The participant sample size was calculated by G*power. The significance level α was set to 0.05, the power 1 − β was set to 0.8, and the effect size d was set to 0.5. As a result, 27 participants were required.

The participants were healthy women in their 20 s to 30 s. The inclusion criteria were having a BMI of higher than 18.5 and less than 25, and being able to answer questions in Japanese. A BMI of 18.5 to 25 is a normal body mass. Approximately 70% of his 20 - 30 s women are within the BMI range of 18.5 - 25 [25] . The exclusion criteria were having a smoking habit, having an underlying disease, and using drugs that influence autonomic nerve activity. Stress level is calculated from heart rate variability. Smoking, including e-cigarettes, affects autonomic nervous system activity [26] . Autonomic activity is known to be affected by oral medications such as antidepressants [27] . The participants were selected by checking with the physicians enrolled in the study group regarding their oral medications.

We recruited research participants through the research group’s website. A summary of the research was posted on the website. The participants who were interested in the research registered their e-mail addresses in the application form created using Google Forms. This application form also included a screening item that confirmed the participant’s selection criteria.

The researchers contacted the participants via the e-mail addresses collected through Google Forms. An overview of the research was explained at a later date to the participants face-to-face, and the participants signed a consent form when consenting to participate in the research. Surveys were conducted among those from whom a consent form was obtained.

The survey was conducted from December 2021 to June 2022.

The following three smartwatch models were used in this survey: Mi smartband 6 (Xiaomi Corporation Inc., China), vivosmart^®4 (Garmin Ltd., USA), and Band 6 (Huawei Technologies Co., Ltd., China). The participants wore the three smartwatches simultaneously on their wrists for 48 hours continuously. They did not remove their smartwatches while bathing or sleeping. The smartwatch wearing method was left up to the discretion of the participant, so the smartwatch wearing order also differed by participant. Many of the participants wore the three smartwatches on the wrist of their non-dominant side. Some participants wore two smartwatches and one smartwatch on each wrist.

In order to take measurements in an environment representing everyday life, surveys were conducted while avoiding days of competitions or events. The participants were allowed to continue activities that they regularly participated in, such as extracurricular and club activities; their habitual exercises were not restricted. The participants recorded activities such as eating, exercising, and learning on an activity chart.

For the participant attributes, we obtained the age, height, weight, and oral medications of the participants.

We extracted and analyzed five datasets from the data measured by each smartwatch: heart rate, stress level, number of steps, distance, and sleep time.

For stress level, number of steps, distance, and sleep time, the daily average values calculated by each smartwatch were used.

For the comparison of the heart rates between the three smartwatches, we used Spearman’s rank correlation coefficient ρ. Next, we created Bland-Altman plots to visually examine the heart rate agreement between the smartwatches. We calculated the mean difference and lower to upper limits of agreement for the heart rate. The lower to upper limits of agreement were calculated by mean ± 1.96 standard deviation (SD).

In order to compare the fitness trackers of each smartwatch, we calculated the mean absolute percentage error (MAPE). In this research, we targeted the five activity data of rest heart rate, stress level, sleep, steps, and distance, which could be measured and calculated by the three smartwatches. The mean error rate was determined as the absolute value of the true value minus the measured value. The true value for MAPE was the average of the data calculated by the three smartwatches. For the sleep time, bed time and wake time were entered in the activity chart, so the sleep time obtained from the activity table was taken as the true value. The absolute value obtained by subtracting the data calculated by the smartwatch from this true value was divided by the true value and multiplied by 100 in order to calculate the MAPE [17] .

Statistical analysis was conducted using IBM statistics SPSS ver. 29, with the significance level set to 5%.

This research was approved by the ethics committee of Tsukuba of University Medical School (approval number 1684). The research was explained to the subjects, and after written consent was obtained from them, we conducted the survey.

There were 25 applications for research participation. Of these, two declined to participate in the research, so 23 people were set as analysis participants.

The average age of the participants was 22.9 (±2.4) years. The average BMI was 20.8 (±2.1). Among the participants, five were taking oral contraceptives, and two were taking anti-allergy drugs.

We calculated the daily heart rate measured by the three smartwatches and compared them using Spearman’s rank correlation coefficient ρ.

A very strong correlation was observed between Mi smartband 6 and vivosmart^®4 (ρ = 0.828, p < 0.001) “Figure 1(a)”. A strong correlation was observed between vivosmart^®4 and Band 6 (ρ = 0.684, p < 0.001) “Figure 1(b)”. Similarly, a strong correlation was observed between Mi smartband 6 and Band 6 (ρ = 0.675, p < 0.001) “Figure 1(c)”.

Next, Table 1 and Figures 2(a)-(c) show the mean difference and lower and upper limits of agreement of the Bland-Altman plots that compare each smartwatch for each pair of models. The mean difference between Mi smartband 6 and vivosmart^®4 was the lowest at −1.74. This indicates that the heart rate error was small.

Figure 3 shows changes in heart rate and stress level calculated by each of a participant’s (ID-20) smartwatches for 24 hours. The time not calculated by the smartwatch was left blank.

Table 2 shows the daily resting heart rate, sleep time, stress level, number of steps, distance traveled, and standard deviation for the fitness trackers calculated

Lower to upper limits of agreement: mean difference ±1.96SD; SD: standard deviation.

MAPE: Mean absolute percentage of error; SD: standard deviation; rest HR: rest heartrate.

by the three smartwatches. For the sleep time, vivosmart^®4 had a high value at 6.58%. This was because vivosmart^®4 had a high sleep time error compared to the other two models. Next, for the resting heart rate, Band 6 had the lowest MAPE at 4.77%. This shows that the error with the average resting heart rate calculated by the three smartwatch models was small. For the stress level, Mi smartband 6 had a mean error of 55.22%. This shows that Mi smartband 6 had a high stress level error compared to those in vivosmart^®4 and Band 6. For the number of steps, Mi smartband 6 had an error of 17.80%, which was a small

value among the three smartwatch models. Similarly, the distance traveled was 18.25% for Mi smartband 6, which had a small error compared to the two other models.

The average BMI for women in their 20s is 21 in Japan [25] . The BMI can be said to be representative of Japanese women. Five participants were taking oral contraceptives. Two participants had hay fever. In Japan, cedar and cypress pollen increases in spring. The survey was conducted from December to June, which is winter and early summer. Therefore, the participants were taking anti-allergic medication internally. These oral medications are common. The participants of this study are considered to be similar to the general population of Japanese women.

The participants wore the smartwatches continuously for 48 hours, and we compared the heart rates measured by the smartwatches during their free activities. Mi smartband 6 and vivosmart^®4 were comparatively able to calculate the heart rate every minute. Meanwhile, Band 6 calculated the heart rate once every 5 - 10 minutes. Looking at the correlations between the heart rates calculated by each smartwatch at the time the heart rate was measured, we found a very strong correlation between Mi smartband 6 and vivosmart^®4. When visually looking at changes in the heart rate, vivosmart^®4 and Mi smartband 6 had sections where the heart rates were not in agreement. This was the same result as in a previous study that compared heart rate accuracy in laboratory environments. In that study, the device manufactured by Polar was used as a gold standard, and it was shown that devices manufactured by Xiaomi and Garmin had a generally accurate PPG [15] . Meanwhile, the agreement between vivosmart^®4 and Band 6 tended to be high in the sections where the Mi smartband 6 and vivosmart^®4 were not in agreement. According to a previous study that compared various models, the device manufactured by Xiaomi had the lowest heart rate accuracy [17] . In previous research, the Mi smartband 2 was used, and the model used in the present research was an improved version of this. Therefore, it is thought that the PPG accuracy improved compared to the device used in previous research.

From the correlation coefficient, vivosmart^®4 and Band 6 had a slightly higher heart rate agreement (ρ = 0.684) than Mi smartband 6 and Band 6 (ρ = 0.675). It can also be seen from the Bland-Altman plots that the mean difference was lowest between Mi smartband 6 and vivosmart^®4. Each of the Bland-Altman plots showed high agreement between Band 6, Mi smartband 6, and vivosmart^®4. It can be said from these results that the heart rate measurement method under free movement was consistent. This is similar to previous research that compared various device models [15] [22] . Therefore, the smartwatches used in the present study are able to measure heart rate at the same level even under free movement in daily life.

In this research, we calculated the MAPE, using the average value of the three smartwatch models as the gold standard. Results showed that the Band 6 manufactured by Huawei was the closest to the average of the three models. However, Huawei has a low agreement of continuous heart rate, and vivosmart^®4 tended to underestimate heart rate more than Mi smartband 6, so it is thought that Band 6 was the average value of smartband 6 and vivosmart^®4. In other fitness tracking indices, the stress level value in the Xiaomi device was the biggest outlier. When confirming the changes in daily activity, heart rate, and stress level, vivosmart^®4 calculated the stress level most often, and the stress level was calculated when the heart rate was calculated. The heart rate of Mi smartband 6 was measured to the same extent as vivosmart^®4, but the stress level was calculated irregularly. Band 6 calculated the stress level when the heart rate could be measured continuously. Previous research has also not reported consistent results on the validity of stress levels, and the algorithms in these devices for calculating stress levels are unclear [23] . A goal in stress monitoring using wearable devices is the generalization of stress models using various forms of machine learning [28] . There have also been studies on psychological and physical stresses [29] [30] , but further research on integrating these stress models is needed. In recent years, there have been not only linear analyses of autonomic nerve activity such as time domain analysis and frequency domain analysis, but also changes in nonlinear heart rate variability measurements [31] . This nonlinear heart rate variability has been studied using accelerometer and gyroscope signals [32] . These sensors are also installed in smartwatches, and it is assumed that future measurements of stress conditions will be from heart rate variability using these sensors. Examining the validity of stress levels requires capturing the autonomic nerve activity from linear and nonlinear analyses for a specified stress load and formulating a model equation that can calculate stress.

Research that examined the fitness tracking of sleep, rest heart rate, steps, distance, calories, and sleep in smartwatches manufactured by Huawei and Xiaomi reported that the MAPE of Xiaomi-manufactured devices was small [17] . In the present research, we showed that Mi smartband 6 had values closest to the average, with the exception of the stress level values. This is thought to be due to the fact that the Xiaomi-manufactured device took the average of data that calculated the fitness tracking in the Garmin- and Huawei-manufactured devices. It has been reported that the heart rate accuracy during cycling of Garmin-manufactured devices was lower than that of Xiaomi-manufactured devices [15] . Meanwhile, research that examined accuracy using the Mi band 4 manufactured by Xiaomi showed that the accuracy of the number of steps or heart rate under free activities was relatively high, and their usability in daily life has been proven [33] .

Presently, various smartwatches have been developed, and they have an influence on fitness tracking, such as data accuracy and user behavior [19] [24] . Wearers need to select a smartwatch that suits their living environment.

This study has several limitations. The first is that the target population is not a representative population. In this study, we recruited research participants through the research group’s website. The participants were students at the university where the research group is located. Therefore, the target population is limited. In order to ensure a diverse and representative sample, it will be effective in the future to use methods such as social media and advertising [34] . In the future, our research should examine ways to collect samples using these methods.

The second is that this research focused on women but did not consider their menstrual cycles. It has been reported that autonomic nerve activity changes according to the menstrual cycle [4] [9] ; therefore, in the future, we will need to create a research protocol that considers the menstrual cycle.

The third is that we were unable to examine fitness tracking other than heart rate with sufficient accuracy. We did not set an exact gold standard, since we calculated MAPE from the true value obtained from the fitness tracking of the three smartwatches and its error. As such, there is a need to set and examine the gold standard for sleep time, and to use pedometers and distance measurements for the number of steps and distance accuracy. In the future, there is a need to verify the accuracy and validity of physical activity using the accelerometers and triaxial sensors installed in smartwatches.