Methods, Reference Data, and Clinical Validation for a DXA-Trained Estimation Model

1. Overview

Visceral adipose tissue (VAT) is the fat depot that surrounds the internal organs. It is metabolically distinct from subcutaneous fat, and elevated levels are associated with insulin resistance, type 2 diabetes, dyslipidaemia, and cardiovascular disease. Because VAT responds to diet and activity changes earlier and more visibly than total body weight does, tracking it gives members a more sensitive signal of metabolic progress than the scale alone.

This whitepaper describes the model Petal uses to estimate visceral fat from a member's midsection (trunk) fat mass together with their chronological age, body weight, and height. Every estimate is reported as five linked outputs: a visceral fat value, a 1-10 Visceral Fat Index (VFI), a plain-language category, an age-matched population percentile, and an expected-error range that communicates the confidence of the estimate.

This document describes Petal's current visceral fat model and supersedes the earlier regression-based description of this metric. The sections below cover the reference data the model was developed on, how it was developed and tested, its measured accuracy against the clinical reference standard, a cross-check against clinical DXA, and the scope within which a result should be interpreted.

2. Data & Statistical Methods

2.1 Reference Data

The model was developed on the National Health and Nutrition Examination Survey (NHANES), conducted by the National Center for Health Statistics at the U.S. Centers for Disease Control and Prevention. NHANES uses a complex, multistage probability sampling design to produce nationally representative estimates for the U.S. civilian non-institutionalised population, and participants receive whole-body DXA scans - the clinical reference standard for regional body composition. Using a nationally representative survey rather than a convenience sample means the model is anchored to the general population rather than to a single clinic or region.

DXA data was pooled across four NHANES cycles (2011-2018) to form an analytic sample of 5,610 women aged 20 to 59, with a mean VAT of 478 g (SD 269 g). Every visceral fat value used to develop and test the model is therefore a DXA measurement, not a derived or self-reported quantity.

2.2 Model Development and Testing

Petal's model learns the relationship between DXA-measured visceral fat and four inputs that Petal either measures directly or already holds in a member's account: trunk fat mass, age, body weight, and height. The specific model form, its parameters, and its calibration are proprietary.

The development process is straightforward to state, and it is what the accuracy figures in Section 3 rest on. The pooled sample was partitioned at the participant level into a development set and a held-out test set of 908 women. All model selection and tuning was performed on the development set only; the held-out women were never seen during development and were scored exactly once, at the end. The figures reported in Section 3 are therefore out-of-sample results on genuinely unseen participants, not in-sample fit statistics.

A final calibration step aligns the model's output with clinical DXA measurement as performed in accredited laboratories, which differs slightly in absolute terms from the NHANES research protocol. This step is what the clinical validation in Section 4 evaluates.

2.2 Population Percentiles

Alongside the estimate itself, each result is placed against an age-stratified population reference distribution, and a percentile is obtained by interpolating the member's value against the reference curve for their age. The reference is stratified on a continuous basis rather than by decade, which avoids the artificial jumps that appear at decade boundaries - a member does not see their percentile shift simply because they had a birthday.

3. Accuracy Against the DXA Reference Standard

On the held-out group of 908 women, Petal's visceral fat estimate explains 82.6% of the variation in DXA-measured visceral fat (R² = 0.826), with a mean absolute error of 74.8 g (0.16 lb) and a root-mean-square error of 118.0 g (0.26 lb). Because these are out-of-sample results, they reflect the accuracy a new member can expect rather than the model's fit to its own training data.

Accuracy is highest in the lower and middle of the range, where most members fall, and decreases at very high visceral fat, where DXA measurements are themselves more variable and the reference data is sparser. In that upper range the model tends to read slightly low.

3.1 Visceral Fat Index and Categories

The estimated visceral fat value is also expressed as a 1-10 Visceral Fat Index and mapped to one of five plain-language categories - Very Low, Healthy, Above Healthy, Elevated, and Very High - so that a member can interpret a result without reference to absolute masses in grams or pounds.

Category assignment was evaluated on the same held-out group. The category Petal reports matches the category derived from the DXA measurement exactly 74% of the time, and falls within one adjacent category 99.2% of the time. In other words, a member is very rarely placed more than one step away from where their DXA scan would place them.

The boundary between the Healthy and Elevated categories reflects the level above which visceral fat is associated with elevated cardiometabolic risk in the published literature, including insulin resistance, dyslipidaemia, hypertension, and cardiovascular disease (Després & Lemieux 2008; Lemieux et al. 2000).

3.2 Expected-Error Range

Every estimate is reported with an expected-error range rather than as a single number, so that the confidence of the estimate is visible alongside the estimate itself. The range is calibrated from the observed error distribution on the held-out set, at two coverage levels. Across the held-out group, roughly 77% of members fall inside the 75% range and 88% inside the 85% range, closely matching the intended coverage - the ranges are neither optimistically narrow nor uninformatively wide.

The range is a prediction interval, not a measurement tolerance. It describes the spread of plausible true values given the estimate; it does not imply that a repeat measurement on the same day would vary by that amount.

4. Cross-Check Against Clinical DXA

Section 3 establishes accuracy at population scale against research-protocol DXA. A separate cross-check confirms that the calibrated estimate also agrees with DXA as performed in clinical practice, where absolute values differ slightly from the NHANES protocol. Petal estimates were compared against paired clinical DXA scans obtained by volunteers at accredited laboratories - eight scans across six women, ages 28 to 68.

The mean absolute difference from the DXA measurement was 0.18 lb (approximately 82 g), consistent with the held-out error reported in Section 3, and the reported expected-error range contained the DXA value in all eight scans. The estimate read below the DXA value on six of the eight scans, by an average of 0.17 lb - the same direction as the high-range behaviour noted above.

At this size the comparison confirms that the clinical calibration is correctly aimed; it is reported as a check rather than as evidence of clinical equivalence.

5. Scope and Interpretation

The boundaries of what this model has been shown to do are stated explicitly, so that a result is neither over-read nor applied outside the population it was developed on. The following applies to every visceral fat estimate Petal reports.

5.1 Who the Estimate Applies To

• Developed for women. The model was developed and validated on female DXA data throughout, and is applied only to women.
• Best supported between ages 20 and 59. The DXA reference data underpinning the model covers this range. For members above 59 the estimate extrapolates beyond it: the direction and movement of the value remain meaningful, but the absolute number carries more uncertainty and should be read together with its expected-error range.
• Referenced to U.S. population norms. Percentiles are expressed relative to the U.S. civilian non-institutionalised population. For members from populations with substantially different body composition distributions, the percentile is best read as indicative rather than exact.

5.2 How to Read the Number

• An estimate with a stated range, not an image. Petal estimates visceral fat from body composition inputs; it does not image the abdomen. The expected-error range is a prediction interval, and by design a defined proportion of members fall outside it.
• Quality follows the trunk fat measurement. Midsection fat mass is the primary input, so the quality of an estimate tracks the quality of that measurement. It is not obtainable from routine anthropometry such as height, weight, and waist circumference alone.
• Widest at very high visceral fat. Error grows in the upper range and the estimate tends to read slightly low there, where reference data is sparser and DXA measurement is itself more variable.

6. References and Data Citations

National Center for Health Statistics. National Health and Nutrition Examination Survey Data. Hyattsville, MD: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, 2011-2018. https://wwwn.cdc.gov/nchs/nhanes/

Després JP, Lemieux I, Bergeron J, et al. Abdominal obesity and the metabolic syndrome: contribution to global cardiometabolic risk. Arteriosclerosis, Thrombosis, and Vascular Biology. 2008;28(6):1039-1049.

Lemieux I, Pascot A, Couillard C, et al. Hypertriglyceridemic waist: a marker of the atherogenic metabolic triad in men? Circulation. 2000;102(2):179-184.

This document is for informational and educational purposes only. The model and its outputs are not a clinical diagnostic tool and should not be used for medical decision-making without professional interpretation.