Evaluation of a smartphone-based visual body composition application with artificial intelligence trained by 4-compartment model

doi:10.21203/rs.3.rs-1758798/v1

Download PDF

Brief Communication

Evaluation of a smartphone-based visual body composition application with artificial intelligence trained by 4-compartment model

https://doi.org/10.21203/rs.3.rs-1758798/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The aim of this study was to assess the precision and agreement of a visual body composition (VBC) application using a 4-compartment (4C) model criterion. Body composition estimations were collected from 102 participants using methods required for a 4C model and a VBC application using two different smartphone types. Intraclass correlation coefficients (0.96–0.99) and root mean square coefficients of variation (0.5%-3.0%) showed good reliability for body fat %, fat mass (FM), and fat-free mass (FFM) between and within smartphone types. There were no significant mean differences between 4C model estimates and the VBC estimates for the total sample, by sex, and for white and Black/African-American participants. VBC estimates demonstrated equivalence with the 4C model for all variables, but revealed proportional biases that underestimated BF% and FM at higher degrees of each variable. VBC applications trained by a 4C model are reliable and produce body composition estimates equivalent with a 4C.

digital anthropometry

mobile health

body composition assessment

obesity

smartphone

artificial intelligence

Altered distributions of fat mass (FM) and fat-free mass (FFM) are associated with several disease states (1) making body composition assessments crucial in determining individual health status. However, comprehensive body composition assessments are rarely utilized in practice due to the cost and limited access of research-grade devices and the recent surge in remote healthcare models that provide treatment regardless of geographical location. (2) The limited availability of high-quality body composition devices and the rise remote healthcare supports the need for remote body composition assessments that can thoroughly evaluate patients without the need for costly devices or frequent in-person visits. Smartphone-based visual body composition (VBC) applications that use artificial intelligence (AI) to remotely automate body composition metrics from 2-dimensional (2D) images may offer a potential solution. To our knowledge, there are no studies investigating the agreement between a VBC application trained by a 4-compartment (4C) model (3) and an actual 4C model. Therefore, the purpose of this study was to evaluate the precision of a 4C model trained smartphone-based VBC application and its agreement with a 4C criterion model.

A total of 102 participants between ages 18–75 completed this cross-sectional study (Supplementary Table 1). The study was approved by an Institutional Review Board (IRB#21–213) and informed consent was obtained prior to participation.

Participants reported to the laboratory after an ≥ 8 h fast and abstention from exercise. Upon arrival, participants verified hydration using an 8-point color chart. (4) Participants then completed body composition assessments necessary for a 4C model estimation including body mass (BM; SECA, Hamburg, Germany), dual-energy x-ray absorptiometry (Lunar iDXA, General Electric, Boston, MA, USA), and bioimpedance spectroscopy (SFB7, ImpediMed, Carlsbad, CA, USA). The 4C body composition model employed the methods described by Graybeal et al. (4) using the equation developed by Wilson et al. (5) for body volume and the 4C equation from Wang et al. (6)

VBC application (MeThreeSixty®, Size Stream LLC, Cary, NC, USA) assessments were conducted, in duplicate, using the front facing camera on an iPhone 12 Pro® (Apple® Inc., Cupertino, CA, USA) and a Samsung Galaxy S21+® (Samsung® Group, Suwon, South Korea) in accordance with the manufacturer guidelines and in a random order. This VBC application has recently been reported to exhibit acceptable reliability for anthropometric variables when compared to tape measurements. (7) Smartphones were placed in a fixed tripod set at an approximate waist height (91.0 cm) for all participants. For the first image, participants faced the camera and stood with arms and feet positioned away from the torso. For the second image, participants turned with their left shoulder facing the camera and arms/hands aligned with and against their trunk/thighs. BM obtained from the calibrated scale was applied to the body fat % (BF%) outputs to yield estimates of FM and FFM.

Variables included BF%, FM, and FFM. Mean difference (MD) was calculated as the device in question minus the reference. Intraclass correlation coefficients (ICC; two-way, random, absolute agreement), precision error (PE), and root mean square coefficient of variation (RMS-%CV) were used to assess precision within and between smartphone types. Agreement was assessed by paired samples t test, Pearson correlation coefficients, root mean square error, standard error of the estimate (SEE), and equivalence testing. Equivalence regions of ± 2% were employed for BF% and ± 1 kg for FM and FFM. Individual accuracy was assessed using the methods of Bland and Altman. (8) Separate analyses were conducted to examine differences by sex and race for non-Hispanic white and Black/African-American individuals. Statistical significance was accepted at p ≤ 0.05.

All ICCs between and within smartphone types ranged from 0.997-1.000 (Supplementary Table 2; all p < 0.001). RMS-%CV were small for each variable within smartphone types (0.5%-1.2%) and slightly higher between smartphone types (1.3%-3.0%). PE for each variable were similar for each smartphone (BF%: 0.3–0.4%; FM: 0.3 kg; FFM: 0.2–0.3 kg) and slightly higher between smartphone types (BF%: 0.8%; FM/FFM: both 0.7 kg).

For the validity analysis, there were no significant differences between the VBC application and the 4C model in the total sample, by sex, or by race (Table 1). Equivalence testing also demonstrated equivalence to 4C model for all variables (Supplementary Fig. 1). RMSE were highest for BF% for the total sample (5.0%-5.1%), for sex groups (4.4%-5.5%), and for race comparisons (4.5%-5.5%) with higher, but similar, errors produced in females and white participants. RMSE for FM and FFM were in a similar range for all groups (3.6 kg to 4.2 kg).

Bland-Altman Analyses (Fig. 1) revealed significant proportional bias for all variables in the total sample other than FFM produced by MeThreeSixty_Samsung. Relative to the 4C model, BF% was significantly underestimated by the VBC application at higher degrees of BF% (both β= -0.25; p < 0.001). Despite significance, proportional bias for FM and FFM was generally low for the total sample (β < 0.10). Both BF% (β= -0.50 to -0.48; both p < 0.001) and FM (β= -0.18 to -0.17; both p < 0.001) were underestimated by the VBC application in females relative to the 4C model which was not observed for males. However, MeThreeSixty_Apple underestimated FFM relative to the 4C model in males at higher degrees of FFM (β= -0.15; p = 0.011). Interestingly, there was no proportional bias observed for any variable in Black/African-American participants whereas significant proportional bias was revealed for every variable in white participants. LOA were high for all variables across groups (BF%: 9.9%-10.8%; FM and FFM: 7.6–8.3 kg).

In conclusion, a VBC application trained by a 4C model produces reliable body composition estimates both within and between smartphone types that are equivalent with a 4C criterion model regardless of sex and the racial groups analyzed in this study. However, individuals with higher degrees of adiposity should use with caution given its potential to underestimate BF%. At present, this method offers a feasible and accurate approach to assess body composition conveniently and remotely, but future research should investigate its ability to assess body composition changes over time. Furthermore, users of this VBC method should employ caution when comparing estimates to non-4C model assessments.

DATA AVAILABILITY

Data are available from the corresponding author on reasonable request.

ACKNOWLEDGMENTS

The authors have no acknowledgements to disclose.

AUTHOR CONTRIBUTION

AJG: Conceptualization, data curation, formal analysis, investigation, methodology, project administration, supervision, roles/writing – original draft.

CFB: Data curation, investigation, methodology, validation, writing – review & editing

GMT: Conceptualization, formal analysis, investigation, methodology, validation, visualization, writing – review & editing.

FUNDING

Funding was provided by the University of Southern Mississippi start-up funds for AJG. There are no other funding sources to report.

ETHICAL APPROVAL

All data collection was performed in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board at University of Southern Mississippi (IRB#21-213). Informed consent was obtained from all subjects.

COMPETING INTERESTS

The authors have no competing interests to declare. Unrelated to the present study, GMT has previously received an in-kind donation of equipment – but not monetary funding – from the manufacturer of the visual body composition application used in the present study (Size Stream LLC, Cary, NC, USA).

Sheean P, Gonzalez MC, Prado CM, McKeever L, Hall AM, Braunschweig CA. American society for parenteral and enteral nutrition clinical guidelines: the validity of body composition assessment in clinical populations. J Parenter Enter Nutr. 2020;44(1):12–43.
Mann DM, Chen J, Chunara R, Testa PA, Nov O. COVID-19 transforms health care through telemedicine: Evidence from the field. J Am Med Inform Assoc JAMIA. 2020 Jul 1;27(7):1132–5.
Harty PS, Sieglinger B, Heymsfield SB, Shepherd JA, Bruner D, Stratton MT, et al. Novel body fat estimation using machine learning and 3-dimensional optical imaging. Eur J Clin Nutr. 2020 May;74(5):842–5.
Graybeal AJ, Moore ML, Cruz MR, Tinsley GM. Body composition assessment in male and female bodybuilders: a 4-compartment model comparison of dual-energy x-ray absorptiometry and impedance-based devices. J Strength Cond Res. 2020 Jun;34(6):1676–89.
Wilson JP, Strauss BJ, Fan B, Duewer FW, Shepherd JA. Improved 4-compartment body-composition model for a clinically accessible measure of total body protein. Am J Clin Nutr. 2013 Mar;97(3):497–504.
Wang Z, Pi-Sunyer FX, Kotler DP, Wielopolski L, Withers RT, Pierson RN, et al. Multicomponent methods: evaluation of new and traditional soft tissue mineral models by in vivo neutron activation analysis. Am J Clin Nutr. 2002 Nov;76(5):968–74.
Smith B, McCarthy C, Dechenaud ME, Wong MC, Shepherd J, Heymsfield SB. Anthropometric evaluation of a 3D scanning mobile application. Obesity. 2022;30(6):1181–8.
Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. The Lancet. 1986 Feb 8;327(8476):307–10.

Table 1 is available in the Supplementary Files section.

There is NO conflict of interest to disclose.

Download PDF

Version 1

posted

You are reading this latest preprint version

Evaluation of a smartphone-based visual body composition application with artificial intelligence trained by 4-compartment model

Status:

Version 1

Abstract

Figures

Introduction

Materials And Methods

Results

Declarations

References

table

Additional Declarations

Supplementary Files

Status:

Version 1