The study design is summarized in Fig. 1. The first study phase involved development of manifold regression prediction models on a sample of 570 healthy adults. Additional information is provided in Methods on development of the manifold regression models. The second prospective phase then followed with comparison of predicted avatar anthropometric dimensions (6 circumferences, 7 volumes, and 7 surface areas) to corresponding ground-truth estimates in a new sample of 84 healthy adults. Ground-truth anthropometric measurements were acquired with a 20-camera 3D optical scanner (SS20, Size Stream, Cary, NC). Predicted avatars were developed by manifold regression using several different exploratory combinations of demographic, physical, and other accessible characteristics as described in the Methods section. Accessible characteristics in the current study were acquired with dual-energy X-ray absorptiometry (DXA, QDR Discovery, Hologic, Marlborough, Massachusetts) and bioimpedance analysis (BIA, InBody S10, Seoul, South Korea). The predicted and actual 3D avatars were analyzed using the same Universal Software [10, 11] developed to identify standard anatomic landmarks.

The first study phase involved development of manifold regression models and the second phase involved comparisons of predicted and ground-truth avatar body circumferences, volumes, and surface areas.

In the first phase of the study, manifold regression model development, participants were evaluated as part of the cross-sectional Shape Up! Adults study (NIH R01 DK109008). The Shape Up! Adults study was designed to investigate associations between body shape and composition with multiple health markers [4, 9]. In the second phase of the study, avatar anthropometric evaluation, participants were a new prospectively evaluated sample of healthy adults at or over the age of 18 years who completed the protocol measurements on the same day. These participants were recruited from the local community through web postings and print media. All participants enrolled in the study self-reported their race/ethnicity. The parent study for this project was approved by the Pennington Biomedical Research Center and University of Hawaii Cancer Center Institutional Review Boards and is posted on ClinicalTrials.gov (ID NCT03637855). The second phase of the current study was approved by the Pennington Biomedical Research Center Institutional Review Board (IRB# PBRC 2022-002). Baseline evaluations included health screening and measurement of body weight and height.

After 3D optical data acquisition, each scan was registered to a 60,001-vertex template using the methods of Allen et al. [12]. This standardization allowed direct anatomical body shape comparisons across the sample. First, seventy-five fiducial points defined in the Civilian American and European Surface Anthropometry Resource Project [13] were manually placed on the raw meshes by trained and validated personnel using Meshlab 1.3.2 (Consiglio Nazionale delle Ricerche, Rome, Italy). Using the software Ganger, developed by Allen et al. [12], the template’s markers were transformed to each target mesh’s markers. The vertices of the template warps to fit the shape of each participant’s mesh [14]. Next, a principal component (PC) transformation of the meshes was performed to create sex-specific statistical shape models. These models described 99% of the body shape variance using fewer than 15 PCs [15].

Manifold regression analysis was performed following the creation of the shape models. The manifold equation is M = P × F+, where M is the manifold, P is the matrix of all PCs for all participants in the shape model, F is the matrix of all feature parameters (e.g., height and weight) for all participants, and + symbolizes the pseudoinverse. Once M was calculated, another matrix was created, W, which contained the target features from a person’s feature parameters (e.g., height = 150 cm and weight = 60 kg). Matrix, M, was then multiplied to matrix W, creating a new PC matrix where the target features of W have modified M. The new PC matrix was then transformed back into Cartesian space from the PC space to generate the manifold images [8, 9].

The manifold regression models can predict 3D humanoid avatars using demographic covariates such as age and physical characteristics including weight and height. Additional characteristics can be included in the equations such as %fat and impedance values from BIA. Adding more covariates usually refines predictions, especially in samples that have highly varied body shapes. In the current study, we found in exploratory evaluations that the simplest model giving good anthropometric predictions relative to ground-truth included age, weight, height, and %fat (DXA) as covariates. Since the shape models were sex-specific, sex was not used as a covariate. This four-variable model was created by modifying F in the manifold equation. An example of the difference in predicted avatars between a model with age, weight, and height and a model that additionally included %fat is shown in Fig. 2 for a young muscular adult male. The three-variable model did not distinguish people in the current study who were muscular from their counterparts with greater relative adiposity as was observed in the participant presented in the figure. Manifold regression analysis was performed in R version 4.2.1 (https://stat.ethz.ch/pipermail/r-announce/2020/000658.html; R Core Team, 2020).

The evaluated man was a body builder (left panel) with a waist circumference of 112 cm (designated by arrows) and hip circumference of 119 cm as measured with a 3D optical scan; his body fat was 19% and body mass index 35.5 kg/m2. Manifold-predicted waist and hip circumferences (middle panel) were 117 cm and 119 cm, respectively, with age, weight, and height as covariates. Manifold-predicted waist and hip circumferences (right panel) were 111 cm and 117 cm, respectively, with %fat added to the age, weight, and height covariates. Adding %fat to the model including age, weight, and height as covariates brought visual appearance closer to actual appearance and improved waist circumference prediction to within 1 cm of that evaluated with a 3D optical scan.

Anthropometric body dimensions were evaluated in the predicted and ground-truth avatars with Universal Software. This software operates on Matlab (Mathworks, Natick, MA) [10, 16] and runs four sequences including pre-processing, landmark detection, body partitioning, and surface area calculation. Initial scan processing repairs gaps or imperfections in the 3D mesh. Major anatomic landmarks are next detected [10] at the crotch, right/left armpits, shoulders, hips, and toes. The software then partitions body mass into six regions including head-neck, trunk, right/left arm, and right/left leg followed by calculation of body lengths, 6 circumferences (waist, hip, right/left mid-upper arm, right/left thigh) and 7 regional/total volumes (head/neck, torso, right/left arms and legs, whole-body), and the same 7 regional/total surface areas. The circumference sites are shown in Supplementary Information I and Fig. S1.

The SS20 3D optical reference system includes twenty structured light infrared depth sensors mounted on four vertical columns. Participants stood in the A-pose at the center of the columns and data was acquired during a 4-second scan. The acquired avatars were analyzed using Universal Software.

The QDR Discovery DXA was operated with software version V8.26a:3.19 and calibrated at regular intervals according to manufacturer specifications. The National Health and Nutrition Examination Survey scanner option was turned off. Two components were evaluated, total body fat and fat-free mass; percentage (%) fat mass was derived as (fat mass/body mass) x 100.

Avatars created using manifold regression (predicted) were compared with the actual (ground truth) participant avatars for selected circumferences, volumes, and surface areas using linear regression analysis (R2) and with means (±SD), root-mean square errors (RMSEs), mean absolute errors (MAEs, X ± SE), concordance correlation coefficients (CCCs), and Bland-Altman analyses [17]. Mean values for circumferences, volumes, and surface areas predicted with manifold regression were compared to their ground-truth counterparts with paired t-tests adjusted for multiple comparisons. The predicted and ground-truth avatar comparisons are presented separately for the circumferences and combined for the volumes and surface area evaluations.