Inter- and Intra-Day Precision of a Low-Cost and Wearable Bioelectrical Impedance Analysis Device

Body composition, including fat mass, fat-free mass (e.g., muscle mass), and hydration status, is of interest to both the clinical and the general populations. The information from body composition measurements can be used, e.g., to predict the outcome and appropriateness of clinical interventions [1], while self-monitoring via digital health solutions, in general, has a positive influence on weight loss [2].

Several methods for measuring body composition exist, including computed tomography (CT), ultrasound, dual-energy X-ray absorptiometry, and dilution-measured total-body water. However, most of these methods are laboratory-bound, invasive, or need a considerable amount of resources [1, 3]. Bioelectrical impedance analysis (BIA) overcomes several of those limitations [4, 5]. In BIA, the impedance generated by sending weak alternating currents through the body at defined frequencies gives indirect information about the composition of the body [6]. Nowadays, with increasing technological progress, cheaper and smaller BIA solutions have entered the market making the method available to a broad audience while simultaneously improving usability. This will improve the implementation of BIA measurements into wearable devices. The prerequisite for any BIA solution within the wearable device is high precision. While the here used device’s measurement accuracy between 1 and 340 kHz has previously been validated against a reference device using several circuits modeling human tissue [7], its reproducibility in human subjects has not been investigated, yet. Thus, in this pilot study, we analyzed the measurement precision regarding the reproducibility of this cheap and commercial BIA solution suitable for integration into a wearable device. Intra-session, intra-day, and inter-day reproducibility of raw impedance values from three subjects at three different body locations (hand-to-hand, hand-to-torso, torso-to-torso), and for three different frequencies (6, 54, and 500 kHz) were reported.

Three healthy subjects (1 female: 178cm, 68.4kg, 25 years; 2 male: 182/185 cm, 86.0/88.7 kg, 31/28 years) participated in the pilot study after they were informed about the scope of the study and gave informed written consent. We collected data over three frequencies (6, 54, and 500 kHz at 63.99 μA) using MAX30009EVKIT (Analog Devices, Inc., Wilmington, USA) for three different electrode locations (wrist-to-wrist, wrist-to-torso, and torso-to-torso), replicating typical location for wearable devices (e.g., watch, belt). The frequencies were selected to cover a wide frequency band to investigate the device’s potential for bioimpedance spectroscopy, multi-frequency, and single-frequency BIA considering its capabilities, and were set up as suggested by the manufacturer. Because setting up each frequency also affected the sampling rate independently, the sampling rate between frequencies differed. The device was calibrated daily using an internal resistor of 600 Ω according to the manufacturer’s guidelines.

We placed the electrode pairs (BlueSensor L, Ambu A/S, Ballerup, Denmark) on the skin with reference to defined bony landmarks (Fig. 1A). For the left wrist, the drive and sense electrodes were placed on the dorsal side of the wrist at the level of the ulnar styloid and five centimeters proximally to the drive electrode, respectively. For the right hand, the sensing electrode was placed at the wrist and the drive electrode was five centimeters distal to the sensing electrode. The sensing electrodes on the torso were placed five centimeters cranial from the anterior superior iliac spine and the driving electrodes were placed five centimeters laterally with respect to the sensing electrodes. For the wrist-to-torso measurements, we used the electrodes on the left hand and the contralateral side of the torso.

Schematic drawing of the electrode locations (A) and overview of the data collection procedure (B). During each session, three consecutive measurements per measurement location (hand-to-hand, hand-to-torso, torso-to-torso) were conducted.

Full size image

We collected data for the three previously defined body locations and in the named order five times in total (Fig. 1B). Within each of the five sessions, we collected three consecutive time series per frequency and location, resulting in a total of nine measurements per session. The first three sessions were conducted on the first day of data collection to measure the intra-day reproducibility. Between each session, we took a 15-min break, in which we reattached new electrodes. Further, we measured inter-day reproducibility by conducting two sessions on the two following days and comparing the data with the data from the first session. During all measurements, we followed the measurement standards by Kyle et al. [5], including e.g., data collection in the morning after overnight fasting in a supine position with abducted limbs, preparation of the skin, and exercise recommendations. Furthermore, we marked the outlines of the electrodes with a waterproof marker to ensure the electrode reattachment at the same location. To minimize the influence of breathing, subjects were told to completely exhale and keep their breath for eight seconds when collecting the data after the signal had been settled. Additionally, the subjects were told not to move during the sessions to minimize the influence of motion artifacts.

We post-processed the data using Python (V 3.11.5). In the first step, we cut and synchronized the raw signals semi-automatically (Fig. 2). To accomplish this, we plotted the raw signals and visually selected the area where the signals were settled after starting the measurement. From the beginning of the selection window, the local minimum was detected automatically, and the next 350 frames were subsequently exported. Signals were not filtered for the following analysis. To evaluate the intra-session precision, we calculated the mean coefficient of variation (CV% = standard deviation/mean*100) of all five sessions combined. Here, we first calculated the mean impedance of each of the time series. Using these values, we then calculated the CV% for each frequency and electrode location per session. The final CV% was then calculated by taking the mean values of all sessions combined. For the intra- and inter-day precision, the mean impedance for the time series of the corresponding frequencies and locations was used to calculate the overall mean and SD and consequently the CV%.

Impedance values were the largest for measuring from one hand to another, and they decreased with increasing frequency (Table 1, Fig. 2). Overall, the female subject demonstrated higher impedance values compared to the two male subjects independent of electrode location and frequency, while the electrode position on the torso throughout all subjects and frequencies demonstrated the smallest impedance values.

Synchronized raw signals from one subject during one session. The columns represent the different electrode locations, and the rows represent the used frequencies.

Full size image

The intra-session variation, regardless of electrode location and frequency, was small with CV% ≤ 1.44% (Table 2). While measurements from one hand to another (meanCV% = 0.14%; minCV% = 0.04% maxCV% = 0.28%) and from the hand to the torso (meanCV% = 0.11%; minCV% = 0.04% maxCV% = 0.22%) overall showed CV% ≤ 0.28%, larger means and standard deviations (SD) for CV% were found for the electrodes placed on both sides of the torso (meanCV% = 0.68%; minCV% = 0.24% maxCV% = 1.44%).

Regarding the intra-day and inter-day measurements (Table 3), the CV% for both hand-to-hand (intra-day: meanCV% = 1.67%; minCV% = 0.30% maxCV% = 2.69%; inter-day: meanCV% = 1.53%; minCV% = 0.75% maxCV% = 2.8%) and hand- to-torso (intra-day: meanCV% = 1.26%; minCV% = 0.77% maxCV% = 2.05%; inter-day: meanCV% = 1.31%; minCV% = 0.53% maxCV% = 1.89%) indicate good performance (CV% ≤ 2.8%). Overall, the torso-to-torso measurements demonstrated larger variations (intra-day: meanCV% = 5.53%; minCV% = 1.71% maxCV% = 24.64%; inter-day: meanCV% = 3.13%; minCV% = 2.20% maxCV% = 4.82%).

In this pilot study, we analyzed the measurement precision of a cheap and commercial BIA solution suitable for integration into a wearable device. Overall, the solution showed high precision in different body locations over a large range of frequencies and, therefore, suggests good usability for this cheap and wearable device.

Mean impedance values for the hand-to-hand and the torso-to-torso measurements ranged on average from 242–670 Ω and 12–35 Ω, respectively. Despite differences in electrode locations, these values are within the range of previous research for hand-to-hand measurements [8, 9] and for the latter setting [9]. While the extremities only account for a small fraction of the body volume, they contribute to the biggest part of the whole-body impedance contrary to the torso [10, 11]. We, therefore, expected the values for the measurement from one hand to the contralateral side of the torso to lay in between the impedance values for the other two settings. In accordance with the previous literature, we found higher impedance values for the female subject independent of electrode position and frequency compared to the two male subjects [12, 13].

The high intra-session precision with CV% values smaller than 0.3% for the hand-to-hand and hand-to-torso measurements is in accordance with results from Hamilton-James et al. [14] who showed similar results comparing percentage fat mass within three measurements using the same clinical hand-to-foot device. Additionally, the intra-day (hand-to-hand-CV% = 1.67%; hand-to-torso-CV% = 1.26%) and inter-day (hand-to-hand-CV% = 1.53%; hand-to-torso-CV% = 1.31%) measurements showed, on average, similar precision as previously reported with intra-day CV% of around 1–2% and inter-day CV% of around 2–3.5% [6]. This underlines that the here-used device can achieve similar precision for those two settings compared to the earlier established and more expensive devices. We were not able to confirm previously observed higher inter-day variation for frequencies lower than 50 kHz [15] on an individual basis, which might be due to the small sample size.

Compared to hand-to-hand and hand-to-torso, the torso-to-torso measurements had comparably larger mean CV% and SD both for intra-session measurements (meanCV% = 0.68%; minCV% = 0.24% maxCV% = 1.44%), within one day (meanCV% = 5.53%; minCV% = 1.71% maxCV% = 24.64%), and between days (meanCV% = 3.13%; minCV% = 2.20% maxCV% = 4.82%). Nevertheless, intra-session and inter-day CV% on average meet similar precision as previously reported for whole-body measurements [14] and [6], respectively. Still, the large SDs for the intra-session measurements and the overall larger values for the intra-day and inter-day CV% compared to the other settings indicate that the measurements on the torso only are more prone to measurement errors than the other locations. Interestingly, we were not able to observe noticeable differences between hand-to-hand and hand-to-torso measurement CV%, even though the latter also included the torso. Differences in absolute impedance due to, e.g., electrode position or small movements, therefore, seem to affect the torso more and these seem to be averaged out for longer conductors.

For the intra-day measurement, the CV% for the lowest frequency and the last subject (CV% = 24.64) was more than four times higher than the second-largest value (CV% = 5.41). We tried to explore the reason for this unexpected outlier, but measurement errors (e.g., wrong electrode placements) seem unlikely to have caused this because the other frequencies should have then been equally affected. When excluding this outlier, both the intra-day (meanCV% = 3.14 ± 1.26%) and inter-day (meanCV% = 3.24 ± 0.94%) variations were close to each other.

Several limitations must be considered when interpreting the results of this pilot study. One limitation is the small sample size which only allows us for a descriptive analysis of the results. Measurement errors and other deviations can, therefore, lead to large variations in data, which can affect the interpretation of the results. In the future, we will collect data from a larger sample size to draw stronger conclusions and to analyze if the trends observed in this study hold true. Another limitation is that we used only three frequencies in the measurements. However, the used frequencies still covered a wide range from low to high frequencies. Consequently, we assume that the device works likewise for frequencies in between. Due to limitations of the device regarding current regulations, we were not able to increase the frequency band without shifting it toward lower or higher frequencies. While the device itself gives stable outcomes for the collected frequencies with only small intra-session variation and has previously been validated against circuits [7], it still needs to be validated using subjects to determine the overall measurement accuracy of the device. Since our results are in accordance with the previous literature, we are confident that the current cheap BIA device provides reasonable results. Finally, all analyses in this study were based on unfiltered data that were cut semi-automatically. Therefore, we expect the results to improve when filtering out the random noise and physical signals, e.g., the heart rate.

On average, in this small subject group, the intra-session, intra-day, and inter-day precision of this commercial BIA solution is high and in accordance with the established devices. This suggests good usability for this cheap and wearable device which could allow for a comprehensive integration of BIA-based body composition measurements in clinical and non-clinical settings. For the torso-to-torso measurements, those results were only met for the intra-session and inter-day measurements, while overall showing a larger variation due to one unexpectedly different measurement result. Future studies need to validate the device and to show if the here-found trends hold for a larger sample.