Validation of non-empirical fat-free mass estimation model for a wrist-worn device

Obesity, one of the most crucial problems in healthcare, has risen dramatically in the past few decades [1]. Among the present methods of body composition evaluation, bioelectrical impedance analysis (also known as bioimpedance analysis or BIA), is one of the most widely used, as it provides a quick and efficient way to estimate and control body fat level.

Fat-free mass (FFM), i.e. the difference between fat mass (FM) and total body mass (TBM), is commonly employed by researchers and specialists studying body composition. In general, FFM, as well as FM, can be used as a basis for creating a personal diet plan in order to prevent obesity or muscular dystrophy. Purely anthropometric FFM estimation methods, such as equations derived by Duerenberg [2], Peterson [3], and the Navy Seal formula [4], have limitations, mostly arising from their strong dependency on the traits of the original experimental sample: ethnicity, sex (gender), age, and general health. The gold standard in FFM estimation, dual-energy X-ray absorptiometry (DEXA), has more recently been widely adopted because of high measurement accuracy, accessibility, and short scanning time [5]. DEXA is, however, not suitable for everyday body composition analysis due to equipment size and cost. Moreover, the maximum effective X-ray radiation dose for a single DEXA scan can reach 1.92 μSv, so this method cannot be recommended for frequent use [6].

Thus, BIA can be a way to overcome the challenges of DEXA and anthropometric FFM estimation, since the current production level of the electronics manufacturing industry makes it possible to use it widely. The method utilizes small variations in the human body’s electric conductivity, altered by the impinged frequencies and body composition [7]. Usually, BIA models are formed by training multiple linear regressions, where DEXA results are used as training data (even though DEXA as an FFM estimation method does have its own bias [8]). An example of this approach is the single-cylinder-based BIA model, in which FFM is determined by the value of height²/R (known as bioimpedance index, BI), where R denotes the real part of bioimpedance (resistance) between the voltage (sensing) electrodes. The model assumes the human body is an isotropic conductor with a constant cross-sectional area, with the length of the conductor being the height. This makes it necessary to add age and sex variables to the prediction models in order to consider age and sex differences in fat and muscle mass body distribution among the population [9, 10]. However, as the human body has a more complex shape than a single cylinder, the model has severe limitations. For accuracy purposes, the human body can be represented as five cylindrical conductors: four limbs and a torso, excluding the head [8]. With such a five-cylinder model, empirical estimation can be minimized due to direct impedance measurement of body parts. The model's main drawback is the requirement of a multielectrode system, contacting each limb, usually incorporated in a bulky full-size device, which is suitable for the laboratory but not for home use. Although laboratory BIA measuring equipment has multiple advantages such as higher accuracy and inter-observer reproducibility, interest in portable BIA devices, especially wearable ones, has been growing in recent years, mainly due to the growing wearable devices market [11].

In order to analyze upper-body impedance, a wearable wrist-worn impedance meter (WIM) was designed as a modification of AURA Strap, a wrist-wearable Apple Watch accessory (AURA Devices, Inc., USA). The WIM is based on CS1259 (analog front-end with 24-bit ADC designed for body impedance analysis, Chipsea Technologies Shenzhen Corp Ltd, China) and uses a tetra-polar BIA measurement scheme where two electrodes contact with the volar wrist and the other two contact with the free hand. The impedance meter is powered by a CR2025 battery in order to avoid noise from an external power source and maintain autonomous work of the device. Control of the WIM and data transfer are performed via Bluetooth protocol.

In this study, we modified the five-cylinder model into a new model, which includes two upper limbs, for the described bioelectrical impedance meter. With this modification, we expect to keep the advantages of the five-cylinder model, i.e., minimize the empirical estimation of body parts and take additional sex and age variables out of consideration, while employing the hand-to-hand measurement scheme.

Our main aim for the present study was to build a novel non-empirical FFM estimation model based on anthropometric parameters and BIA values measured with the WIM. Since electrode location and measuring frequency affects both BIA accuracy and precision, we also investigated the effect of these factors on the model’s performance. Finally, we compared the non-empirical model’s effectiveness against the population-based model.

In this study, we developed the novel TTC model of FFM estimation based on hand-to-hand BIA measurements made by a wrist-worn impedance meter. We also investigated the effects of measuring frequency and electrode positions on model performance and demonstrated the agreement between the model and reference DEXA estimates of fat-free mass.

We derived both empirical and non-empirical regression equations based on the single cylinder and TTC models. One of the main advantages of the TTC model turned out to be the uniformity of distribution of maximal residual error. The population-based single cylinder model error followed a multimodal distribution with several peaks resulting from the influence of the sex variable. In practical use, this can lead to a significant FFM prediction error. The TTC model, on the other hand, did not have this disadvantage.

Bioimpedance values (Table 2) measured using different placements of the WIM electrodes and at different frequencies varies. The highest value was measured through the fingers (Figure 2a) for all frequencies since they represent the longest body pathway among the ones tested. Measurement frequency also affects the bioimpedance value: a 10 kHz current resulted in the highest resistance for all electrode placements.

The effect of measuring frequency and electrode placements on the performance of FFM models was also analyzed. The bioimpedance measured through the proximal phalanx of the index finger of the right hand (Figure 2b) resulted in the worst performance with the lowest r² and the highest MAPE. Although measurements taken through the index and middle fingers showed better performance, it was still low compared with measurements taken through the thenar. This could be related to the high influence of a non-relevance additional impedance of the whole hand including contacting fingers, whose impedances do not strongly depend on a person’s body composition.

It was found that 50 kHz measurements on the right hand thenar (Figure 2a) resulted in the lowest MAPE and maximal residual error and the highest r² for both models. It is possible that the higher amount of muscle tissue (e.g., flexor pollicis brexis, adductor pollicis brevis) under the thenar electrode is associated with more stable values of measured bioelectrical impedance, probably due to noise reduction and lesser effect of skin conductance.

Comparison of the SC and TTC models’ performance with the results obtained by DEXA scan demonstrated the high accuracy of both models, with mean bias between the methods being not significantly different from zero. These results demonstrating the new TTC (as well as SC) model for the WIM suggest that it can provide an accurate and effective method to assess body fat-free mass.

There are, however, limitations to the uses of the new models, the main one being related to the physical principles behind bioimpedance analysis. As only single-frequency BIA (50 kHz) measured impedance was utilized in building the models, their ability to distinguish between extracellular and intracellular water is hindered. As a result, FFM estimation could be easily affected by electrolyte imbalances or body hydration status [17]. As a typical empirical BIA model, our single cylinder model depends on the sex, age, and ethnicity of the experimental sample, and can possibly perform worse for the rest of the population. The TTC model is non-empirical but has a practical limitation due to the requirement of specific measurements, such as wrist and shoulder circumferences.

In the future, the TTC model can be enhanced by employing bioimpedance spectroscopy (also known as multifrequency BIA), since the use of several frequencies makes it possible to differentiate total body water into extracellular and intracellular spaces and avoid the effect of body hydration status on FFM prediction [18]. Moreover, since the WIM prototype is able to assess phase angle, as well as both the real (resistance) and imaginary (reactance) part of bioimpedance, this data can be used to reveal water distribution between the extra- and intracellular spaces [19]. This will further improve the performance of the TTC model. Last but not least, the geometrical shape of the arms can be more closely approximated: instead of a truncated cone, each arm can be represented as a more realistic complex geometrical object consisting of several truncated cones.