range (WEAR) scale; responses range from 1 (strongly disagree) to 5 (strongly agree).

^cReverse-coded; higher scores indicate higher levels of disagreement with the statement.

Statistical Analysis

Demographic, clinical, and sleep characteristics were described using means and standard deviations (SD) for continuous variables and counts with frequency (%) for categorical variables. To evaluate the performance of the HB and ACT compared with PSG, we used a standardized framework for testing the validity of sleep devices developed by the Sleep Research Society.²⁹ All sleep measures compared between the three devices were assessed on the night of PSG. TST, WASO, SOL, and SE were compared for HB-PSG and ACT-PSG, while additional measures (light sleep, deep sleep, REM) were compared for HB-PSG only. Because HB and PSG recordings were started at different times and may have been separated in time by one to two hours, we used the lights-off and light-on times from polysomnography to align recording times across all three devices. For analyses of SOL, 10 additional participants were removed from analyses as they had sleep onset occur before the automatic PSG start time. Pearson’s correlation coefficients were estimated. Absolute differences in sleep measures between HB or ACT vs. PSG were analyzed using paired t-tests (difference= device [headband or actigraphy]-polysomnography). Two-way mixed effects intraclass correlation coefficients (ICCs) were used to assess the agreement for HB-PSG and ACT-PSG for sleep measures. ICC reliability values <0.5 indicate poor agreement, 0.5–0.75 indicate moderate agreement, and values ≥0.75 indicate good to excellent agreement.³⁰ Bland-Altman plots were used to graphically show differences in sleep measures for HB-PSG and ACT-PSG.³¹ Bland-Altman plots display the agreement between two different measurement methods by plotting the differences against the averages of the measurements. Bias, limits of agreement (LoA), and minimum detectable change (MDC) were also computed.²⁹ For bias (i.e., mean difference), a negative value represents underestimation by the device compared with PSG. LoAs were calculated using the classic Bland–Altman approach (LoA= bias ± 1.96 × SD of differences). Confidence intervals were derived via non-parametric bootstrapping (10,000 resamples), without assuming normality, homoscedasticity, or proportional bias. The MDC represents the smallest detectable change in an estimate that exceeds the device’s measurement error. Assumptions for Bland–Altman analysis were evaluated by visual inspection of Bland–Altman plots and Q–Q plots of the differences. Normality of the differences was assessed with Shapiro–Wilk tests. Proportional bias (i.e., how bias was affected by the magnitude of the measures) was examined by regressing the mean difference on the average of the two methods. Homoscedasticity was assessed by visual inspection for trends in variance across the range of measurements.

To examine the sensitivity of results to HB signal quality, we repeated the above discrepancy analyses after excluding poor quality HB recordings. Poor quality recordings were defined as having a proportion of “good quality segments” at or below 50% across all 4 channels on the night of PSG or an off-head time >10%. We ultimately excluded 13 participants (n=13 had quality ≤50% across all channels; n=1 had off-head time >10% but also met the former criteria). For each night of HB use, we calculated the average HB signal quality by averaging quality scores across the four EEG channels.

Because increasing levels of sleep disturbance have been shown to lower agreement between sleep devices and PSG,^32,33 we evaluated the performance of the HB or ACT with PSG among subgroups of the poorest sleepers using the following thresholds: AHI ≥15, ISI ≥15, PSQI >5, and PLMI ≥15, and self-reported reported daily napping (vs. no napping or non-daily napping). We also examined agreement in persons with vs. without cognitive impairment (i.e., MoCA short form score < 12 vs. 13–16). Adherence and usability were assessed via descriptive statistics. To compare measures from methodologies providing multiple nights of data, we calculated means (±SD) for sleep measures for the HB, ACT, and sleep diaries. We excluded the PSG night for these calculations and included participants who had ≥4 nights of data (n=60). All data from the HB on these nights was automatically estimated and variables were generated using the HB’s proprietary autoscoring algorithm. Automatic estimations of SOL by the headband assume that the start of the HB recording coincides with the time that a participant starts trying to fall asleep.

All analyses were performed using SAS 9.4 (SAS Institute Inc., Cary, NC). All tests had statistical significance set at p < .05.

RESULTS

The characteristics of the 61 participants who had complete data from the three sleep devices are shown in Table 1. The average age was 72.6 (standard deviation [SD] 6.5) years, 62.3% were female, and 77.1% identified as non-Hispanic White. Overall, participants had poor sleep, with average scores on the PSQI and ISI of 9.4 [3.9] and 12.9 [4.7], respectively. Over 30% of participants had a pre-existing sleep disorder and reported use of a sleep aide in the past two weeks. PSG-related sleep characteristics included averages for TST and WASO of 365.4 [100.8] and 83.4 [62.8] minutes, respectively, and an average AHI of 19.6 [15.3] events per hour. Among the 6 participants who dropped out of the study, reasons included difficulty starting the HB, inability to sleep with the HB, and poor fit of the HB. Those who dropped out tended to be older and were more likely to have cognitive impairment (see Supplemental Table S2).

Table 1.

Demographic and clinical characteristics of the sample

Characteristics	N=61 Mean (SD) or n (%)
Age, years	72.6 (6.5)
Female	38 (62.3%)
Non-Hispanic White	47 (77.1%)
College graduate	51 (83.6%)
Retired	42 (68.9%)
Body mass index	28.5 (7.2)
Cognitive impairmenta	13 (21.3%)
History of diabetesb	10 (16.4%)
History of coronary artery diseaseb	6 (9.8%)
History of heart failureb	4 (6.6%)
History of chronic obstructive pulmonary diseaseb	8 (13.1%)
History of a psychiatric disorderc	24 (39.3%)
History of sleep disorderb	20 (32.8%)
Sleep Apnead	15 (24.6%)
Insomnia	7 (11.5%)
Restless legs syndrome	2 (3.3%)
Self-reported daily nappinge	9 (14.8%)
Medication countf	7.6 (4.8)
Reported use of a sleep aid in the past two weeksg	19 (31.7%)
Pittsburgh Sleep Quality Index score	9.4 (3.9)
Epworth Sleepiness Scale score	7.9 (5.0)
Insomnia Severity Index score	12.9 (5.0)
Total sleep time, minutesh	365.4 (100.8)
Wake after sleep onset, minutesh	83.4 (62.8)
Apnea hypopnea index, events/hourh	19.9 (15.4)
Periodic limb movement index, events/hourh	7.6 (14.4)

^aScore <12 on the Montreal Cognitive Assessment short form.

^bSelf-reported physician diagnosis

^cSelf-reported physician diagnosis of at least one of the following: depression (13), anxiety (17), bipolar disorder (1), post-traumatic stress disorder (3), substance use disorders (1).

^d2 participants used continuous positive airway pressure therapy during the study.

^eNaps average around 44 minutes a day, with the shortest nap being 10 minutes and the longest being 75 minutes.

^fTotal number of medications taken in the past two weeks, including prescription medications and over-the-counter use of aspirin, pain medication (e.g., acetaminophen, ibuprofen, naproxen), allergy/cold medicines, or antacids (e.g., tums, famotidine, proton pump inhibitor); does not include use of bowel regimens, eyedrops, topical lotions, nasal sprays, multivitamins or probiotics.

^gInquired by asking, “In the past two weeks, have you taken a sleep aide to help you sleep during the night?” and includes prescription or over-the-counter medications or supplements.

^hFrom polysomnography

Bland–Altman plots showing HB-PSG biases for sleep measures are presented in Figure 1. For the HB, the mean differences when compared to PSG for measurements of TST, WASO, SOL, and SE were not significantly different (Table 2). The HB overestimated WASO and underestimated TST, SOL, and SE. The HB showed excellent agreement with PSG for TST, WASO, and SE (ICCs ranging 0.82–0.91). SOL and sleep stage measures were less accurately estimated by the HB (ICCs ranging 0.44–0.66). Proportional bias was found for all measures except WASO and SE. Normality was violated for all variables except REM and heteroscedasticity was found for SOL and deep sleep. Log transformation of SOL resulted in a normal distribution, resolution of bias, and homoscedasticity, but did not improve the distribution or bias for other variables.

Figure 1.

Bland-Altman plots comparing measures of sleep from the HB with PSG

Plots show the agreement of the HB compared with PSG for TST (1a), WASO (1b), SOL (1c), SE (1d), light sleep (stages N1 and N2; 1e), deep sleep (stage N3; 1f), and REM sleep (1g). The x-axis represents the average of the sleep measures from the two devices, while the y-axis represents the difference between the two devices (i.e., HB - PSG). Solid gray lines indicate 95% limits of agreement. Dotted lines indicate 95% confidence intervals. Each point represents a measurement pair.

Abbreviations: HB= Headband; PSG=Polysomnography; REM=rapid eye movement; SE=sleep efficiency; SOL=sleep onset latency; TST=total sleep time; WASO=wake after sleep onset.

Table 2.

Agreement metrics for HB vs. PSG and ACT vs. PSG

Sleep Metric	Mean (SD)	r a	Bias (SD)b [95% CI]	p value	LoA lower bound [95% CI]	LoA upper bound [95% CI]	MDC	ICC (95% CI)
Headband c
TST	358.6 (96.4)	0.9	−6.7 (40.7) [−17.3, 2.9]	0.20	−86.0 [−112.8, −57.1]	72.6 [55.9, 87.1]	79.3	0.91 (0.86, 0.95)
WASO	92.8 (66.9)	0.8	9.4 (37.6) [−0.4, 18.3]	0.06	−64.3 [−74.1, −55.4]	83.1 [73.2, 92.1]	73.7	0.82 (0.73, 0.89)
SOLd	28.4 (23.37)	0.6	−2.3 (24.1) [−9.5, 3.4]	0.49	−49.5 [−78.3, −21.6]	44.9 [20.5, 64.5]	47.2	0.59 (0.38, 0.74)
SE (%)	75.7 (14.3)	0.8	−1.1 (7.8) [−2.9, 0.9]	0.29	−16.3 [−18.2, −14.3]	14.2 [12.4, 16.1]	15.3	0.84 (0.75, 0.90)
Lighte	251.6 (83.2)	0.2	−34.9 (60.0) [−50.6, −20.8]	<0.01	−152.5 [−198.4, −106.4]	82.6 [53.0, 109.4]	117.5	0.66 (0.50, 0.78)
Deepf	23.9 (26.8)	0.4	5.5 (25.8) [−1.0, 11.9]	0.10	−45.1 [−63.4, −28.3]	56.1 [41.4, 69.1]	67.8	0.51 (0.31, 0.67)
REMg	83.2 (47.5)	0.4	22.7 (40.2) [12.4, 33.0]	<0.01	−56.0 [−73.9, −38.2]	101.4 [83.6, 119.3]	78.7	0.44 (0.22, 0.62)
Actigraphy h
TST	417.3 (100.2)	0.8	52.0 (55.4) [38.0, 65.7]	< 0.01	−56.6 [−95.8, −22.9]	160.5 [126.1, 191.6]	108.6	0.73 (0.60, 0.83)
WASO	55.5 (44.0)	0.5	−27.9 (54.6) [−41.3, −13.9]	< 0.01	−134.8 [−163.4, −103.4]	79.0 [46.7, 118.1]	106.9	0.41 (0.18, 0.59)
SOL	2.4 (5.4)	0.3	−24.1 (27.4) [−31.2, −17.7]	< 0.01	−77.7 [−98.6, −55.3]	29.5 [16.8, 40.4]	53.6	−0.17 (−0.40, 0.08)
SE (%)	87.7 (9.1)	0.6	10.9 (10.9) [8.2, 13.6]	< 0.01	−10.3 [−18.4, −3.6]	32.2 [26.0, 38.0]	21.3	0.28 (0.03, 0.49)

Abbreviations. ACT=Actigraphy; CI=Confidence Interval; HB=Headband; ICC=Intraclass Correlation Coefficient; LoA=Limit of agreement; MDC=Minimum detectable change; PSG=Polysomnography; REM=Rapid eye movement sleep; SD=Standard Deviation; SE=Sleep efficiency; SOL=Sleep onset latency; TST=Total Sleep Time; WASO=Wake After Sleep Onset.

^aPearson correlation coefficient

^bDifference= device (headband or actigraphy) – polysomnography

^cApart from REM sleep, all HB variables showed proportional bias and non-normality. SOL and deep sleep showed heteroscedasticity. Classic bias and LoA estimates are reported for consistency and comparability across metrics. Bias and LoAs were reported using bootstrapped confidence intervals except for REM sleep.

^dn=51; 10 participants removed for sleep onset that preceded start of polysomnography.

^eN1 and N2 sleep stages; mean (SD) of 69.8 (12.5) and 79.1 (9.3) percent in headband and polysomnography, respectively.

^fN3 sleep stage; mean (SD) of 6.9 (8.7) and 4.7 (6.3) percent in headband and polysomnography, respectively.

^gMean (SD) of 23.3 (12.6) and 16.2 (8.6) percent in headband and polysomnography, respectively.

^hAll ACT variables showed proportional bias and non-normality. TST and SOL showed heteroscedasticity. Classic bias and LoA estimates are reported for consistency and comparability across metrics. Bias and LoAs were reported using bootstrapped confidence intervals.

In the overall sample of N=61 participants, the average HB quality scores on non-PSG nights and PSG nights were 53.0 [17.5] and 46.8 [21.9], respectively (t(60)=2.25, p =0.014). After removing n=13 persons with poor quality signals on the PSG night, average HB quality scores on non-PSG nights and PSG nights were 55.7 [16.1] and 53.8 [18.4], respectively (t(47)=0.64, p=0.26). The 13 participants excluded from the sensitivity analysis had similar age and self-reported and objective sleep measures as individuals included in the sensitivity analysis, less cognitive impairment, and a higher proportion of female sex (see Supplemental Table S4). In examining daily variations in signal quality in persons excluded from the sensitivity analysis, we found that 7 persons had consistent poor quality (i.e., average signal quality <50% across all nights), 3 had poor quality only on the PSG night, and 3 had variable quality across all nights. In the sensitivity analysis (n=48), agreement with PSG improved across all variables (see Supplemental Tables S4 and S5 and Supplemental Figures S1a–S1g). Bias was reduced in magnitude, distributions of differences appeared more symmetric, heteroscedasticity was less pronounced, and the number and severity of outliers decreased. This restriction led to narrower limits of agreement for HB-PSG.

Bland–Altman plots showing ACT-PSG biases for sleep measures are presented in Figure 2. Mean differences for ACT when compared to PSG were significantly different for all sleep measures, with ACT overestimating TST and SE and underestimating WASO and SOL. ICCs for ACT compared with PSG showed moderate agreement for TST (0.73 [0.60, 0.83]) and poor agreement for WASO, SOL, and SE (ICCs below 0.50). Proportional bias was found for all measures. Normality was violated for all variables and heteroscedasticity was found for TST and SOL.

Figure 2.

Bland-Altman plots comparing sleep measures from ACT with PSG

Plots show the agreement of the ACT compared with PSG for TST (2a), WASO (2b), SOL (2c), and SE (2d). The x-axis represents the average of the sleep measures from the two devices, while the y-axis represents the difference between the two devices (i.e., ACT - PSG). Solid gray lines indicate 95% limits of agreement. Dotted lines indicate 95% confidence intervals. Each point represents a measurement pair.

Abbreviations: ACT=Actigraph; PSG=Polysomnography; SE=sleep efficiency; SOL=sleep onset latency; TST=total sleep time; WASO=wake after sleep onset.

Table 3 shows the performance of the HB-PSG and ACT-PSG for TST and WASO among subgroups of interest. These included the poorest sleepers and persons with cognitive impairment, defined according to well-accepted thresholds from validated questionnaires or PSG. Among persons with sleep apnea (AHI ≥15), moderate to severe insomnia (ISI ≥15), poor sleep quality (PSQI>5), daily napping, and cognitive impairment (MoCA short form <12), the HB showed good to excellent agreement with PSG for TST (ICCs 0.87–0.91). Agreement of TST for HB-PSG in persons with increased leg movements during sleep (PLMI ≥15) was moderate (ICC 0.56 [−0.08, 0.87]). For all subgroups, comparisons of HB-PSG for WASO showed moderate to good agreement (ICCs 0.54–0.90). In contrast, ACT showed lower levels of agreement with PSG among these subgroups, with ICCs ranging 0.55–0.80 for TST and ICCs ranging 0.23–0.59 for WASO.

Table 3.

Agreement of sleep measures for HB-PSG or ACT-PSG among subgroups of interest

Index	TST			WASO
	Difference, min (95% CI)	p-value	ICC (95%CI)	Difference, min (95% CI)	p-value	ICC (95%CI)
Headband
AHI ≥15 (n=35)	12.9 (−29.0, 3.3)	0.1	0.90 (0.82, 0.95)	13.7 (0.5, 26.9)	0.04	0.83 (0.70, 0.91)
ISI ≥15 (n=20)	−13.8 (−38.2, 10.6)	0.3	0.87 (0.72, 0.94)	18.8 (−4.0, 41.7)	0.1	0.54 (0.18, 0.78)
PSQI >5 (n=52)	−7.2 (−18.9, 4.6)	0.2	0.91 (0.85, 0.93)	9.7 (−1.3, 20.6)	0.08	0.82 (0.71, 0.89)
PLMI ≥15 (n=8)	−47.4 (−98.1, 3.2)	0.06	0.56 (−0.08, 0.87)	44.9 (0.05, 89.8)	0.05	0.64 (0.05, 0.90)
Daily Napping (n=9)	1.5 (−36.5, 39.5)	0.9	0.89 (0.63, 0.97)	6.1 (−26.6, 38.8)	0.7	0.90 (0.67, 0.97)
MoCA score <12 (n=13)	8.3 (−23.2, 39.8)	0.6	0.88 (0.69, 0.96)	8.3 (−25.1, 41.7)	0.6	0.73 (0.36, 0.90)
Actigraphy
AHI ≥15 (n=35)	46.6 (25.5, 67.7)	<0.01	0.78 (0.61, 0.88)	−21.1 (−41.2, −1.1)	0.04	0.48 (0.19, 0.70)
ISI ≥15 (n=20)	34.3 (6.5, 62.1)	0.02	0.80 (0.60, 0.92)	−7.2 (−34.5, 20.1)	0.6	0.33 (−0.09, 0.65)
PSQI >5 (n=52)	54.3 (38.1, 70.6)	<0.01	0.71 (0.55, 0.82)	−29.2 (−45.2, −13.2)	<0.01	0.39 (0.14, 0.60)
PLMI ≥15 (n=8)	37.0 (−3.8, 77.8)	0.07	0.74 (0.23, 0.93)	−20.8 (−60.2, 18.6)	0.3	0.59 (−0.04, 0.88)
Daily Napping (n=9)	38.8 (−33.5, 111.0)	0.3	0.55 (−0.05, 0.86)	−18.4 (−94.1, 57.2)	0.6	0.31 (−0.34, 0.76)
MoCA score <12 (n=13)	57.3 (3.2, 111.4)	0.04	0.61 (0.15, 0.85)	−20.9 (−73.4, 31.6)	0.4	0.23 (−0.31, 0.66)

Abbreviations: ACT= Actigraphy; CI= Confidence Interval; HB=Headband; ICC= Intraclass Correlation Coefficient; PSG= Polysomnography; SD= Standard Deviation; TST= Total Sleep Time; WASO= Wake After Sleep Onset

Adherence and acceptability data for the HB are shown in Table 4. On average, participants wore the HB for 6.5 [0.8] nights. Sixty of the participants (98.4%) wore the HB for ≥four nights and 38 participants (62.3%) wore the HB for seven nights. Median scores for usability for each WEAR scale item were 4 and above (i.e., agree or strongly agree). When persons with and without cognitive impairment were compared, we found no differences in the average nights worn or percentage wearing the HB for 7 nights, and mean scores for items on the usability questionnaire were similar (see Supplemental Table S3).

Table 5 presents measures of sleep and night-to-night variability from the HB, ACT and sleep diary on non-PSG nights. Across non-PSG nights, ACT demonstrated the most consistent estimates of sleep, with the lowest night-to-night variability (as reflected by SD) for WASO, SOL, and SE. The HB showed comparable variability to ACT for TST, but exhibited greater variability for WASO and SOL. Sleep diary estimates were the most variable overall, particularly for TST and SE.

Table 5.

Measures of sleep and night-to-night variability from the HB, ACT and sleep diary on non-PSG nights ^*

Device	TST (min)	WASO (min)	SOL (min)	SE (%)
	Mean (SD)
HB	358.8 (61.2)	74.7 (38.5)	42.7 (25.1)	73.7 (8.1)
ACT	401.8 (59.8)	44.2 (20.9)	22.5 (19.4)	82.9 (6.8)
Sleep diary	384.7 (78.2)	39.9 (28.6)	31.3 (20.8)	82.2 (10.7)

Abbreviations: ACT=Actigraphy; HB=Headband; Min=Minutes; SD=Standard deviation; SE=Sleep efficiency; SOL=Sleep onset latency; TST=Total sleep time; WASO=Wake after sleep onset

^*From n=60 participants who had ≥4 nights of data outside of the PSG night.

DISCUSSION

Among 61 community-dwelling older adults with symptoms of insomnia and/or daytime sleepiness, we found that an EEG-HB out-performed ACT, providing measurements of sleep that were more consistent with PSG. This was true even among the poorest sleepers, including persons with sleep apnea or moderate-to-severe insomnia, in whom accurate measures are essential and for whom ACT may be less reliable. The HB showed lower accuracy for identification of sleep stages. Our study population was able to wear the HB over multiple nights and rated its usability highly. These results support the use of the HB or similar EEG-based devices to provide measures of sleep with increased accuracy in this population.

The present study extends prior validation work on the HB by focusing on a population of older adults with sleep disturbances, diagnosed sleep disorders, and concurrent use of hypnotics. Several studies, focused on younger populations without sleep problems, demonstrate the staging accuracy of the HB.^34,35 In a study by Ong et al., the HB (Dreem 3 version), Oura ring, a research-grade actigraph, and the Fitbit were compared with PSG in 40 healthy adults (average age 38 years) without sleep problems.³⁵ Among devices, the HB showed the smallest discrepancies in TST, WASO, SOL, and SE, narrower limits of agreement, and outperformed other devices with respect to sleep staging compared with PSG.³⁵ Another study by Beau et al. compared the performance of the HB (Dreem 3 version) to PSG in 31 adults (mean age 46 years) with insomnia disorder or comorbid insomnia and OSA.³⁶ Of note, persons taking hypnotics or with other health conditions that might affect sleep were excluded. The authors found that the HB accurately summarized SOL and WASO but significantly overestimated TST and SE.³⁶ Overestimation of TST and SE may be related to a tendency for the HB to misclassify wake epochs as sleep, an issue that would be especially prominent in persons with insomnia disorder. Studies by Gonzalez et al. (in 10 persons with Parkinson’s disease) and Ravindran et al. (in 62 older adults, 12 of whom were living with dementia) demonstrate acceptable accuracy of the HB across sleep stages in older adults, though neither study included persons with sleep disturbances.^18,19 Apart from the study by Beau et al., prior studies and findings from our work suggest that the HB performs well for summary measures, especially TST, WASO, and SE. However, similar to prior studies,^18,19,35,36 we found lower stage-specific accuracy. A consistent finding across studies is the over-estimation of REM,^18,19,35,36 which may relate to the HB’s restricted EEG montage. In the absence of electrooculography and with reliance mainly on frontal electrodes, algorithms may have difficulty distinguishing REM from light sleep, with ambiguous epochs being classified as REM. Previous studies also found, as we did, a tendency for the HB to overestimate deep sleep,^18,35 and underestimate light sleep (especially N1).^19,35,36 Because slow waves are maximal in frontal EEG, the algorithm may over-detect slow-wave activity, inflating estimates of N3, especially in older adults who have less true deep sleep. N1 may be underestimated because true N1 epochs may be misclassified as wake or REM. Upon examining outliers, we noted a trend for the HB to detect one or two epochs of sleep early in the recording, followed by long periods of wakefulness, which led to lower estimates of SOL and inflated estimates of WASO by the HB. In addition, outliers often had much lower estimates of TST and higher estimates of WASO compared with PSG, which may suggest the HB had difficulty differentiating between wake and light sleep in these individuals.

Few studies examined feasibility of the HB among study participants. In a study among 21 adults with chronic pain (average age 44 years), 86% of participants said they would wear the HB longer than the 2-night minimum requirement, and over half of the sample rated the HB as comfortable while sleeping.³⁷ Ong et al. noted that 25% of participants were unable to tolerate the HB.³⁵ As opposed to other studies, which examined HB performance over one night in a sleep lab,^18,19,35,36 our protocol may have enhanced feasibility by allowing participants to use the HB at home for several nights prior to PSG. While some might assume an older population to have more difficulty adapting to new technologies, we show high rates of feasibility and acceptability in this group. Importantly, our inclusion criteria required that participants identified as having a sleep problem, a population in whom successful use of a wearable device to assess sleep should not be assumed. Nevertheless, most of our participants could sleep with the HB, as evidenced by 98.4% of participants wearing the HB for ≥4 nights. The most common reason for participants doing fewer than 7 nights of recording was that they encountered difficulties doing the manual start for the recording. Most often, this occurred during the first few nights of their week-long participation. Participants were given the option to continue recording additional nights if they encountered difficulties in the beginning of the protocol, but this was not required. Two additional issues are important to note. The first is that among the 71 enrolled participants, six (8.5%) dropped out of the study due to inability to use the HB. Accounting for this group of early dropouts provides a better expectation of feasibility. Second, we did not require participants to use a mobile device to manage the HB. Doing so would have limited our ability to study a more generalized population and may have led to more difficulties with HB use. The manual start for recordings on the newer generation of headbands is easier and individuals who are comfortable with smartphone technology may find that the mobile device increases feasibility by making it easier to begin and end recordings.

A separate but related issue to feasibility is HB signal quality. Prior studies, despite being conducted in supervised settings, excluded up to 30% of recordings due to poor signal quality.^18,35,36 We opted to present results for all participants, regardless of signal quality, to provide a pragmatic assessment of the HB and because signal quality was assessed using a proprietary algorithm that has not been validated. As noted in our sensitivity analyses, focusing on higher-quality signals improves HB performance, suggesting that poor-quality signals were a driver of error in the overall sample. Our examination of the inter-day consistency of HB signal quality sheds light on potential drivers. First, persons with the poorest signal quality on the PSG night were not different from individuals with higher quality with respect to age or sleep characteristics and had less cognitive impairment. They were more likely to be women, which may have led to poor HB fit due to smaller head size.³⁸ Poor HB fit was frequently observed in men and women, despite the device’s adjustable sizing, and likely contributed to poor quality because of inconsistent electrode–scalp contact. We regularly used an elastic fabric headband over the top of the HB device to improve fit. Because signal quality was not assessed until equipment return, further adjustments could not be made in the field. Second, use of the HB and PSG concurrently decreased HB signal quality, especially among persons with poor signal quality on the night of PSG. While the HB and PSG equipment were positioned by research staff on the night of PSG, the protocol was otherwise unsupervised and adjustments to the HB fit throughout the night were not possible. To optimize signal quality, researchers could consider requiring participants to use an elastic headband over the EEG headband to improve contact with the scalp, incorporating use of a wireless modem with the HB to enable daily monitoring and real-time fit adjustments, or a supervised protocol. It should be noted that the newer HB model (i.e., Dreem 3) has updates to improve adjustability and fit for smaller head sizes. Improved signal quality may lead to higher sleep staging accuracy.

Our study is one of the first to simultaneously compare the performance of the HB and ACT, the current reference standard for assessment of sleep over multiple nights. Compared with ACT, the HB demonstrated narrower limits of agreement and closer correspondence with PSG, though still with clinically meaningful variability at the individual level. We also present measures of night-to-night variability from the HB, ACT and sleep diary on non-PSG nights. These results suggest that ACT provided the most stable nightly estimates, while the headband and diary measures showed greater fluctuation, especially for nocturnal wakefulness and sleep onset. Known limitations of ACT include its bias towards overestimation of TST and underestimation of WASO in all persons, as well as its tendency to perform worse among persons with multiple chronic conditions and sleep disturbances.^16,32,33 ACT relies on accelerometry to record movements and infers sleep periods via specialized algorithms. Thus, it is prone to errors of overestimation of TST in persons who may be awake but trying to sleep (e.g., persons with insomnia) but also to underestimation of TST among persons who may move more during sleep (e.g., persons with sleep apnea or leg movements during sleep). Conversely, the HB’s ability to measure EEG activity assesses sleep directly and may explain its superior performance, especially among the poorest sleepers. Thus, the HB may be more appropriate for studies requiring accurate sleep metrics among persons with disturbed sleep, while ACT may be better used as a tool to provide objective, longitudinal patterns of sleep in persons without major sleep disturbances or when 24-hour rest-activity patterns are the study focus. Future work might also examine the HB as a tool for assessing daytime napping, which is prevalent among older adults and may not be reliably measured via actigraphy. However, the HB’s algorithm cannot autoscore daytime sleep and its battery life may not accommodate 24 hours of recording on a full charge. With respect to variability measures over multiple nights, it is important to recognize that greater variability from a particular methodology does not necessarily indicate inaccuracy. Rather, the differences in variability may reflect the ways in which each method captures or smooths underlying fluctuations in sleep. Lower variability from ACT may be attributable to its tendency to misclassify quiet wake as sleep, thereby dampening true night-to-night differences. Conversely, the headband may be more sensitive to nocturnal awakenings, which could exaggerate variability. Without a gold standard across multiple nights, it is not possible to determine which method best reflects the true extent of night-to-night variation.

While the HB shows promising results with respect to measures of sleep duration, WASO, and SE, additional work is needed to improve its use for other sleep measures, especially sleep stage measures. As noted above, improvements in signal quality may improve staging accuracy. Because raw EEG data is available from the HB, one might consider direct scoring of sleep data by a trained sleep technologist rather than relying on the HB’s autoscoring algorithm. However, a preferred approach may be to improve the algorithm, as this would increase feasibility and scalability in future studies. Thus, an essential next step of our work will be to complete an epoch-by-epoch analysis for the sleep staging of the HB compared with PSG, which will contribute to improved performance of the algorithm. Additional steps to improve the algorithm may also include optimizing training data through additional studies that provide high-quality, well-labeled data from diverse sources; refining preprocessing of data through signal filtering and artifact removal; incorporation of multimodal data (e.g., from EOG or EMG) to improve accuracy; and continued external validation from different populations and environments. The HB’s autoscoring algorithm and the ability to provide data on sleep staging are major advantages compared to other wearable devices, and improvements to its performance have the potential to lead to significant advances in the field of sleep medicine.

This study’s strengths include its focus on a generalizable sample of community-living older adults, high representation of minorities, high retention rates, a pragmatic approach to data collection and analysis, and simultaneous comparison of multiple sleep devices on the same night in the home environment. Furthermore, the inclusion of participants with known sleep disorders and use of hypnotic medications enhances the validity of our findings. However, several limitations are important to note. First, the sample size, while larger than in previous studies, is modest and predominantly composed of highly educated individuals. Second, we could not complete an epoch-by-epoch analysis due to technical limitations that prevent reliable temporal alignment between the Dreem 2 HB and the PSG recordings. Without access to sample timestamps, precise synchronization is currently not feasible due to the sampling rate drift and greater inter-sample interval instability of the Dreem 2 device. While a previously available synchronization tool from the developers of the Dreem 2 is no longer available, we are working with engineers at Beacon Biosignals to overcome these limitations. An epoch-by-epoch analysis is needed to enhance our understanding of the sensitivity and specificity of the HB’s sleep staging compared with PSG among older persons with sleep disturbances. Third, the data distributions for most sleep variables in our analysis were not normal. While Bland–Altman analyses are thought to be robust to non-normality, the limits of agreement reported in Table 2 may not fully capture variability at extremes. Fourth, our usability data should be interpreted with caution. As noted above, several participants dropped out of the study due to difficulties using the HB. In addition, we adapted our protocol for the study population (e.g., not requiring participants to use a mobile device). Others may find that mobile devices increase or decrease feasibility depending on the population being studied. Finally, we studied the “Dreem 2” device, which has been replaced by the “Dreem 3”; thus, our performance and usability data may not be directly applicable to the newer device.

In summary, we found that an EEG-measuring HB demonstrated good to excellent agreement with PSG for measures of TST, WASO, and SE, and outperformed ACT, especially among the poorest sleepers. Additional validation of the HB is needed to improve the identification of SOL and specific sleep stages. Devices like the HB might provide accurate and feasible measures of sleep in habitual sleep environments, enhancing the evaluation and management of sleep health in older adults.

Statement of Significance.

In older adults, self-reported sleep measures may be inaccurate, but polysomnography is burdensome. We compared the performance of an electroencephalography-measuring headband vs. actigraphy with polysomnography in older adults. Participants had an average age of 73 years and reported insomnia and/or daytime sleepiness ≥once/week. We included persons taking hypnotic medications and with known diagnoses of sleep disorders. Compared with polysomnography, the headband outperformed actigraphy, including among the poorest sleepers (e.g., persons with sleep apnea or severe insomnia). Sleep stage durations from the HB had lower levels of agreement with polysomnography. Participants were able to use the headband over multiple nights and rated its usability highly. Future research should validate sleep staging of the headband compared with polysomnography.

DATA SHARING STATEMENT

The data supporting this study will be made available upon reasonable request by contacting the corresponding author.