Sixty middle-aged and older adults (64.0 ± 8.0 years, BMI 24.4 ± 2.8 kg · m⁻², 48% females) participated in this study (Table 1). Participants were recruited through word-of-mouth, advertisements on the university campus, and in local retirement homes. All 60 participants recruited also completed the study.

To qualify for inclusion, subjects were required to be 45 years or older, in good health, and not taking any medication (except for some allowable health conditions and medications for participants aged 65 years and above—refer to exclusion criteria). Additional requirements included a BMI greater than 18.5 kg · m⁻², the ability to perform all activities outlined in the study protocol, normal hearing or at max. mild hearing loss, and willingness to comply with the study's procedural guidelines (i.e., to refrain from intense exercise 48 h before testing; to abstain from any exercise 24 h before testing; to ensure a minimum of 7 h of sleep on the two nights before testing; to avoid alcohol on the evening before testing, as well as on the day of testing; to not consume any caffeinated food or beverages before testing on the day of the experimental visit; and to arrive fasted after at least 10 h without food intake for visit 2).

Subjects were excluded if they had a history of heart, cardiovascular, metabolic, or neurological disease (incl. seizures and cognitive impairment), a biomechanical dysfunction affecting the ability to perform all activities, ear canal pathologies, moderate or severe hearing loss, an existing implanted medical device that may interfere with data collection, skin allergies or sensitivity to materials or devices used in the experiments, and—specifically for the participants aged 65 years and above—medication influencing heart rate or EE and neurological, orthopedic, rheumatologic, or metabolic disorders influencing upper or lower limb function.

We aimed to recruit 30 adults aged between 45 and 64 years and 30 participants aged 65 years or older, with each group containing a BMI distribution encompassing normal weight, overweight, and obese categories and males and females in equal numbers (≥40% for each sex). The study was approved by the local ethics committee of ETH Zurich (EK 2022-N-44). Every participant gave written informed consent in accordance with the Declaration of Helsinki before participating in the experiment.

Thirty middle-aged participants of median [range] age 59 [46–64] years (57% females) and 30 older participants [71 (65–78) years; 40% females] completed the study (see Table 1). On average, older participants had lower bone mineral density and higher systolic blood pressure and pulse wave velocity compared to middle-aged participants. There were no significant differences between the calibration and validation groups for the variables tested (see Supplementary Table S1).

Figure 3 shows the reference EE, as measured by the ergospirometric device, in METs for all activities. The median METs for most activities fell within the expected range (i.e., light intensity), except for some household activities, such as vacuum cleaning, cleaning with a mop, dust wiping, and hanging laundry, which were categorized into the moderate intensity category.

Mean bias for the Audéo class-estimated model was −0.23 METs (see Figure 4), differing significantly from zero (p = 0.031). Lower and upper LoA were −3.56 and 3.10 METs, respectively. MAE amounted to 1.19 METs.

The performance metrics for Audéo (class-estimated) vs. Fitbit vs. Garmin are shown in Table 3 and Figure 5. Note that for this comparison, two participants had to be excluded from the analysis as their Fitbit data could not be downloaded. To facilitate the comparison to the 6 min data from Fitbit and Garmin, the three stages of activity 18 (Cycle_ergo) were averaged into a single value.

Performance metrics for Audéo vs. ZurichMove (both class-known) are shown in Table 4. Note that with the approach used in this study, it was not possible to build an EE estimation model for the wrist sensors because of the missing linear relationship between counts and METs in the wrist data for the majority of the activity classes.

Across all activities, the Audéo prediction model (class-estimated) demonstrated a minor underestimation of 0.23 METs, along with a wide LoA of ± 3.33 METs. These findings align with other studies that used ear-level accelerometers to predict EE. For example, Atallah et al. (26) used an ear-worn inertia sensor during 11 ADLs (lying, standing, computer work, vacuuming, stairs, walking, running, cycling, and rowing) in 25 healthy young subjects. This resulted in an overall mean bias of 0.04 METs and a LoA of ± 3.65 METs. Similarly, Bouarfa et al. (27) developed an EE prediction model using 25 young subjects involving 10 ADLs (same as in Atallah et al., but without rowing). They reported a mean absolute deviation below 1.2 METs, i.e., identical to the MAE found in this study.

Good accuracy and low precision are typically also observed in other studies that use research or consumer devices to predict EE. For example, Crouter et al. (40) investigated the performance of the Actigraph and Actical devices (both worn on the waist), and the AMP-331 monitor (worn on the ankle) during 18 different leisure and sporting activities (e.g., lying, computer work, vacuuming, walking, running, stairs, basketball) to predict MET in 48 younger to middle-aged adults. The Bland-Altman plots show mean biases of about −0.5, −1.0, and −2.5 METs for Actigraph, Actical, and AMP-331, respectively, and LoAs of about ± 3.5, ± 3.0, and ± 4.0 METs. This data also aligns with our finding that sensors located closer to the body's center of mass (including the ear), tend to outperform devices positioned on the limbs when the goal is to predict EE in a wide range of ADL. Literature supports this observation, indicating that wrist-worn accelerometers generally yield less accurate EE estimates compared to those worn on the hip (16–18). In theory, this difference could be attributable to different device grades, as research devices are typically being worn on the hip or chest, while consumer devices are worn on the wrist. For example, Chowdhury et al. (41), comparing consumer monitors (Microsoft Band, Apple Watch, and Fitbit Charge HR) with a research device (Actiheart) during 9 ADL in 30 young subjects, concluded that consumer devices are not yet at the level of the best research devices: Mean bias and 95% LoA values for consumer monitors amounted to −0.55 ± 3.65 METs while for Actiheart they were 0.51 ± 2.11 METs, respectively. However, this finding is more likely confounded by the sensor location: while the Actiheart was worn on the chest, all other sensors were worn on the wrist. Indeed, when consumer and research devices are worn at the same location, differences tend to disappear. For example, a meta-analysis investigating the accuracy of wrist-worn devices found no significant overall differences between research and consumer devices in estimating EE (6). This finding is consistent with the notion that the latest generation of consumer devices incorporates similar technology to that of established research devices (41). Our findings thus contribute to the existing body of literature by demonstrating that an accelerometer integrated into a hearing aid performs comparably to a research device placed on the hip, despite exhibiting a broad LoA—a characteristic consistent with other studies in the field.

Our observation that the Audéo model marginally surpasses Fitbit and Garmin in performance should be interpreted with caution. We developed and validated our model with the assumption that 1 MET = V̇O₂ measured individually at rest. Reanalyzing the data using 1 MET = 3.5 ml · min⁻¹ · kg⁻¹, mean biases for Fitbit and Garmin change from about −0.6 METs to about −0.1 METs. Precision also improves but remains slightly inferior to Audéo. More importantly, the calibration and validation groups are very similar in this study. This represents a possible disadvantage as Fitbit and Garmin were likely calibrated in a population that is not as comparable to the one used in this study. For example, EE estimation in an older population is compromised when a model is trained on younger subjects and improves substantially when the model is trained on older participants (unpublished data). This study shows that the EE estimation is slightly better in the older subgroup and that training the model on the older subgroup leads to better performance when evaluated in the older subgroup than when the model is trained on the middle-aged subgroup (data not shown). Due to age-related physiological and functional changes, e.g., changes in speed of movement, gait mechanics, and body composition (42–44), algorithms validated in younger adults may not accurately apply to older age groups (18). Whether Audéo outperforms Garmin and Fitbit (or other consumer sensors) in an independent sample in a laboratory setting or under free-living conditions, needs to be tested according to existing validation frameworks [e.g., (30)].

The broad LoA observed in this study for Audéo (and other sensors) can be attributed to the underestimation and overestimation of EE for distinct activities. Specifically, EE associated with ascending activities (e.g., climbing stairs, walking uphill) and stationary activities (e.g., cycling on an ergometer, squatting) was commonly underestimated; whereas EE for descending activities was overestimated, as illustrated in Figure 6 and Supplementary Figure S1. This phenomenon is likely attributable to the disparity between accelerometer counts and actual EE for certain activities—namely, those involving low accelerations with high EE (e.g., cycling) and those with high accelerations but low intensity (e.g., descending stairs). The variability in accuracy for single activities has also been demonstrated in reviews on the topic (6, 45) and is a known limitation when using accelerometers to predict EE. When comparing different Actigraph equations, Crouter et al. (40), for example, concluded that no single equation is valid for the EE estimation of all activities and that equations work best only in the activity subgroup they were developed. In this study, mean bias and LoA increased with activity intensity, a trend likely attributable to the nature of intense activities—predominantly ascending and stationary—which are associated with the largest errors.

In the model employed in this study, we operated under the assumption of a linear relationship between counts and METs within the different activity classes. Although the adoption of this two-step approach—initial classification followed by application of a class-specific model—is acknowledged to enhance estimation performance (46), it is not without its challenges. For example, for some activity classes, there were no or only weak correlations between counts and METs, rendering EE prediction challenging. Similarly, in a meta-analysis on the validity of the Actigraph device (worn either on the hip or wrist) for measuring EE in healthy adults, Wu et al. (45) found no correlation between activity counts and EE for some activities, e.g., during cycling, standing, walking at a moderate speed, and fast running (47). Because of the missing linear relationship between counts and METs in the wrist data within most of the activity classes in this study, we refrained from developing an EE-estimation model for ZurichMove sensors on the wrist. It is important to note that our methodology was specifically tailored to enhance the EE prediction for a sensor integrated into a hearing aid with limited memory and computational power, suggesting that different strategies might have been effective for wrist-worn sensors or devices with more computational power.

Despite this variability in accuracy for single activities, the errors might cancel out under the assumption that a wide range of activities are performed over a day. Indeed, the calculated mean bias for Audéo for a 10 h wear time amounted to an overestimation of 61 kcal (∼5% of 10h-EE). Similarly, Härtel et al. (48), investigating the kmsMove sensor (worn on the hip) during rehabilitation activities over 7 h in 7 middle-aged adults, found an average underestimation of 14 kcal. Berntsen et al. (49), using Actigraph (worn on the hip), ActiReg (chest and thigh), and ikcal (chest) monitors during free-living lifestyle and working activities over 2 h in 20 younger and middle-aged adults, found mean biases ranging from −34 to −111 kcal. Despite good overall accuracy, drawing inferences for individual subjects remains challenging, as evidenced by our data (LoA 302 kcal) and Härtel's and Bernsten's findings (LoA ranging from 261 to 397 kcal). Nonetheless, it can be argued that the level of accuracy and precision shown by Audéo, and other devices is acceptable in the context of health interventions. For example, in order to achieve weight loss, energy intake should be reduced by about 500–1,000 kcal a day (50). This change is higher than the reported LoA of this study, hinting at the potential for detecting such changes. Furthermore, the observed significant within-subject correlation between an increase in intensity from walking to running, as measured by both the ergospirometric device and the Audéo sensor, underscores the device's ability to detect MET changes. However, this detection is relative rather than absolute, as indicated by the regression line's deviation from the identity line. This deviation from the identity line is the result of an overestimation of METs during level walking activities.

Our findings from a sensor located at ear-level open the door for other application fields in this age range, e.g., IMU integration in in-ear headphones worn during sporting activities (29). Accuracy and precision for the Audéo sensor, and in general for accelerometers aiming to predict EE in a variety of activities (including ADLs, sporting activities), can be improved by the incorporation of heart rate (HR) data, as evidenced by O’Driscoll et al.'s (6) meta-analysis. This is particularly relevant for activities exhibiting a disparity between accelerometer counts and actual EE. It would be interesting to explore whether an ear-worn device that can detect HR via photoplethysmography can provide better performance. In addition, incorporating an altimeter has the potential to improve accuracy given that the largest errors in this study were found for ascending and descending activities. This is supported by Duncan et al. (51) who found that using barometers and global positioning systems improved EE estimation accuracy during field-based activities, compared to accelerometry alone, by 11%.

This study was performed in a laboratory setting, following the framework proposed by Keadle and colleagues (30). However, it is known that the accuracy of models validated under laboratory conditions decreases when applied in free-living conditions (21). Also, the modeling approach adopted in this study did not enable the development of an EE estimation model for the ZurichMove sensors positioned on the wrists, thereby precluding a comparative analysis of their performance with the Audéo sensor.

This study demonstrates that an accelerometer integrated into a hearing aid (Audéo) can accurately estimate EE across a broad range of ADL in a middle-aged to older population. However, the precision of these estimates is limited, making personal-level inferences challenging, though still offering valuable insights on a population level. Moreover, the Audéo sensor's performance in EE prediction, using the same modeling approach, matched that of a research device worn on the hip and slightly outperformed two wrist-worn consumer monitors. This indicates that an accelerometer integrated into a hearing aid can serve as an equivalent alternative for monitoring physical activity. This opens the door to unobtrusive evaluation of energy expenditure during daily life in older individuals that are already using hearing aids, and eventually to implementation of personalized interventions promoting healthier aging.