Body Composition Reports as Bayesian Priors: Handling DXA, BIA, and Skinfold Noise in Clinical Workflow

Body composition is the single most over-trusted number in a sports nutrition chart. A DXA report with two decimal places of body fat percentage looks definitive. It is not. It is one estimate from one machine on one day, and the question for clinical work is not "what does it say?" — it is "how much should I update my prior in response to it?"

This is a Bayesian framing, and it is the correct framing.

The athlete is not a body fat percentage. The athlete is a probability distribution over possible body fat percentages, narrowed by every measurement you have, weighted by how much each measurement deserves to be trusted. The DXA, BIA, and skinfold reports on your desk are likelihoods. Your job is to combine them with everything else you know and produce a posterior that informs the plan.

Most RDs do not work this way. They take the most recent number and write it in the chart. Then six weeks later they take the next number and compute a difference. If the difference exceeds some threshold they declare progress or regression. This procedure produces noise-driven plan changes, and it is responsible for a meaningful share of the dietary thrash that erodes athlete trust.

Here is how I handle it instead.

Each Method Has a Different Error Structure

DXA, BIA, and skinfolds do not share a common scale and do not share a common bias. Treating them as interchangeable readings of the same underlying truth is the central mistake.

DXA has the lowest within-day measurement noise — typically ±1.0–1.5% body fat for repeat scans on the same machine, same operator, same hydration state. But it has substantial between-machine and between-software-version bias. A Hologic Horizon and a GE Lunar iDXA can disagree by 2–3 percentage points on the same body. Software updates reset the calibration and can shift trend lines artificially. Glycogen and water content affect bone-corrected fat mass estimates by roughly 0.5–1.0% per 100g of glycogen change, which means a depleted athlete and a fully fueled athlete will not produce comparable scans.

BIA has higher within-day noise — ±2–4% body fat depending on the device tier — but it is fast, cheap, and repeatable. It is also extremely sensitive to hydration, food, and exercise within the prior 4 hours. A morning fasted measurement and an afternoon post-training measurement will disagree by enough to mimic a month of "progress" in either direction. Multi-frequency bioimpedance (e.g., InBody, Seca mBCA) reduces this somewhat by separating intracellular from extracellular water, but the underlying assumption that the impedance pattern maps cleanly to fat mass remains indirect.

Skinfolds have the highest method-bias risk because operator skill is variable, but a well-trained operator using ISAK protocol can produce the most repeatable inter-day signal of any of the three methods for tracking change over time in a known athlete. The absolute body fat percentage from a skinfold equation is wrong (every equation is wrong; some equations are useful), but the skinfold sum itself is a real, mechanically grounded measurement of subcutaneous fat under specific anatomical landmarks.

The implication is that each method answers a different question. DXA is best for one-off baseline phenotyping. BIA is best for high-frequency monitoring under tightly controlled conditions. Skinfolds are best for tracking change in a single athlete over time when operator and protocol are held constant.

The Prior Comes First

Before you read the report, you should already have a prior on what the number should be. Build it from three sources:

Sport and position. A male collegiate distance runner sits in 6–10% by skinfold, 8–12% by DXA. A male collegiate offensive lineman sits in 22–30% by DXA. A female heavyweight grappler sits in 25–35% by DXA. These are population priors. Knowing them lets you spot reports that should not be plausible before you spend a session on a flawed number.

History. The athlete's last three valid measurements are a tighter prior than any population estimate. If three DXA scans over six months sat at 14.2, 14.7, and 14.4%, your prior for the next scan is roughly N(14.4, 0.5) — meaning you should be surprised by a 17.8% reading and you should not commit to a plan change based on it without re-measurement.

Mechanism. A 2.3% reduction in DXA body fat over six weeks during a kcal deficit of 250/day is mechanistically improbable. The thermodynamics do not support that velocity of fat loss in an already-lean athlete. If the report shows it, the report is wrong, the deficit is wrong, or both.

When the new number is far outside the prior, the right move is almost always to remeasure, not to update the plan. The cost of a repeat scan is one hour and a copay. The cost of a wrong plan change driven by measurement noise is the next four to twelve weeks of the athlete's training cycle.

How to Update on the Posterior

Bayes' rule in a clinical workflow does not require math. It requires three habits.

Always log the conditions, not just the number. Every measurement entry in the chart needs: device, operator, time of day, hours since last meal, hours since last training session, [hydration state](/blog/hydration-status-assessment-in-clinical-workflow) (urine specific gravity if available, otherwise self-reported), and menstrual cycle phase if relevant. A naked body fat percentage is not a measurement; it is a rumor. The conditions are what let you reweight the likelihood.

Compare like to like. A DXA at 14.2% does not compare to a BIA at 16.8%. They are answering different questions and they have different biases. If you must reconcile them, anchor to the lower-noise method (DXA for baseline, skinfolds for trend) and treat the others as confirmatory or contradictory signal — not as alternative readouts of the same truth.

Weight new evidence by its precision. A new DXA on the same machine with the same conditions as the previous scan is a strong update. A new BIA on a different device after a hard training session is a weak update. The chart entry should reflect that — use language like "consistent with prior trend," "weak signal due to hydration confound," "remeasure recommended" — and resist the temptation to round every reading to a fact.

A Worked Example

Female track athlete, 19 years old, 800m specialist. Prior body comp from skinfolds taken every six weeks by the same RD using ISAK 7-site sum: 62, 60, 59, 61, 60. Prior is approximately N(60.4, 1.1) on the sum, which the regression maps to 12.5% body fat ±0.5.

She comes in for a sponsored DXA in the off-season. Result: 18.4% body fat. The athlete and her coach panic. The RD does not.

The DXA-vs-skinfold gap is real but not surprising at this magnitude — DXA systematically reads higher than skinfold-derived BF% for lean athletes, often by 4–6 percentage points, and the literature is consistent on this. The athlete's ISAK sum from the same week is 61, fully consistent with the trend. The DXA is a one-time baseline, not a trend signal.

The clinical action is not to change the plan. It is:

• Note the DXA as a baseline reference point with a clear method-bias caveat in the chart
• Continue tracking the ISAK sum on the same six-week cadence
• Schedule the next DXA on the same machine in 12 weeks for a within-method comparison

This is what Bayesian thinking looks like in a clinical chart. The new evidence is logged. The prior is preserved. The plan does not move on noise.

The Documentation Pattern

For audit-defensibility and for future-you reading the chart cold six months later, every body composition entry in the SOAP should follow this shape:

• Method: DXA / BIA / Skinfold ISAK 7-site
• Device/Operator: Hologic Horizon machine ID #2 / NM, ISAK Level 2
• Conditions: time of day, hours fasted, hours post-training, hydration
• Reading: primary metric (e.g., 14.2% BF; or skinfold sum 60mm)
• Comparison: vs N prior measurements — trend / outlier / remeasure pending
• Interpretation: strong update / weak update / discard
• Plan impact: none / minor / triggers re-evaluation

The "Interpretation" line is the Bayesian payload. It is the line that separates a clinician who treats numbers as facts from one who treats them as evidence.

What This Costs You

Working this way feels slower because it forces you to second-guess reports that the athlete and the coach treat as definitive. The athlete wants a number. The coach wants a target. The DXA technician confidently delivers 18.4% and the screen says so.

Your job is not to tell the room a story the screen agrees with. Your job is to estimate the truth as well as you can with the evidence in hand and to update the plan only when the evidence merits it. That is the work.

The next time you put a body composition reading into a chart, ask: "How much should this update my prior?" Not "what is this number?"

The first question is clinical. The second is just a readout.

The body composition test is also one of three streams a [member compliance audit](/blog/member-compliance-auditing-in-sports-nutrition) reads against — alongside the food log and the performance data — before the prescription gets adjusted. Read in isolation, body comp data overweights single-test noise; triangulated against the other streams, it surfaces what no single number can on its own.