{
 "name": "Whole-body fat mass from body measures, against the scan: a mass-closing model, against a published formula, a linear equation and a learned model",
 "registered": "2026-11-16",
 "data": {
  "source": "NHANES 2011-2018 (National Center for Health Statistics, public domain): whole-body dual-energy X-ray absorptiometry (DXX), body measures (BMX), demographics (DEMO); people aged 8 to 59",
  "root": "/home/hilarl/lw-health-1116/data/nhanes",
  "development": ["G", "H", "I"],
  "sealed": "J",
  "split": "by survey cycle, fixed before any value of the sealed cycle was read: development = 2011-2012 (G), 2013-2014 (H), 2015-2016 (I); sealed = 2017-2018 (J), other people in other years. The sealed files were downloaded and hashed, and not opened beyond the file header.",
  "sha256": {
   "DXX_G.xpt": "df4a9741ef55866ca039b0a01f55ab4d3dd55f83f4ee845ca03e8bfca9a65099",
   "DXX_H.xpt": "792ebf3791156367832a01be0c96286ef8da1f540a1d674077f5af12951e99e5",
   "DXX_I.xpt": "56623a1b1c790e43d2c3d9792d30418de8553db53b55ef5e38ab2d00a8e983e8",
   "DXX_J.xpt": "de16420ff3f978f72dbcc6a83c64de041440f29ad1f516f5e1cc5476af994ff1",
   "BMX_G.xpt": "4814bfc3047ed400b9d43d285f8c3ea7c940ac6489404a9b699579715d158ec3",
   "BMX_H.xpt": "fd5e9fc6e6aab0a4aee6e699f51497bbc9b62101f7f43aee924c473e38fd9442",
   "BMX_I.xpt": "d31da84e14212b4e58e8340598a5b8e2144fac83333563e966eb1b332e4141d6",
   "BMX_J.xpt": "8d675e42d8826ac98714b2c3dd4c5138a5e353fb4424f7eff5e6db4a01ce838a",
   "DEMO_G.xpt": "eaf0525d1952626885af3e935415a1f66ad62c18698080e7354789c125af252d",
   "DEMO_H.xpt": "f8f0cbb3085a323d4cde22349b164878fea1e64dbc404e65b5815c7816b547d7",
   "DEMO_I.xpt": "c9297c6c37ae8f78f29be9568fa2a03cf3b112616a39afee04030fc775a66a0d",
   "DEMO_J.xpt": "c0b46e0345ea19404928656277c8b0d10b0cca348a9b2fe4fc3c67e8b7ee73ec"
  }
 },
 "items": "one person per row: a valid whole-body scan (DXAEXSTS = 1) with total fat (DXDTOFAT), lean excluding bone mineral (DXDTOLE) and bone mineral content (DXDTOBMC), and all of weight, standing height, waist circumference, upper-arm circumference, upper-arm length, upper-leg length, age and sex present. People missing any of these are left out of every arm.",
 "arms": {
  "published": "body fat percent from body mass index, age and sex (Deurenberg, Weststrate and Seidell 1991): adults 1.20 BMI + 0.23 age - 10.8 sex - 5.4; children to 15 years 1.51 BMI - 0.70 age - 3.6 sex + 1.4; times weight. Fat only.",
  "linear": "per sex, least squares of each tissue on the seven measures, fitted on the development people",
  "learned": "per tissue, histogram gradient boosting (400 iterations, rate 0.05, seed 0) on the eight measures, trained on the development people",
  "living_worlds": "the body-composition backend of livingworlds/map: lean tissue and bone mineral as power laws with an age curve per sex, fitted in logarithms on the development people; fat is the scanned total less lean and bone, so the three always add up"
 },
 "selection": "the form of the model was chosen on development data only: fitted on 2011-2014 (9,424 people), checked on 2015-2016 (4,537 people). Fat RMSE there: living_worlds 2.49 kg, learned 2.59, linear 2.63, published 5.17. No alternative form was scored on the sealed cycle.",
 "metric": "RMSE in kg of whole-body fat mass over the sealed people (validate.score of livingworlds/map, which refuses a holdout used for fitting). Lean and bone are reported with the same arms. Reported floor = 3% of the mean measured fat mass (the minimum acceptable precision of a scan for total fat mass, ISCD 2013), for reading only. Fat RMSE is also reported by sex, age group and body mass index group.",
 "bar": {
  "on": "fat mass, sealed cycle",
  "rule": "living_worlds RMSE lower than published RMSE AND lower than linear RMSE. The learned arm is reported beside them and is not part of the bar: the backend ships as a few dozen coefficients, and the claim is that those beat the equations in use, not that they beat every model."
 },
 "procedure": "nhanes.py run protocol.json, once, on server2; results.json is committed whatever it says. The coefficients fitted on the development people in that run are the ones the backend then carries.",
 "results": "results.json",
 "code": {
  "map": "livingworlds/map 6df0107e37aec02710b77921144f54ba5349f74b (livingworlds_map/backends/body_composition.py)",
  "engine": "experiment/body/nhanes.py and this file, at the commit that adds this protocol"
 }
}
