Is my smart scale's body-fat number accurate?
Not for one reading: vs a lab 4-compartment model, the best consumer scales could miss by about ±6.5 points, most foot-to-foot ones by ±11-15, a 3-month change by ±3.5-5 (our arithmetic).
Siedler et al. asked how repeatable consumer bioimpedance scales are, how close their body-fat percentage comes to a laboratory four-compartment model, and whether they track change.
Seventy-three healthy adults, standardised for fasting, hydration and exercise, stood on 15 devices in random order — one hand-to-hand device, ten foot-to-foot scales and four octapolar (hands and feet) models, three of them consumer-grade and one research-grade — and were measured by a 4C model built from air-displacement plethysmography (body volume), DXA (bone mineral) and bioimpedance spectroscopy (total body water).
Thirty-seven came back 12-16 weeks later. Immediate test-retest precision was high (precision error 0.0-0.49%), but in five foot-to-foot scales suspiciously so (repeat readings never more than 0.1% apart), which the authors read as the device repeating its last number for the same user.
Against the 4C model the constant error ran from −3.5 to +11.7 body-fat points; only five devices were within ±2 points of it at the group level (the research-grade Seca mBCA, InBody H20N, Withings Body Cardio, Tanita BC-554 and Tanita UM-081).
For change over 12-16 weeks, when the true mean change was +0.2 points (SD 2.9), the devices' constant error was small (−0.4 to +1.3) but individual agreement was poor (standard error of the estimate 1.7-2.6 points, concordance 0.37-0.78), and seven devices compressed change systematically.
Our reading: a well-standardised study with a gold-standard criterion; for an athlete, a single body-fat reading from a scale carries an individual error of several points even on the best models, and a three-month change of a few points is inside the noise of every device tested.
Protocol
Design: prospectively registered (ClinicalTrials.gov NCT05026697), Texas Tech University IRB (IRB2021–107); two visits, the second optional, 12-16 weeks apart (mean 89.2 ± 5.8 d); participants could lose, gain or keep weight in between.
Who: n = 73 (35 men, 38 women) at the first visit, age 26.2 ± 7.3 y, 72.0 ± 16.8 kg, BMI 25.3 ± 4.7, 4C body fat 24.7 ± 9.1%; 37 (21 men, 16 women) at follow-up. Exclusions included implanted electrical devices and anything beyond the devices' height or weight limits.
Standardisation: 24 h without exercise; 8 h without food, fluid, caffeine, alcohol or nicotine, after 1 L of water before the fast; visits between 06:00 and 14:00; bladder voided, urine specific gravity 1.018 ± 0.005; skin-tight clothing; at least 10 min standing before the first device.
Devices (Table 2 order; letters A-O as in the text, mapped by our reading of Table 2): A Omron HBF-306 (hand-to-hand); foot-to-foot B RENPHO, C HAWANA, D Wyze, E INEVIFIT, F Weight Watchers, G Tanita UM-081, H Vitagoods Form Fit, J Seca 804, K Withings Body Cardio, L Tanita BC-554 Ironman; octapolar I Omron HBF-516, M Tanita BC-568 InnerScan, N InBody H20N and the research-grade O Seca mBCA 515/514.Chosen from Amazon best-sellers and makers known to the authors, about $24-349. Height entered to the device's precision; birthday as 1 January of the birth year; activity level at the lowest setting.
Measurement: each device twice in succession at visit 1 (stepping off and on), the first reading used for validity; once at visit 2.
Criterion: 4C model (Wang 2002 equation) from body mass, body volume by BodPod, bone mineral by DXA (GE Lunar Prodigy) and total body water by bioimpedance spectroscopy (ImpediMed SFB7).
Analysis: precision error and least significant change (2.77 × precision error); constant error, total error (RMSE), standard error of the estimate, Lin's CCC, Bland-Altman limits of agreement with proportional-bias regression, OLS and Deming regression; equivalence by two one-sided tests with ±2 points (cross-sectional) and ±1 point (change); a 1-15 ranking per domain summed to a global score.
Funding and interests: paid from Texas Tech University research funds given to the principal investigator (G.M. Tinsley); no external entity provided equipment or money for this study. The authors declare no relevant conflicts; Tinsley has received in-kind equipment loans or donations from body-composition device makers including Biospace, Inc. (InBody) and research funding from Prism Labs, stated as playing no role here.
Key Results
Body-fat percentage, each device vs the 4C model (Tables 2 and 3; constant error = device minus 4C, in body-fat points):
| Device (configuration) | Cross-sectional constant error | Total error | Equivalent ±2 | Limits of agreement, single reading (our arithmetic) | Change over 12-16 wk: constant error | Total error of change | Limits of agreement for change (our arithmetic) |
|---|---|---|---|---|---|---|---|
| Seca mBCA 515/514 (octapolar, research) | −0.2 | 3.3 | Yes | −6.7 to +6.3 | 0.3 | 2.1 | −3.8 to +4.4 |
| InBody H20N (octapolar) | −0.1 | 3.3 | Yes | −6.6 to +6.4 | −0.4 | 2.1 | −4.5 to +3.7 |
| Tanita BC-568 InnerScan (octapolar) | −1.4 | 4.4 | No | −9.6 to +6.8 | 0.7 | 1.9 | −2.8 to +4.2 |
| Omron HBF-516 (octapolar) | 3.3 | 5.6 | No | −5.6 to +12.2 | 0.8 | 1.9 | −2.6 to +4.2 |
| Withings Body Cardio (foot-to-foot) | −0.4 | 5.8 | Yes | −11.8 to +11.0 | 0.6 | 1.9 | −3.0 to +4.2 |
| Tanita BC-554 Ironman (foot-to-foot) | 0.7 | 5.8 | Yes | −10.7 to +12.1 | 0.6 | 2.0 | −3.2 to +4.4 |
| Tanita UM-081 (foot-to-foot) | 0.6 | 5.8 | Yes | −10.8 to +12.0 | 0.7 | 1.9 | −2.8 to +4.2 |
| Omron HBF-306 (hand-to-hand) | −3.5 | 5.4 | No | −11.6 to +4.6 | 0.2 | 2.1 | −4.0 to +4.4 |
| Other foot-to-foot scales (RENPHO, HAWANA, Wyze, INEVIFIT, Weight Watchers, Vitagoods) | −1.4 to +3.6 | 6.8-7.8 | No | about ±12-15 around the constant error | −0.3 to +0.3 | 2.1-2.6 | about ±4-5 |
| Seca 804 (foot-to-foot) | 11.7 | 12.6 | No | +2.5 to +20.9 | 1.3 | 2.9 | −4.0 to +6.6 |
The paper prints each device's limits of agreement only inside its Bland-Altman figures.
The limits above are constant error ± 1.96 × the SD of the differences, with that SD recovered from the total error (and cross-checked against the 95% CI of the constant error); this reproduces the cross-sectional SDs the abstract gives (4.1 for the Omron HBF-306, 4.7 for the Seca 804) and its change SD of 2.1 for the InBody H20N; for the Seca 804's change the reconstruction gives 2.6, so the abstract's 2.7 is used.
Our arithmetic, not the authors' figures.
Reliability: mean difference between immediate repeats −0.06 to +0.07%, maximum absolute difference 0.0-2.5%, precision error 0.0-0.49%, least significant change 0.0-1.36%.Five foot-to-foot devices (RENPHO, HAWANA, INEVIFIT, Weight Watchers, Vitagoods) never differed by more than 0.1% between repeats, which the authors treat as non-independent readings (the scale likely repeats a recent result for the same user account) and rank last for reliability.The octapolar models M-O had the highest precision error, which the authors attribute to hand placement and more measured segments.
Cross-sectional validity: constant error −3.5 (sd 4.1) to 11.7 (sd 4.7) points, standard error of the estimate 3.1-7.5, total error 3.3-12.6, CCC 0.48-0.94. Without the Seca 804 the largest constant error was 3.6 and the largest total error 7.8. Six devices (the hand-to-hand Omron and five foot-to-foot scales: RENPHO, HAWANA, Wyze, Weight Watchers, Vitagoods) showed significant negative proportional bias: they read lean people too high and fatter people too low.
Longitudinal validity: true (4C) change +0.2 (sd 2.9) points, range −5.0 to +7.5; device constant error −0.4 (sd 2.1) to 1.3 (sd 2.7), standard error of the estimate 1.7-2.6, total error 1.9-2.9, CCC 0.37-0.78; nine devices equivalent within ±1 point at the group level; seven (the hand-to-hand device and six foot-to-foot scales) showed significant negative proportional bias, i.e. they shrank real gains and losses toward zero.
Global ranking (sum of five domain ranks): 1 Omron HBF-516, 2 Tanita BC-568, 3 Tanita UM-081, 4 Seca mBCA, 5 InBody H20N, 6= Withings Body Cardio and Tanita BC-554; last HAWANA (15) and RENPHO and Weight Watchers (13=).
Mechanisms
Measured: octapolar models and some foot-to-foot scales agreed best with the 4C model, both for a single reading and for change; the hand-to-hand device and most cheaper foot-to-foot scales showed proportional bias.
Authors' interpretation: the octapolar models measure resistance and reactance through all four limbs rather than two, which, with variable hand placement, may explain their slightly worse repeatability (the authors do not offer it as the reason for their better validity); the prediction equations are not published, and devices that did well may have been developed on samples like this one (speculative, in the authors' words).
Authors' interpretation (bias and repeats): negative proportional bias in six devices suggests an output range restricted toward "normal" values, or prediction equations built on a narrow reference population; very low repeat variability in five scales suggests a built-in limit on how much repeated readings may differ for the same user within minutes.
Inference: an endurance athlete is lean compared with this sample (mean 24.7%), and the devices with negative proportional bias read lean people high; such a scale would tend to overstate an athlete's body fat. The "athlete" modes on many scales use different equations and were not tested, so their error is unknown in either direction.
Inference: bioimpedance reads body water; anything that moves water — the session just done, glycogen loading or depletion, a hot day, sodium, the menstrual cycle — moves the body-fat number without moving fat. This study removed those effects by standardisation; a home measurement usually does not.
Practical Application
Is my smart scale's body-fat number accurate? Not for one reading: vs a lab 4-compartment model, the best consumer scales could miss by about ±6.5 points, most foot-to-foot ones by ±11-15, a 3-month change by ±3.5-5 (our arithmetic).
Do not set or chase a body-fat target from a scale (B, 73 healthy non-athletes): even the best-agreeing consumer device (InBody H20N) put an individual anywhere within about ±6.5 points of the laboratory value, and the Withings Body Cardio within about ±11 (our arithmetic from the paper's tables).This gives measured support to SYNTH-54's rule never to set a body-fat-percentage target; scale weight itself was not tested here, and the base's own sweat-rate and weight-trend practice (SYNTH-68, SYNTH-54) rests on body mass, not on bioimpedance.
A change of a few points over a few months is not a finding: true change in this sample was +0.2 ± 2.9 points over 12-16 weeks, and every device's individual error for change (about ±3.5-5 points) was larger. Seven devices also shrank real gains and losses toward zero.
If a scale is used at all, use it the way the authors advise: the same device, fasted, after voiding, minimal clothing, every morning, a multi-day average rather than one reading; octapolar models over foot-to-foot or hand-to-hand. A scale that gives exactly the same number twice in a minute may simply be repeating itself.
Naming matters: these numbers belong to the named models, on firmware the paper does not report (one device's algorithm changed mid-study), in "normal" mode; Garmin Index S2 and the scales' athlete modes were not tested.
Limitations
Not athletes, and athlete settings not used. Healthy adults with a mean body fat of 24.7%; activity level set to the lowest setting on every device. The authors call for work on "athlete" modes.
Strict standardisation. Fasted, hydrated, rested, morning-to-early-afternoon measurements; home use without that control would add day-to-day biological error, which the immediate-repeat precision here does not include (the authors' point).
Limits of agreement not in the text. Device-level limits of agreement are printed only in the figures; the values in this entry are our reconstruction from the tables.
Firmware. Firmware versions are not reported; one device (Wyze, device D by our mapping) received an algorithm-changing firmware update during data collection.
Small change, few returners. 37 of 73 returned, and the true mean change was +0.2 points, so group-level equivalence for change (±1 point, nine devices) mostly shows that a device reported little change on average; it was the devices that were not equivalent by this test (Tanita, Omron HBF-516, Withings) that the authors rank best for tracking change, by total error. Both views are in the paper; the individual limits for change are the athlete's number.
Reporting. The methods describe four octapolar models (I, M, N, O); the reliability results and the discussion refer to "three of the five" octapolar devices (M-O).
Interests. Internal university funding and no external role in this study; the principal investigator has received in-kind equipment from InBody (Biospace), whose H20N was among the two best-agreeing consumer devices, and from other body-composition makers, and research funding from Prism Labs.
How this entry was checked and changed
Why this level
cross-sectional measurement study with a gold-standard criterion, n = 73 (37 re-tested after 12-16 weeks), read from the full text (open access, CC BY 4.0). The rubric places a cross-sectional or test-validation study at B and needs ≥100 for B+; one lab, healthy non-athletes, and the devices' "athlete" settings were not used
Full text (checked 2026-10-27)
Seventy-three healthy participants aged 19–50 years were assessed by each of fifteen BIA devices; thirty-seven returned 12–16 weeks later (full text)
Precision error ranging from 0.0 to 0.49 %; constant error relative to the 4C model ranging from −3.5 (sd 4.1) % to 11.7 (sd 4.7) %, standard error of the estimate values of 3.1–7.5 % and Lin's CCC of 0.48–0.94 (full text)
For longitudinal validity, constant error ranged from −0.4 (sd 2.1) % to 1.3 (sd 2.7) %, standard error of the estimate 1.7–2.6 % and Lin's CCC of 0.37–0.78 (full text)
The mean change in BFP as assessed with the 4C model was 0.2 (sd 2.9) %, ranging from −5.0 % to 7.5 % (full text)
Table 2 constant error and total error: InBody H20N –0.1 and 3.3; Seca mBCA 515/514 –0.2 and 3.3; Withings Body Cardio –0.4 and 5.8; Tanita BC-554 Ironman 0.7 and 5.8; Omron HBF-516 3.3 and 5.6; Seca 804 11.7 and 12.6 (full text)
Five foot-to-foot devices (B, C, E, F and H): maximum absolute difference 0.0 % to 0.1 %, PE 0.0–0.04 % and least significant change 0.0–0.11 % (full text)
Six of the fifteen devices exhibited negative proportional bias cross-sectionally; seven (47 %) for change over time (full text)
Activity level "was entered as the lowest possible setting (e.g. '1'/sedentary or 'normal' mode)" (full text)
"For one device (D), a firmware update occurred during the initial data collection period, and this update was confirmed to include a change to the algorithm" (full text)
Funded from research funds "provided to the principal investigator (GMT) by Texas Tech University"; G. M. T. has received in-kind support from makers including "Biospace, Inc. (DBA InBody)" (full text)