A convenience sample was recruited by word of mouth through friends and colleagues of the authors (H Shimura and SO) from communities across several prefectures in Japan, including both the National Capital Region (Tokyo, Kanagawa, Saitama, and Yamanashi) and other regions (Nagano and Gifu). Eligibility criteria included age between 20 and 89 years, no gait abnormalities, having one’s own smartphone, and willingness to wear activity monitors during the monitoring period. Participants aged 70 years and older were recruited from the National Capital Region, while those aged 70 years and younger were recruited from both the National Capital Region and other regions. Efforts were made to achieve a balanced distribution of participants across 10-year age groups (20s to 80s), and to recruit men and women in roughly similar numbers within each age group wherever possible.

Participants who attended the introductory session were given a full explanation of the purpose and potential risks of the study. They signed an institutionally approved informed consent form before data collection. This study was conducted after obtaining approval from the institutional review board of the Tokyo Metropolitan Institute for Geriatrics and Gerontology (R22-013). It was part of the Smart Watch Innovation for Next Geriatrics and Gerontology (SWING) study, an intensive longitudinal study designed to investigate the use of behavioral data, such as physical activity and sleep monitored by wearable devices, to promote health among older adults.

The data from the TDKsw and the AG were merged using the participant ID and the timestamp as keys. In doing this, records (minute-by-minute observations) were excluded if there were any missing values (eg, due to TDKsw device malfunction) or if the AG data were classified as nonwear [19,20]. There was no validated algorithm for TDKsw to identify nonwear; therefore, it was deemed that the participants followed the instruction to wear or remove both AG and TDKsw simultaneously. A valid day was defined as having ≥10 hours of accumulated AG wear time (convention for compliant wear time) [21] with an accumulation of ≥100 and <50,000 steps (for exclusion of outliers) [22] from both AG and TDKsw.

Data from each device were aggregated to daily step counts and time spent in MVPA (≥3 METs for the TDKsw and ≥1952 vertical axis activity counts/minute for the AG [13]) for each participant. Also, each valid day was classified as either an active day or an inactive day, using the data of step counts and time spent in MVPA from each device; valid days with ≥10,000 steps/day [23,24] or ≥21.4 minutes MVPA per day (corresponds to ≥150 minutes MVPA per week [5]) were classified as active days.

Baseline characteristic data were summarized as mean (SD) for continuous variables and as frequency and percentage for categorical variables.

The dataset in this study had a hierarchical structure, with daily aggregated data (maximum 7 valid daily observations per participant) nested within individuals. Therefore, multilevel models were used to appropriately account for this clustering. An intercept-only multilevel model was employed to estimate the overall mean (SE) of step counts and time spent in MVPA. Bland-Altman plots [25] were used to visualize the agreement between the TDKsw and the AG for each measure across each day where valid data were available. An intercept-only multilevel model was also applied to estimate the overall mean of the paired differences between the devices (ie, fixed bias) and the 95% limits of agreement (LoA) [26-28]. Proportional bias was evaluated based on the slope parameter of a multilevel model that included the average value between the devices for each outcome as a fixed effect.

To address potential model misspecification, robust inference was applied to all multilevel model analyses [29,30]. In addition, residual normality and homoscedasticity were assessed using diagnostic analyses for the Bland-Altman analysis models.

The proportion of agreement in classification as active day or inactive day was assessed using κ statistics. For interpretation of the κ statistics, we followed the ratings suggested by McHugh [31]: none (0.00‐0.20), minimal (0.21‐0.39), weak (0.40‐0.59), moderate (0.60‐0.79), strong (0.80‐0.90), and almost perfect (above 0.90).

Statistical analyses were performed using SPSS 29.0.1.0 (IBM Corporation) and R (version 4.5.0; R Foundation for Statistical Computing). The lme4 and clubSandwich packages in R were used for multilevel modeling [32,33].

This study assessed the agreement between a wristband-type wearable IoT device, the TDKsw, and a research-grade activity monitor, the AG, worn on the waist, in estimating the daily accumulation of step counts and time spent in MVPA. Agreement was investigated at the day level under free-living conditions in young to older adults. Our results showed some discrepancies between the devices. Compared with the AG, the TDKsw underestimated step counts by approximately 1200 steps per day, whereas it overestimated the time spent in MVPA by 5 minutes per day. These differences seemed to be mostly constant over the measurement range. In addition, the results indicated moderate agreement in classifying whether a given person was active or not on that day; the levels of agreement were similar whether we used the data of daily step count or time spent in MVPA.

Our findings seemed to be counterintuitive, given that the daily step counts are expected to increase as people spend more time in MVPA [34]. A possible explanation for why step counts were underestimated, whereas MVPA was overestimated in this study, is that the MVPA measured by the TDKsw was not explicitly identified as walking. However, the exact reason remains unclear because this study used data that were preprocessed by the manufacturers’ proprietary algorithms.

The differences between the two devices are not trivial, equivalent to approximately 12 minutes of walking per day (assuming 100 steps per minute [35]) and one-fifth of the recommended level of physical activity (150 minutes of MVPA per week [5]). In addition, our results may have important implications for promoting physical activity. That is, some TDKsw users may be identified as active on that day even if in fact they are not, and vice versa. Such misclassification, which stems from the fixed bias observed in step counts and MVPA estimates, may be considered unacceptable for research or clinical purposes when assessing the amount or intensity of physical activity at a given point in time. However, when examining within-individual fluctuations of physical activity behavior over time, such differences would not substantially affect the interpretation of the data. The key advantage of the TDKsw is that it enables long-term monitoring and remote collection of large amounts of data by using IoT technology.

Although the κ statistics indicated moderate agreement, whether this level is acceptable depends on the intended use. As noted above, it may be insufficient for point-in-time assessments in clinical or research settings, but potentially acceptable for tracking temporal trends within individuals.

Comparing the present results with other free-living studies was challenging because there were differences in terms of methodology and the devices used. The only study to date that has investigated the accuracy of the TDKsw is Kimura et al [11], which reported a significant correlation (r=0.987; P<.001) between TDKsw step count and 9 minutes of video observation in older adults (n=20). Other related studies include research on the agreement between wristband-type activity trackers and waist-worn research-grade devices on a day-to-day basis. Chu et al [36] (n=35 males and n=69 females; median 31.0, IQR 26.0‐42.8 years; 7 days) reported that the Fitbit Flex overestimated steps/day relative to the waist-worn AG (median difference 1300 steps/day). Mikkelsen et al [37] (n=10 males and n=31 females; mean 47.6, SD 10.4 years; 7 days) reported that compared with the waist-worn ActiGraph GT3X, the Fitbit Charge 2 measured 1492 more steps per day (LoA −2250 to 5234), and 31 fewer MVPA minutes per day (LoA −132 to 71) with increasing difference as time in MVPA increased, suggesting that counterintuitive results might be obtained, as observed in our study. It should be noted that there was variability among the studies in the choice of ActiGraph count data (vertical axis or vector magnitude) and data-processing algorithms, which might have contributed to the mixed findings.

This study has several strengths. First, it included a relatively large sample of participants with a wide range of ages, equal gender balance, and various sociodemographic backgrounds. Second, the study was conducted under free-living conditions, which improved the ecological validity and the generalizability of the findings to real-life settings [38]. Third, we kept the observations (ie, daily aggregated data) as repeated measures for each participant and used multilevel analysis to calculate the LoA. This provides a more accurate representation of the actual agreement by lowering the risk of producing LoA that is too narrow [26]. Earlier agreement studies have rarely used multilevel analysis, despite the nested structure of the repeated measurement [27,37].

The major limitation of this study is the use of non–gold standard methods for comparing steps and MVPA. Having said that, there is no feasible gold standard method for assessing these measures in free-living conditions. In this study, we regarded the waist-worn AG as the best reference because the algorithms it uses to estimate METs and step count have been validated [13-15] and frequently used in earlier studies. Furthermore, although we instructed the participants to wear and remove both TDKsw and AG at exactly the same time, there is no way to ensure they followed this instruction. In addition, because the algorithms used in the TDKsw are proprietary and not publicly available, we were unable to provide a direct comparison of the algorithms.

Differences were found in estimating the daily accumulation of step counts and time spent in MVPA when comparing outputs between the wrist-worn TDKsw and the research-grade waist-worn AG. It was noted that the TDKsw underestimated step counts and overestimated the time spent in MVPA, which might lead to misclassification of active days and inactive days. Wearable IoT devices such as the TDKsw are expected to be used for large-scale epidemiological studies and for promoting physical activity through long-term monitoring. When using TDKsw data over relatively short periods, caution is warranted, as misclassification—particularly when compared with data from research-grade monitors such as the ActiGraph—may affect feedback or goal-setting. Adjustments to classification thresholds may be necessary depending on the intended application.