Advanced search
Publication Cover
Expert Review of Medical Devices Volume 18, 2021 - Issue sup1: Digital Health
Submit an article Journal homepage
5,261
Views
5
CrossRef citations to date
0
Altmetric
Review

Hearables, in-ear sensing devices for bio-signal acquisition: a narrative review

Colver Ken Howe Nea School of Clinical Medicine, University of Cambridge, Cambridge, UKCorrespondence[email protected]
ORCID Iconhttps://orcid.org/0000-0003-0126-705X
ORCID Icon,
Jameel Muzaffarb Department of Clinical Neurosciences, University of Cambridge, Cambridge, UK
,
Aakash Amlanic Department of Ear, Nose and Throat Surgery, Birmingham Children’s Hospital, Birmingham, UK
&
Manohar Banceb Department of Clinical Neurosciences, University of Cambridge, Cambridge, UK
Pages 95-128 | Received 01 Sep 2021, Accepted 30 Nov 2021, Published online: 03 Jan 2022
1

ABSTRACT

Introduction

Hearables are ear devices used for multiple purposes including ubiquitous/remote monitoring of vital signals. This can support early detection, prevention, and management of urgent/non-urgent healthcare needs. This review therefore seeks to analyze the challenges and capabilities of hearables used to monitor human physiological signals.

Areas covered

Studies were identified via search (Medline, Embase, Web of Science, Cochrane Library, Scopus) and conducted in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses. Bias assessment used the Mixed Methods Appraisal Tool 2018 and Quality Assessment of Diagnostic Accuracy Studies 2nd Edition. 92/631 studies met the inclusion criteria and were qualitatively analyzed. The outcomes, applications, advantages, and limitations were discussed according to the vital signal measured. The bias risk ranged from low to high, with most studies facing moderate-to-high risk in subject selection due to small sample sizes.

Expert opinion

Most studies reported good outcomes for ear signal acquisition compared to reference devices. To improve practicability and implementation, wireless connectivity, battery life, impact of motion/environmental artifacts and comfort need to be addressed going forward. Hearable technologies have also shown potential synergies with hearing aids. In future, multimodal ear-sensing devices opens the possibility of comprehensive health monitoring within daily life.

1. Introduction

Assessing and monitoring the physiological and neural signals of individuals over extended periods of time and in a variety of settings, including the community, is key to the operation of future health systems. “Hearables’ are wearable electronic devices that can be placed in, on or around the ear. In some publications they are termed ‘earables,’ but we will use Hearables as our nomenclature here, as more publications appear under this name. They can be used for a wide range of purposes, including health monitoring by recording bio-physiological signals. It is an appealing technology for use in areas where portable, discreet and user-friendly monitoring is advantageous over conventional means of physiological signal acquisition. These include ambulatory monitoring, sports medicine, sleep monitoring/staging and occupational medicine [Citation1]. Furthermore, electroencephalography (EEG) based brain monitoring has also been investigated for several emerging interactive applications like brain–computer interface (BCI) and biometric authentication [Citation2,Citation3].

1.1. The ear as a sensor mount

The ear offers numerous advantages for bio-signal acquisition. The ear canal has a relatively stable position from which multiple vital signs and bio-signals [Citation4] can be derived. These include heart rate (HR), respiratory rate, core body temperature, blood oxygenation (SpO2), estimated blood pressure, electrocardiogram (ECG), EEG and electromyography (EMG) [Citation5]. Moreover, wearable ear devices can offer a discreet and unobtrusive means to monitor such vital signals compared to their bulkier conventional measurement methods. This enables data recording over extended periods in the community, with minimal restriction for the user. The concept of fitting multiple biosensors within a hearable to enable multimodality has also been a topic of interest [Citation6]. Apart from user convenience, multimodal ear sensing also has the potential for bio-signal integration that may aid in scientific, diagnostic and therapeutic purposes.

A problem inherent to hearable technologies and the ear as a sensor location is interference arising from motion and physiological artifacts. These can arise from activities such as eye movements and jaw-related muscle artifacts in the context of an ear-EEG, for instance [Citation7]. Moreover, for EEG and ECG, the reduction in number of electrodes on the ear compared to conventional measurement entails a trade-off between signal resolution and wearability/portability. These factors result in a lower signal-to-noise ratio (SNR) obtained from ear sensors compared to their conventional counterparts. Furthermore, for EEG in particular, the restriction of electrode positions around or in the ear also limits spatial coverage of the potential field on the scalp, which limits spatial resolution [Citation8]. Therefore, it is crucial to explore and better understand the challenges and capability of ear bio-signal sensing compared to standard methods of measurement.

1.2. Objectives

This review aims to identify and analyze the studies from the past 20 years on the use of wearable ear devices in measurement of physiological signals in human subjects. The timeframe was chosen to capture recent developments in the field. Devices included prototypes and validated designs used in a multitude of situations that include both clinical and non-clinical/home settings. The focus will be on the types of hearables available, their respective forms of vital signs measurement, and an assessment of the feasibility and effectiveness in recording and monitoring physiological signals.

Population: Human subjects from any age group, healthy or with existing medical conditions.

Intervention: Hearable device physiological signal measurement

Comparison: No specific comparator (can be reference measurements taken from standard/conventional devices)

Outcome: Reported feasibility and efficacy of hearable devices in recording bio-physiological signals compared to their conventional forms of measurement. Descriptive synthesis of results, advantages/disadvantages, challenges, and complications.

2. Methods

2.1. Design

The protocol for this systematic review and narrative synthesis was registered on the International Prospective Register of Systematic Reviews (224314) and has been created according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines () [Citation9].

Figure 1. PRISMA Flowchart.

Figure 1. PRISMA Flowchart.

2.2. Inclusion and exclusion criteria

Studies were eligible for inclusion met the following criteria: 1) trials including randomized clinical trials and quasi-experimental studies evaluating the efficacy of ear devices used for continuous physiological signal monitoring. 2) Studies published after 2001. Review articles were not included but reference lists were searched to find further primary studies. Studies were excluded if: 1) they involved animal experiments. 2) no abstract or full-text was available following attempts to contact the study authors. 3) No English translation was available.

2.3. Search strategy

A systematic literature search of Medline via Ovid, Embase via Ovid, Web of Science, Cochrane Library and Scopus was performed on 31 December 2020. The following search terms were used:

  1. Ear

  2. In-ear

  3. 1 or 2

  4. Hearable*

  5. Earphone*

  6. Earpiece*

  7. “Ear Device?”

  8. “Ear Sensor?”

  9. ‘Ear Sensing’

  10. ‘Ear Technolog*’

  11. Wearable*

12) 4 OR 5 OR 6 OR 7 OR 8 OR 9 OR 10 OR 11

  1. EEG OR Electroencephalogram

  2. ECG OR Electrocardiogram

  3. Vital sign*

  4. Biosignal?

  5. Bio-signal?

  6. Physiological signal?

  7. ‘Eye tracking’

  8. ‘Visuomotor tracking’

  9. ‘Heart rate’

  10. ‘Respiratory rate’

  11. ‘Breathing rate’

  12. ‘Cardiac rhythm’

  13. ‘Heart rhythm’

  14. ‘Sleep scoring’

  15. Speech

28) 13 OR 14 OR 15 OR 16 OR 17 OR 18 OR 19 OR 20 OR 21 OR 22 OR 23 OR 24 OR 25 OR 26 OR 27

  1. Implant

30) 3 AND 12 AND 28 NOT 29

2.4. Study selection

Two reviewers (CN and AA) independently screened all records by title and abstract identified from database searches. Any disagreement was resolved by discussion with a third reviewer (JM). Potentially relevant studies identified from the initial searches and abstract screening underwent full-text screening by the first reviewer (CN) prior to data extraction. Any disagreements was resolved by discussion between the two reviewers (CN/AA) and a third (JM) reviewer if required.

2.5. Data extraction

Data were extracted by the first reviewer (CN) and then checked by a second reviewer (AA). Extracted data were arranged in a spreadsheet (Excel, Microsoft Corp., Redmond, WA, USA). A study characteristic table will summarize the authors, date of publications, study design, setting and comparators (if any). A participant characteristic table will extract sample sizes, study population and age group. Device characteristic tables for each modality will be created, which summarizes bio-signal measurement method, wearable location, connection type, battery life, usage setting and main outcomes/conclusions.

2.6. Risk of bias (quality) assessment

Studies were assessed for risk of bias by the first reviewer (CN) and then checked by the second reviewer (AA). Discrepancies between the reviewers were resolved by discussion and arbitration by a third reviewer (JM). Due to the diverse and heterogenous study designs, two different tools were used for assessment. The Mixed Methods Appraisal Tool (MMAT) 2018 [Citation10] was used for 5 study designs, namely qualitative, quantitative randomized controlled, quantitative nonrandomised, quantitative descriptive, and mixed methods. For each study design, there were five criteria, resulting in a score of 0 to 5 accordingly. The higher the score the lower the risk of bias. The Quality Assessment of Diagnostic Accuracy Studies 2nd edition (QUADAS-2) [Citation11] was used to evaluate diagnostic accuracy study designs. This consists of 4 main categories for which risk of bias was assessed: patient selection, index test, reference standard and flow and timing. Additionally, the categories patient selection, index test and reference standard were also assessed for applicability concerns.

3. Results

Four hundred and thirteen potentially relevant studies were identified from the database search after duplicate removal, of which 4 studies were identified from the reference list of the database identified studies. Title and abstract screening resulted in the exclusion of 266 studies, while 138 reports had their full-text retrieved. After full-text screening, 46 studies were excluded because they did not fulfil the inclusion criteria, mostly due to studies not focused on hearable devices (n = 11), studies with no/little experimental data (n = 9) and studies that reported the same data that has already been extracted (n = 9). This resulted in the inclusion of 92 studies that met the eligibility criteria. A PRISMA flowchart of the search is shown in .

3.1. Description of studies

Ninety-two studies met the inclusion criteria with a total of 1129–1173 subjects (some papers described multiple experiments with different number of subjects for each part). Subjects were primarily healthy subjects, but also included subjects with chronic conditions, cardiac/neurological conditions such as suspected epilepsy and surgical patients. A study characteristic table is shown in . 1 study was conducted in neonates, another 1 included 1 child (4 years old) subject and 6 studies were conducted on elderly subjects. The other studies were conducted on adults, as represented in .

Table 1. Study Characteristics

Table 2. Participant characteristics

Of the 92 studies, 22 were from the US, 18 from UK, 11 from Korea, 8 from Germany, 9 from Denmark, and the rest were from Japan (n = 4), China (n = 4), Canada (n = 3), Switzerland (n = 3), Belgium (n = 2), Italy (n = 2), Thailand (n = 1), South Africa (n = 1), Netherlands (n = 1), India (n = 1), Taiwan (n = 1) and Hong Kong (n = 1). A large majority of the studies were preliminary/pilot studies (n = 69) of hearable devices or measurement of vital signs from around the ear. 20 were validation studies and 3 were proof-of-concept studies. 16 studies declared no conflicts of interest and the remaining 76 studies did not declare any potential conflicts of interest.

3.2. Quality of studies

summarize the quality assessment for all included studies using the QUADAS-2 or MMAT 2018 tools respectively. Sixty-eight studies had a measurement/diagnostic accuracy design and were evaluated using the QUADAS-2 while 21 studies had a quantitative descriptive design and 3 studies had a mixed-methods design, which were evaluated using the MMAT 2018. Heterogeneity of measured outcomes and experimental design precluded a meta-analysis.

Figure 2. QUADAS-2 Bias assessment.

Figure 2. QUADAS-2 Bias assessment.

Figure 3. MMAT Bias assessment.

Figure 3. MMAT Bias assessment.

3.3. Types and location of hearables, sensing principles, and potential applications

There are numerous form factors for hearable devices. This includes headphones, earbuds, earring sensors, earlobe clips, and other customized ear-fitted devices. They record physiological signals from a variety of locations around and from the ear, such as within the ear canal (external auditory meatus), from the cavum and cymba conchae, the earlobe, and other locations behind or around the ear (most commonly at the mastoid area). shows a visual of the outer ear anatomy reflecting these areas. The type of physiological signal that can be recorded at these locations and their advantages and disadvantages are discussed below.

Figure 4. Anatomy of the outer ear.

Figure 4. Anatomy of the outer ear.

3.4. Heart Rate (HR) measurements

The gold standard for heart rate measurements is the ECG, which records electrical signals originating from the heart through leads placed on the chest. This produces an ECG waveform over time from which the heart rate can be determined. Photoplethysmography (PPG) is one method by which the heart rate can be estimated using wearables [Citation12]. It optically measures volumetric changes in blood vessels by measuring light absorption with a light source and detector [Citation13]. Other means of calculating the heart rate involves analysis of waveform readings from ECGs [Citation14] and Ballistocardiogram (BCG) methods, which involves the measurement of mechanical changes/forces as blood is pumped during systole [Citation15]. Interestingly, one study detailed the use of thermometry to estimate HR [Citation16], which is based on measurement of heat transfer from the blood to the surrounding environment. Another device [Citation17] recorded sounds from the ear canal using an in-ear microphone and determined the HR through an algorithm. One other study also investigated the use of piezoelectric sensors to capture in-ear pulse waves corresponding to blood flow from the ear canal [Citation18]. Of the 92 studies, 22 studies reported measurement of HR from subjects. This data is shown in (single modality) and 10 (for multimodal devices).

Table 3. Device characteristics and study outcomes – Heart rate

The sensors for HR measurement were placed in several locations, ranging from within the ear canal to the earlobe and in areas behind and around the ear. The main advantage of HR measurements within the ear canal are its potential for reducing motion artifacts and environmental influences on noise level, given the relatively fixed position of in-ear sensors compared to sensors placed on the periphery (like the earlobe) [Citation19]. Moreover, sensors placed within the ear canal are less subject to the effects of centralization, where blood perfusion to the periphery is reduced in critical conditions like sepsis or hypothermia. For earplug-like devices or devices that cover the ear, however, there may be reduction in hearing, especially important within the elderly population who may have preexisting hearing loss [Citation12].

A total of 14 studies reported devices using the PPG technique to measure HR. Multiple studies also reported more than one main outcome as reflected in . Across these studies, all reported acceptable outcomes for HR measurement with mean absolute errors ranging from 1.8 to 2.77 bpm [Citation20,Citation21]. Bland-Altman analyses also generally showed good agreement with mean differences ranging from −0.613 to 0.78 bpm with acceptable limits of agreement, with a tendency for the devices to underestimate compared to reference devices [Citation12,Citation19,Citation20,Citation22]. Passler, et al. [Citation23] evaluated the performance of 2 currently available fitness tracking commercial in-ear devices Dash Pro and CosinussOne and concluded that while the devices showed good/excellent agreement levels, they tended to underestimate ECG derived HR values and were sensitive to motion artifacts. This was also concluded by Vogel, et al. in another study [Citation19], who attempted to reduce artifacts by developing an artifact recording system for monitoring photoplethysmography curves and acceleration at the head and hip.

Table 4. Device characteristics and study outcomes – Body temperature

Table 5. Device characteristics and study outcomes – Blood pressure

Table 6. Device characteristics and study outcomes – Respiration rate

Table 7. Device characteristics and study outcomes – SpO2.

Table 8. Device characteristics and study outcomes – ECG

Table 9. Device characteristics and study outcomes – EEG

Table 10. Device characteristics and study outcomes – Multimodal devices

He et al. [Citation24] report three studies using the ballistocardiogram method to measure the HR. The studies were mainly focused on measuring the pre-ejection fraction from BCG signals, but also used the frequency of J waves on the BCG corresponds to calculate heart rate.

ECG methods of HR measurement were used in three studies. Jacob, et al. [Citation25] recorded ear-ECG signals from behind the ear and reported that the fundamental heart beat frequency was only present in half of the cases considered, when compared to a reference PPG device. However, a second harmonic was present in all records and could accurately extract HR. In another study, Martin et al. [Citation17] reported a mean difference estimate of −0.44 BPM with limits of agreement (LoA) interval of −14.3 to 13.4 BPM compared to a wearable chest belt heart rate reference system.

Regarding HR measurement from thermometry, the authors [Citation16] reported that initial measurements had low accuracy and concluded that more electronic and sensor placement optimization is required.

HR measurement via in-ear pressure variance demonstrated promising results with a mean absolute difference of 0.62 compared to a commercial ECG device [Citation18].

In general, the above discussed Hearables demonstrated promising capabilities for recording HR from the ear, although Bland-Altman analyses showed a tendency of several devices to underestimate HR. This primarily applies to HR determined by PPG and ECG methods, although HR derived from in-ear pressure variance also demonstrated good results.

3.5. Body temperature measurements

The most commonly cited gold standard for core body temperature measurement is the pulmonary artery temperature which is taken using a pulmonary artery catheter with a thermistor at the distal port. This is typically used in critical care. In clinical settings other alternatives such as esophageal measurements with a thermistor are considered gold standard, due to the proximity of the esophagus to the heart [Citation26]. These methods, however, are too invasive for routine use and require specialist equipment. In most occasions, the tympanic temperature using the thermistor technique is considered a viable noninvasive alternative [Citation27].

Body temperature measurements were taken in 7 different studies, represented in . This is done through thermometry using thermopile sensors placed within the ear canal to measure the tympanic membrane temperature, which directly reflects the core temperature of blood [Citation28]. The key advantage is its noninvasive and unobtrusive nature compared to other forms of core temperature measurements such as rectal/oral measurements which are unsuitable for ubiquitous monitoring [Citation29]. Only 2 out of 7 studies used a comparator for in-ear temperature (both using commercial infrared ear thermometers). Bestbier, et al. [Citation13] concluded that the ear was a suitable location for core temperature monitoring with a mean error of 0.018 ± 0.516°C. Ota, et al. [Citation29] also reported temperature fluctuations with biking recorded with the in-ear device compared to skin temperature measurements and a commercial IR ear thermometer. They concluded that the device was able to track core body temperature regardless of activity and environment of the user, although they did not compare it to any invasive probes (e.g. Rectal) which may more accurately reflect core body temperature. In summary, ear canal Hearables demonstrate a good ability to record body temperature measurements with low mean error and is less sensitive to environmental influences. However, only a small number of studies used a comparator which makes evaluation of measurement errors/accuracy difficult.

3.6. Blood pressure estimations

Blood pressure measurements were traditionally taken with the mercury sphygmomanometer using the auscultatory method, which is regarded as the gold standard. In this method a cuff is wrapped around the arm and inflated, while a stethoscope is placed along the brachial artery and Korotkoff sounds are used to indicate the systolic and diastolic blood pressure [Citation30]. However, these devices are now being phased out in favor of non-mercury devices such as the aneroid and electronic sphygmomanometer. The former involves an aneroid (liquid-free) gauge that detects pressure in place of the mercury manometer, but otherwise uses the auscultatory method. The latter uses the oscillometric technique, where pressure oscillations in the cuff are recorded during gradual deflation to obtain an automatic reading [Citation30]. Both of these techniques are used for clinic blood pressure readings, and although they are noninvasive, the measurement is highly prone to motion artifacts (requires subjects to be still), changes in arm elevation and necessitates a cuff around their arm. Moreover, these devices record clinical readings that may not be reflective of the patient’s actual blood pressure, as patients can demonstrate the white coat effect. This phenomenon occurs when clinic blood pressure readings are artificially raised due to the anxiety of having blood pressure monitored by a clinician. For ubiquitous monitoring purposes, an alternative method for BP estimation is via calculation of the Pulse Transit Time (PTT), which is inversely correlated with BP [Citation31]. This measures the blood wave propagation time between 2 arterial sites on the body. The most widely used measurement techniques estimate PTT start and end time via ECG and PPG. The ECG heartbeat peak signals the PTT start time (correlating to pressure wave occurrence at a proximal site) while the PPG measures volumetric changes as blood reaches the distal site indicating PTT end time.

2 studies reported data on BP estimation via PTT, while 1 study only mentioned the possibility of measuring PTT from acquired data. Only 1 out of 3 studies took both ECG and PPG measurements from behind the ear (left mastoid area), 1 study recorded PPG from the earlobe while having a portable chest ECG device and the other study recorded PPG from the earlobe and wrist simultaneously. These data are represented in .

Only one study, by Zhang, Q., et al. [Citation20] reported the performance of their device’s BP measurement against a reference, with a mean error and standard deviation of −1.4 ± 5.2 mmHg, mean absolute error of 4.2 mmHg and root-mean-square error of 5.4 mmHg. Separately, Zhang, Y., et al. [Citation32] reported a maximum correlation factor for the Diastolic BP/ln(PTT) at about 0.3. Although this is low, the authors concluded that increasing the number of participants may improve reliability. These studies demonstrate the potential feasibility of continuous BP measurement via pulse arrival time differences.

A key advantage of PTT BP estimation methods is that calculations can be done on software automatically and measurement techniques require no cuffs/pressure, thus enabling devices to be more portable and unobtrusive compared to standard auscultatory/oscillometry methods. PTT measurements, however, require at least two sensors for signal acquisition and their tolerance to motion artifacts are not well studied [Citation33]. This is an issue that needs to be considered when applying such measuring principles for ambulatory/continuous monitoring purposes.

3.7. Respiration rate measurements

RR measurements are typically taken via manual counting in non-emergency settings. This is regarded as the ‘industry standard’ while the gold standard is via capnography. This method involves recording the concentration of carbon dioxide in respiratory gases through attaching a capnograph (carbon dioxide sensor) to a plastic cannula worn over the nose and mouth [Citation34]. Understandably, these measurement techniques are not suitable for ubiquitous monitoring of RR.

A total of 5 studies reported devices/prototypes measuring RR (). A variety of measurement methods have been used including in-ear sound microphones (n = 2), use of a LED Infrared radiation sensor to measure ear canal form changes (n = 1) and utilizing the principle of respiratory sinus arrhythmia (RSA) combined with algorithms to extract respiration rate from PPG signals. The RSA is a natural phenomenon where heart rate is increased during inspiration and decreased during expiration. Only 1 study reported using conchae electrodes with a custom fitted ear mold to measure PPG signals [Citation22], while the rest used in-ear sensors.

In-ear microphones generally suffered from poor signal quality and susceptibility to multiple artifacts. Goverdovsky, et al [Citation6]. reported that at low rates (4–8 breaths per minute; bpm), the acoustic signal was too low to be measured reliably. Only at 12 bpm or more does the measurement become more reliable. Martin, et al [Citation17]. utilized an algorithm to extract breathing sounds from the in-ear microphone signal. The algorithm had an absolute mean error of 2.7 cycles per minute and a Bland–Altmann analysis showed a mean difference of 2.40 CPM with a limit of agreement (LoA) interval of −2.62 to 7.48 CPM. Notably, both studies reported that poor earplug fitting and low breathing sound amplitudes were limitations of this measurement method. More optimization of the prototypes and algorithms will be necessary before such technologies can be implemented for ambulatory/work place safety monitoring purposes.

For PPG methods of RR estimation, the limitations are closely tied to heart rate variability. Bestbier, et al. reported that elevated hearts rates >80 bpm can reduce the magnitude of the RSA and cause false detection of breaths [Citation13]. Moreover, if the RR approaches half of the heart rate, the sampling frequency would be too low for reliable breath detection. They concluded that PPG methods of RR estimation are most accurate at low heart and respiratory rates. Venema, et al. [Citation22] used a binary Naïve Bayes’ classifier to estimate moments of inspiration and expiration based on heart rate variability and signal amplitude variation. With reference to a thorax polysomnography belt, they achieved a sensitivity and specificity of 81.4% and 86%, respectively, across 3215 breathing cycles, with an estimated average breathing frequency of 13.9 per minute across 8 subjects. These results indicate that such devices would be useful in day-to-day monitoring of healthy subjects, although their efficacy in measuring deviations from the norm will need to be improved.

RR measurement using in-ear LED IR sensors was reported to be more resistant to environmental noise compared to in-ear microphones [Citation35]. In their study, Taniguchi, et al. reported that their device achieved an accuracy of 100% for nose breathing of 12 bpm, 93.8% at 16 bpm, and 93.8% at 20 bpm. However, the device is intended more as a point-of-care testing tool rather than a long, continuous measurement of RR.

3.8. SpO2 measurements

SpO2 is a percentage estimate of arterial blood oxygen saturation. The gold standard measurement for this parameter is the arterial blood gas analysis, which is both invasive and painful for the subject. In most clinical settings it is measured through pulse oximetry, which utilizes PPG techniques for measurement [Citation36]. In this case, 2 wavelengths are used simultaneously (red light 660 nm and infrared light, 880–940 nm) and the ratio of absorbance of both wavelengths enables an estimation of blood oxygen saturation due to differential absorption of infrared light by oxygenated and deoxygenated hemoglobin [Citation37]. There are two methods to detect this – transmittance and reflectance oximetry. The transmittance method directly measures light transmitted through tissue and is typically used peripherally such as the finger. The reflectance method detects the backscatter of light using a sensor placed adjacent to the emitter [Citation38]. The main advantage of the reflectance method is that unlike transmittance, it is not restricted to peripheral sites where tissue is thin. This is important because such peripheral body sites are affected by hypothermia and vasoconstriction that can impair accuracy of SpO2 readings [Citation38]. In terms of hearables, this also enables recording from sites like the ear canal which as mentioned above has the advantage of greater stability.

Eight studies reported devices that measured SpO2, 7 of which used reflectance pulse oximetry/PPG and 1 which used transmittance-reflectance oximetry as shown in . The sensors are located predominantly in 2 locations – the earlobe (n = 4) and in the ear canal (n = 3), with 1 study reporting measurements from the tragus.

The results were generally favorable with 1 study reporting a bias of 0.26% and precision of 4.14% for the ‘TOSCA’ sensor (using transmittance-reflectance) [Citation39] and another study [Citation40] reporting a Bland-Altman bias of −1.7% with limits of agreement of −6.8/+3.9% for their:V-Sign 2” sensor (using the reflectance method). Measurement errors were also low, with Venema, et al. [Citation41] investigating the effect of high altitude on SpO2 measurement, reporting mean errors of 1.17% (sea-level), 1.14–1.54% (High altitude, 2500–5300 m) and 2.19% (Very high altitude, 5300–8848 m). However, this study reported that measurements were strongly affected by motion artifacts and could be difficult under cold and wet conditions. Bestbier, et al. [Citation13] reported a mean error of −0.22 ± 1.50% but concluded that further testing was required to evaluate the SpO2 measuring capabilities of their device due to lack of measurable variations in SpO2 in healthy subjects. Moreover, Davies, et al. [Citation42] compared in-ear SpO2 measurements with measurements obtained from the finger, reporting a low root-mean-square difference of 1.47% and concluded that in-ear SpO2 monitoring is both convenient and feasible, especially with its robustness to temperature-induced vasoconstriction compared to finger measurements. These results demonstrate the feasibility of reflectance-based ear SpO2 monitoring for purposes of ambulatory/continuous hospital/continuous monitoring, although there are similar concerns regarding motion artifacts and recording extreme deviations from the norm.

3.9. ECG and integrated sensing applications

An ECG is the process of obtaining electrical signals originating from the heart through electrodes placed at different positions on the body, for which the 12-lead ECG is considered to be the gold standard. Including the studies discussed earlier for HR/BP measurements that used an ECG method, a total of 14 studies involved ECG measurement from a hearable device (). In this context ECGs were recorded from various positions around and within the ear. Sensor positions vary, with 11 studies reporting electrodes behind the ear (8 at the mastoid region, 3 unspecified), 2 studies within the ear canal and 1 study below the earlobe.

8 studies recorded the ECG using electrodes placed on both ears thus obtaining an across-ears signal, while 5 studies placed 1 electrode on the ear and another on the neck. One study described obtaining the ECG by placing 1 electrode each on the right ear and left arm respectively.

Apart from their use in HR/BP estimation, ECGs obtained from the ear have also been investigated for uses such as stress monitoring and biometric identification [Citation43,Citation44].

The main outcomes reported for the acquisition of ear-ECG, excluding the outcomes from studies already discussed in the HR/BP measurements section, are as follows: ECG graphs/morphology analysis (n = 7), mean difference in ECG time readings (n = 3), biometric identification rates (n = 1), mean error and standard deviation compared to a standard chest ECG (n = 1), signal quality/SNR (n = 2), classification accuracy for mental state in combination with EEG results (n = 2).

Overall, there is generally a good degree of concordance between ear-ECG and reference measurements, with capability to detect some or all components of the ECG waveform (PQRST) [Citation45,Citation46]. These were studies that used ear-neck electrodes and bilateral in-ear canal electrodes respectively. 1 study [Citation47] reports that ear-ECG derived from electrodes placed on the right ear and left arm is a linear combination of regular ECG leads I and III, with high correlation for lead I. Nonetheless, numerous studies reflected that ear-ECG signal quality tends to be weak with a low signal-to-noise ratio [Citation46,Citation48]. Celik, et al. [Citation45] explored various configurations of electrode positions (ear-neck/across-chest) and designs (gel/dry) and concluded that the combination of behind ear-neck electrode placement and dry electrodes was able to produce a SNR which is comparable with across-chest ECG recordings. They however noted that the current wired ear-neck measurements are susceptible to motion artifacts that may be reduced with a wireless sensing option.

3.9.1. Multimodal sensing and cardiac/exercise monitoring applications

In combination with the above discussed physiological signals, ear-ECG can be applied in areas of ambulatory cardiac monitoring and exercise monitoring. He, et al. [Citation24] provided proof-of-concept using a monitor that measures BCG, ECG, and PPG and estimates a variety of cardiac parameters, such as HR, pre-ejection period, stroke volume, cardiac output and PTT. Testing was performed in 13 subjects and the authors highlighted the need for larger studies to better access inter-parameter relationships like BCG and stroke volume. In another study, Celik, et al. [Citation45] proposed a multi-smart sensor system that integrates signals from ear-ECG, ear body temperature sensing and finger PPG, demonstrating the potential and feasibility of multimodal measurement from the ear in cardiac monitoring. Separately, Winokur, et al. [Citation49] also presented a wireless ear device that measures ECG, BCG and PPG from behind the ear, demonstrating a correlation between -ln(PTT) and arterial blood pressure. The PTT derived arterial blood pressure had a mean error of −0.07 mmHg and standard deviation of 3.64 mmHg compared to a reference finger blood pressure monitor. Further, they also described a correlation between R-J interval (from BCG and ECG) and pre-ejection period. In the field of exercise monitoring, Gil, et al. [Citation14] proposed a device which could simultaneously measure ECG (to determine HR) and lactate and pH sensing from the ear. They conducted an indoor-cycling trial demonstrated that the device could record characteristic changes, albeit with a low SNR. These studies have highlighted important considerations for feasibility of ear devices in cardiac ambulatory and exercise monitoring, namely: portability/wireless connectivity, multimodality, accounting for motion artifacts and low SNR.

3.10. EEG signals and potential applications

EEG provides a direct measurement of the electrical activity in the brain. In conventional clinical practice this is measured from the scalp, which involves attachment of multiple scalp electrodes to the head and generally requires patients to be still during measurement (not motion tolerant); it is also stigmatizing if used in daily-life due to multiple wires and the cumbersome apparatus [Citation50]. This precludes its use in ambulatory/continuous monitoring. The potential of recording such signals from the ear offers a discreet, unobtrusive, portable means of long-term brain activity monitoring [Citation2].

The ability to record EEG signals accurately from the ear is investigated in a total of 48 studies. These studies were quite heterogenous and reported multiple different outcomes () depending on the proposed use of the device. This includes ambulatory/continuous monitoring, stress monitoring, epilepsy/seizure monitoring, sleep monitoring and staging, BCI applications, and identity authentication.

For studies that investigated the feasibility of recording EEG signals from the ear, the main outcomes were related to EEG graph analyses (n = 5) such as grand average power spectral density (PSD) graphs for one or more EEG paradigms (evoked potentials/event-related potentials). Stress monitoring studies (n = 3) all reported classification accuracies and 1 study reported sensitivity and specificity. Epilepsy monitoring studies (n = 4) reported sensitivity and/or specificity measures. Sleep staging/monitoring studies (n = 14) reported Kappa agreement (n = 6), classification accuracy (n = 6), sensitivity and specificity (n = 2), and hypnogram analysis (n = 1). Studies describing BCI applications (n = 15) mainly reported classification accuracy (n = 9), Information transfer rate (n = 3), SNR (n = 2), EEG graph analysis/PSD analyses (n = 7) and sensitivity and specificity (n = 1). Identity authentication studies (n = 3) all reported authentication accuracy as the main outcome with 2 studies reporting half total error rate. 2 studies used EEG for attention monitoring, and both reported classification accuracy. 1 study described emotion classification accuracy while another used EEG for hearing threshold estimation in subjects with sensorineural hearing loss.

Overall, the studies that compared EEG readings between the ear and conventional scalp electrodes have reported good outcomes. This has been tested against several EEG steady state and evoked paradigms, including auditory brainstem response (ABR), envelope following response (EFR), auditory steady state potentials (ASSRs), steady state visual evoked response (SSVEP) and alpha band modulation. Two studies reported devices that were able to consistently record ASSRs from subjects, both within the laboratory [Citation51] and in real-life environment [Citation52]. The authors of the latter study [Citation52] noted, however, that ear-EEG could not produce statistically significant recordings for auditory onset response and alpha band modulation when applied in real-life settings.

Another study investigated the ‘cEEGrid’ device and demonstrated that it could record ABRs and EFRs within a controlled lab environment [Citation53]. The same study also noted that hearing impaired subjects showed less detectable responses, which must be considered when trying to apply ear-EEG in synergy with hearing aids. Further studies on the device have demonstrated the possibility to disentangle various non-brain artifacts from ear-EEG data using a method of independent component decomposition [Citation54]. This includes heartbeats, eye-blinks and eye movements. However, the ability to remove such artifacts from ear-EEG data without impacting the EEG signals of interest has not been fully explored.

A separate device from another study [Citation55] showed that using viscoelastic earpieces could render the device immune to motion artifacts generated by pulsatile ear canal wall movements while being able to record lower EEG frequencies. These studies provide proof-of-concept and pilot data on the feasibility of recording different forms of ear-EEG.

There are a number of commercial ear-EEG headsets available, for example, the Neurable [Citation56] headset and the mBrainTrain Smartfones [Citation57]. These devices feature around the ear recording, with 16 channels and 8 channels, respectively. They have been proposed for various BCI and monitoring functions. For example, the Neurable includes a focus tracking function and the mBrainTrain Smartfones have been investigated for a music-based stimuli BCI system [Citation58].

3.10.1. Stress/mental workload monitoring

Ear-EEG has been used in conjunction with ECG by 2 separate studies [Citation43,Citation59] to monitor stressed and relaxed states. The first study used a support vector machine to classify stressed (Stroop test S1 and Arithmetic test S2) and relaxed (R1 and R2) states. They concluded that with a sensitivity of 90%, specificity of 85% and accuracy of 87.5%, ear-EEG signals together with heart rate variability could assess and classify stress in 14 subjects. The second study came to a similar conclusion and developed a neural network that could classify stress states with 75% accuracy using EEG and ECG data. Furthermore, another study demonstrated a two-channel EEG system that could distinguish between resting and visuomotor tracking states with an average accuracy of 79.30 ± 4.85% across 10 subjects [Citation60]. More recently, Hölle et al. [Citation61] studied the feasibility of using long-term ear-EEG recordings (more than 6 hours) in an office environment. This involved running an auditory oddball task to assess participants’ attention throughout the day, using the cEEGrids technology. This study demonstrated that it was feasible to continuously record brain activity over extended periods beyond the laboratory environment, adding preliminary insight into the potential of ubiquitous monitoring using ear-EEG. To this end, the authors concluded that a multimodal approach with sensors to characterize and account for interference from the environment was important to fully translate from the laboratory to ear-EEG monitoring in daily life.

3.10.2. Seizure and epilepsy home monitoring

In terms of seizure monitoring, ear-EEG has been shown to be effective in detecting both focal and generalized tonic-clonic seizures. Using a behind-the-ear device, 1 study reported a median sensitivity of 94.5% and a false detection rate of 0.52 per hour for focal epilepsy [Citation62]. Another study, also using a behind-the-ear device, used an automated algorithm that could detect temporal lobe epilepsy with a 64.1% sensitivity and 2.8 false-positive detections (FPs)/24 hours [Citation63]. Other authors report that ictal morphology and frequency dynamics can be observed from visual inspection and time-frequency analysis [Citation64]. Using a custom fitted ear mold which recorded signals from within the ear canal and the conchae, the same authors later reported that generalized tonic-clonic seizures could be detected within 4.2–12.9 s of onset with 100% sensitivity (CI 29.2–100%) without false positives, in 820.7 hours of ear-EEG.

Ear-EEG has also been demonstrated to be able to reliably detect typical absence seizures. Swinnen et al. [Citation65] recently evaluated a behind-the-ear 2-channel EEG hearable, the ‘Sensor Dot,’ in detecting typical absence seizures in comparison to a 24-hour 25-channel video-EEG. This device demonstrated robust performance, with a sensitivity of 81%, precision of 89% and F1 score of 73% on blind reading of ear-EEG data. Combined with a developed algorithm, this significantly reduced the review time and showed better performance (sensitivity 83%, precision 89% and F1 score 87%).

It has to be noted that all the above studies had not tested their device in a home environment, only within a hospital setting or laboratory setting. Swinnen et al. [Citation65] found that a majority of false positives on the ‘Sensor Dot’ were due to chewing artifacts.The interference of motion artifacts on the ability to recognize seizures in a ubiquitous monitoring setting has not been evaluated and is one of the main issues that will have to be addressed going forward. Regardless, these preliminary studies highlight the potential of using ear-EEG recordings for seizure/epilepsy home monitoring.

3.10.3. Sleep staging

The gold standard for assessing sleep is polysomnography. This technique combines EEG, ECG, EOG, and EMG to identify sleep stages and wakefulness [Citation66]. All devices recorded EEG signals from within the ear canal and sometimes also from the conchae, except for 1 device that recorded it from behind the ear. Notably, all studies that compared ear-EEG with scalp-EEG with a kappa agreement noted substantial agreement (>0.60) between the two recordings [Citation6,Citation67–71]. Nakamura, et al. [Citation71] demonstrated a method for automatic sleep-stage prediction from ear-EEG data that showed 74.1% accuracy over 5 sleep-stage classifications in comparison with full polysomnography. Moreover, Goverdovsky, et al. [Citation6] Alqurashi, et al. [Citation67] and Looney, et al. [Citation69] All reported devices that could distinguish between wakefulness and sleep stages (SWS/N2, N1-3, all non-REM sleep vs. wake respectively) in healthy participants. Moreover, Goverdovsky, et al. presented a device which was multimodal, recording a mix of cardiac, respiratory and EEG signals. They also mitigated motion artifacts and blood vessel pulsation artifacts via a viscoelastic earpiece, and further demonstrated how cross-modal information enabled their device to deal with jaw clench artifacts. Ha, et al. [Citation72] also demonstrated a multimodal device recording both EEG and NIRS (near IR spectroscopy) which could provide 88.1%, 77.9% and 65.9% classification accuracy of 1, 2 or 3 missed stimuli, respectively, in the Oxford Sleep Resistance Test, proposing a possible auditory alarm function for their device.

Mikkelsen, et al. investigated the use of wet electrode ear-EEG in 9 full night recordings which showed promising results compared to manually scored PSG [Citation73]. They then followed up with a comprehensive study comparing 80 full-night (from 20 subjects over 4 nights) concurrent ear-EEG and partial PSG recordings [Citation74]. This was conducted in the home setting and with the use of dry electrodes, which are more practical for general home use. This involved a custom-designed earpiece with 6 electrodes to record EEG from within the ear, at the conchae and the tragus. The study demonstrated that their automatic sleep scoring of ear-EEG data had a kappa coefficient of 0.73 when compared to manual scoring of scalp-EEG, indicating good agreement. Moreover, the participants also reported good comfort outcomes which did not affect their sleep, highlighting the prospect of an unobtrusive, yet effective ear-EEG device for continuous home sleep monitoring.

More recently, the cEEGrid ear-EEG sensor was also evaluated in a home environment for sleep staging. This involved a cEEGrid device placed around the right ear, consisting of 8 electrodes, complemented with a single electrode on the Fpz location and 2 EOG electrodes near the eyes. In this study, da Silva Souto et al. [Citation75] reported that hypnograms derived from cEEGrid only and cEEGrid + EOG channels both showed a substantial agreement (kappa coefficient 0.67 and 0.75 respectively) compared to the Fpz + EOG electrodes (not a full PSG). Notably, the authors found that grapho-elements (like K-complexes and sleep spindles) were best represented by electrode combinations pointing toward Fpz (front facing) with an average correlation coefficient of 0.65. Likewise, this provides data that demonstrate the feasibility of sleep staging using ear-EEG within a home environment, although future studies evaluating its efficacy against the gold standard full PSG will be necessary.

Of all the devices, Ha, et al. [Citation72], Jorgensen, et al. [Citation68], Pham, et al. [Citation76] and da Silva Souto et al. [Citation75] reported devices, which had wireless capability. This is a key factor for unobtrusive monitoring and user comfort that will need to be addressed in future works.

The concept of multimodal signal acquisition for sleep monitoring was also reflected in three studies, which used EEG and EMG signals together for microsleep/sleep detection. These will be discussed under the EMG section below.

3.10.4. Brain computer interface

BCI describes a technology involving connections between the brain and external devices [Citation77]. It involves recording user brain activities and translating them into digital commands that can be applied for numerous functions, such as alternative means of communication, touch-free movement and neuroprosthetics [Citation78]. In this field, both Ahn, et al. [Citation79] and Wang, YT, et al. [Citation80] demonstrated devices that could monitor SSVEPs, with Ahn, et al. reporting 79.9% accuracy and information transfer rate (ITR) of 11.03 bits/min in an offline BCI visual target experiment and Wang, Y.T. et al. reporting 87.92% accuracy and an ITR of 16.6 bits/min. Additionally, Kaongoen, et al. [Citation81] developed a silicone earpiece that could classify auditory P300 signals (this is an event related potential elicited by brain recognition of odd-ball stimuli) with an accuracy of 95.61% and ITR of 2.9685 bits/min.

Kaveh, et al [Citation82]. reported a wireless, dry electrode earpiece that could record a mean alpha modulation that outperforms state-of-the-art dry electrode in-ear EEG. In this case, the use of dry electrodes, wireless connectivity and user-generic earpiece shows promise for unobtrusive, comfortable (dry) and portable applications.

Schroeder, et al. [Citation83] fabricated an External and Internal analog-digital converter (and foam) ear-EEG prototypes and demonstrated good performance against existing consumer products Neurosky MWM® and Muse®, with an image recognition classification accuracy of 97.78% and 96.13% respectively, compared to 96.64% (Neurosky MWM) and 97.32% (Muse).

Lee, JH et al. [Citation84] demonstrated a carbon nanotube polydimethylsiloxane (CNT/PDMS)-based canal-type ear electrodes (CEE) which could record multiple EEG paradigms, including alpha rhythms, N100 AEP, SSVEP, and ASSR. A wearability survey has also shown that the CNT/PDMS CEE was comfortable and felt like a standard headphone.

In terms of manoeuvrability applications, Wang, K.J. et al. [Citation85] demonstrated a wireless, dry electrode device EXGBuds® which could classify eye and facial gestures with approximately 95% accuracy, showing satisfactory performance in applications like controlling a camera-mounted robot in combination with a VR headset and hand gesture recognition technologies, or integration with a wheelchair to allow touch-free directional movement [Citation86].

The main limitations of BCI are the limited spatial resolution that reduces the effect of using common-spatial-filtering techniques, inherently lower SNR of ear-EEG signals and contamination with motion artifacts that can impair classification accuracies. Moreover, most EEG classifiers are non-adaptive, which means that the user will have to adapt instead to the classification methods of the machine which is time-consuming and tedious [Citation87]. Moreover, there are challenges with maintaining electrical coupling to the skin without gels, as dry-electrodes tend to have higher electrode-body impedance which must be taken into consideration when designing an instrumentation amplifier [Citation88]. This is because the formation of a double layer between the electrode and the body/skin is limited by the amount of moisture from the body condensed on the electrode surface [Citation88].

3.10.5. Identification/authentication

Ear-EEG has also been investigated for use in biometric authentication. Merrill, et al. reported an earpiece that achieved 0% false acceptance rate, 0.36% false rejection rates and 99.82% accuracy through a method using participants ‘passthoughts’ (a list of simple and personal mental tasks such as imagining a portion of a favorite song or relaxed breathing) for authentication. This study was only conducted in 7 subjects, however. Similarly, Curran, et al. [Citation89] utilized a list of mental tasks (like imagining a face with eyes closed or relaxed breathing) for authentication in 12 subjects and reported an accuracy of 72% for within-participant analyses and 80% accuracy in between-participant analyses. Moreover, Nakamura, et al. [Citation90] reported an experiment involving 15 subjects using robust power spectrum density and autoregressive features to identify individuals. They demonstrated an authentication accuracy of 95.7%. All the studies presented here had small sample sizes and would benefit from larger validation studies to further evaluate device performance.

3.10.6. Attention monitoring

2 devices have shown promise for use in attention monitoring. Bleichner, et al. [Citation91] used cEEGrid technology and demonstrated a method by which the direction of attention could be decoded significantly above chance (50%) for at least 16 out of 20 participants with a median of 66% and range of 57% to 85%). Additionally, Jeong, et al. developed a machine learning algorithm that could discriminate between attentive and resting states with approximately 81.16% accuracy.

3.11. EMG

3 studies documented the recording of EMG signals from hearable devices as shown in . 2 of these were from the same author. It has primarily been used in conjunction with EEG signals for sleep/microsleep detection via algorithms. The EMG signals were recorded from within the ear canal (n = 2) and behind the ear (n = 1). The main outcomes discussed were classification accuracy (n = 1), Raw graphs of EMG signals (n = 1) and sensitivity and specificity (n = 1). Generally, all studies reported good extraction of EMG signals. Nguyen, et al. [Citation92] concluded that their device was promising to monitor brain activity (EEG), eye and facial muscle signals with reasonable fidelity, with a 94% accuracy on sleep-stage classification on 8 subjects. Similarly, Pham, et al. [Citation76] reported an average precision of 76% and recall of 85% when comparing their device, WAKE, with gold standard devices for microsleep detection in 19 sleep-deprived and narcoleptic subjects.

3.12. Other signals

shows the 5 studies that record signals not directly covered by the above sections. In summary, 2 studies reported using an accelerometer for measurement of physical activity/posture/gait asymmetry, 1 study reported a device that measured cephalic blood flow and beat-to-beat systolic blood pressure, 1 study used PPG methods to identify speech and 1 study utilized BCG to monitor stroke volume. Main outcomes were highly heterogenous and depended on the type of signal measured. Accelerometry was evaluated for gait asymmetry and for measurement of physical activity/energy expenditure. Linear correlation (n = 1) was used for cephalic blood flow measurement, classification accuracy (n = 1) for speech detection via PPG and BCG waveform analyses (n = 1) for stroke volume monitoring.

Table 11. Device characteristics and study outcomes – Other signals

For ear accelerometry studies, 1 study [Citation93] reported that their earpiece could predict free-living walking velocity with substantial accuracy, allowing reliable monitoring of free-living physical activity and its associated energy expenditure. Additionally, the authors of another study [Citation94] found that gait cycle time and step-period asymmetry measures from the ear sensor had good matching with those of the treadmill, concluding that their device could potentially be used to replace expensive equipment used for rehabilitation monitoring.

Regarding cephalic blood flow measurement, the authors [Citation95] found a positive correlation (r = 0.557) between the drop ratio of cephalic blow flow and systolic blood pressure upon rising to an erect position, and that the blood flow drop ratio was significantly associated with dizziness (p = 0.023). This suggests that the device could potentially be used to evaluate for transient decreases in systolic blood pressure and cerebral ischemic symptoms during postural change to an erect position.

The study on speech detection [Citation96] obtained an accuracy of 83%. However, the study only tested a limited vocabulary of Chinese words and from 1 subject, thus more research will be necessary to validate the detection algorithm against spoken words in other languages and in a variety of subjects.

4. Conclusion

This narrative synthesis describes the main outcomes of hearable devices/prototypes used in the acquisition of physiological signals in the past 20 years. To the authors’ knowledge, this is the first review on this topic, and a majority of studies are on prototype devices. Across 92 studies, most have demonstrated good outcomes of recording a range of physiological signals from the ear with reasonable fidelity compared to conventional means of measurement. This is also true for studies describing classification accuracy and those which used Bland-Altman and Kappa agreements. The commonly proposed applications are cardiac/seizure monitoring, brain–computer interface and sleep staging. However, only seven studies have reported comfort outcomes of their devices in subjects. This is a key aspect that determines the applicability and wearability of such hearable devices in the context of ubiquitous monitoring.

The heterogeneity in protocols and reference devices make pooled evaluations of outcomes difficult. Furthermore, most studies were preliminary, pilot or proof-of-concept studies that were underpowered to demonstrate significant effects. The risk of bias assessment revealed that quality varied considerably between studies, with numerous studies showing a high risk of bias from subject selection due to small sample sizes. Moreover, the reference devices used to compare measurements were highly heterogenous between studies that made cross-study comparisons difficult and presents as a potential risk of bias. Several studies also did not report using a comparator.

5. Expert opinion

Most studies (74 out of 92) described measuring outcomes within a controlled laboratory environment or within a single-hospital setting (9 studies) as opposed to real-life settings (9 studies). This means that the full impact of artifacts from daily activities like walking or facial muscle electrical activity has not been fully evaluated in many studies. Given the generic consensus that weaker signals are recorded from the ear than in body locations closer to the signal of interest, this serves as a major roadblock toward device practicability that needs to be addressed. 3 studies reported a means by which their devices could denoise real-world motion artifacts and environmental noise. These included methods which required additional sensors like accelerometry [Citation97] and use of electret condenser microphones [Citation6]. These methods require additional computation and electrodes to be placed on the device, which may impact its wearability and practicability. Interestingly, Pham, et al. [Citation76] described a novel threefold cascaded amplifying technique that could alleviate motion and environmental artifacts through hardware optimization and demonstrated its feasibility by showing it could denoise by 9.74–19.47 dB in different practical situations like walking and driving.

Connectivity is also an important factor in ensuring device portability. Forty-three out of 92 studies reported devices that had wireless potential, these ranged from the use of low energy Bluetooth to short range radio connections. A few studies that used wired devices also noted the need for future work on wireless capability. Some studies did not report connectivity outcomes. Adding such wireless capability also comes at an expense of battery life, which favors the use of low energy wireless methods. Notably, battery life was a measure reported only in 8 studies – these ranged from 5 to 168 hours. Hearable devices will need to account for the appropriate battery life for their intended purposes.

Multimodality was also a feature investigated in 18 studies. Not all studies recorded all signals from the ear, with some complementing ear-sensing with sensors attached to other parts of the body. The benefits of multimodality have been demonstrated in numerous applications, such as in the monitoring of cardiac parameters, exercise monitoring, stress monitoring, sleep staging and reduction of motion artifacts. Considerations will have to be made about the hardware and physical limitations of multimodal devices, which may take up more space physically and require greater computing power.

Hearable devices have also shown considerable synergies with hearing aids. Some of the proposed devices were manufactured in the shape of a hearing aid for better anchoring to the ear [Citation49]. This opens the possibility of integrating sensors within current hearing aids to enable physiological signal monitoring in hearing impaired individuals [Citation25]. Ota, et al. [Citation29] demonstrated a similar concept, presenting a device that could record tympanic temperature and function as a bone conduction hearing aid simultaneously. Moreover, Christensen, et al. [Citation98] reported a device that utilized ear-EEG to estimate auditory threshold levels in individuals with sensorineural hearing loss, concluding that with further refinements, it could be possible for a device that could automatically alter its auditory processing according to progressive hearing loss of its user. Other possible synergies include BCI applications such as autonomous audio steering, which can improve hearing in background noise by controlling the direction of hearing aid sound amplification based on user brain activity [Citation99].

Future work will have to address several issues concerning the applicability of hearables in daily life. First, the development of multimodal sensing hearables can provide opportunities for a comprehensive sensing system that can report health outcomes remotely. It is particularly important to develop sensors with better SNR profiles. Second, improvements to connectivity and battery life issues will improve the portability and practicability of such devices. Third, it is necessary for more in-depth research on methods of denoising motion and environmental artifacts that can impair signal quality. Fourth, robust dry electrodes that offer low electrode-body impedance and are motion tolerant have to be optimized. Finally, comfort and wearability outcomes must be taken into consideration – this includes among others, material choice for devices and the use of dry instead of wet electrodes.

Declaration of interest

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Funding

This paper was not funded.

References

Download PDF

Related research

People also read lists articles that other readers of this article have read.

Recommended articles lists articles that we recommend and is powered by our AI driven recommendation engine.

Cited by lists all citing articles based on Crossref citations.
Articles with the Crossref icon will open in a new tab.