What was compared
Every method below was applied to the healthy children of the PHN public database (after the same cleaning as the model). This model, cross-validated scores each child with a model fitted without them (10 folds), so it has no home advantage. The PHN, Mahgerefteh and CSA/height methods were derived from these same children, so they do. Lauzon-Schnittka is the only reference built on a different cohort (a Canadian multicentre clinical cohort read by many readers), computed with the authors' own R function; it tests how well a reference transfers to another lab. Mahgerefteh is published for ages 2 and up, and Chaudhary CSA/height covers the sinuses and ascending aorta only.
All figures are group summaries (whole-cohort statistics and 20 equal-count groups of about 160 children); no individual records are published. Comparison generated —.
What it shows
- This model holds up out of sample. Cross-validated z-scores keep SD 1.01, 2.1–2.7% beyond each of ±2 and no trend against height, age, build, BMIz or sex. The sinotubular junction has slightly heavy tails.
- Ignoring weight costs about 0.2 z per SD of build (height-and-sex model), and changes about 1 in 4 of the z > 2 calls.
- BSA0.5 indices overcorrect for weight (heavier children score about 0.1–0.15 z lower per SD of build), are right-skewed, and leave boys about 0.1–0.5 z above girls.
- Dividing by height overcorrects for height (Mahgerefteh: taller children score about 0.2–0.3 z lower per SD), because the aorta grows less than in proportion to height.
- Lauzon-Schnittka transfers in shape but not in level. Its trends against body size are nearly flat in these children, but the spread is narrower (SD about 0.9 at the sinuses and junction) and its ascending aorta reference sits about half a z-score higher than these measurements, so it flags only a third of the children this model places above +2. A different measurement convention or reader population is a likely reason.
- CSA/height is right-skewed (skewness about 0.5): 3–4% of healthy children land above +2 and under 1% below −2, with trends left against build and BMIz. The continuous version as distributed is also offset (sinuses −0.5 z; ascending aorta +0.3 z, flagging 6.5% of healthy children).
- Extra weight raises the expected diameter only a little: at BMIz +3, about 2–3.5% in this model and 3–5% for Lauzon-Schnittka at the sinuses and ascending aorta, against 7–17% implied by BSA0.5.
Scorecard:
Trend columns give the change in mean z per standard deviation of each variable (a slope of 0.2 means the tallest children run about 0.8 z higher than the shortest). Boy − girl is the difference in mean z. Shaded cells are off target: ! worth noting !! clearly off. Thresholds: mean beyond ±0.1 / ±0.25; SD outside 0.95–1.05 / 0.9–1.1; skewness beyond ±0.1 / ±0.25; tail outside 1.6–3.0% / 1.2–4.0%; trends and sex difference beyond ±0.05 / ±0.15.
| Method | Children | Mean | SD | Skew | Beyond (%) | Trend in z per SD of | Boy − girl | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| +2 | −2 | Height | Age | Build | BMIz | ||||||
Build is weight for height (weight relative to the typical weight for the child's height). BMIz is the WHO BMI-for-age z-score.
Where each method drifts:
Mean z (or its SD) in 20 equal-count groups along each variable. A good reference stays flat at 0 (or at 1 for SD). Choose methods with the buttons; hover a chart for the values in each group.
Who gets flagged above +2:
Agreement with the cross-validated model on z > 2. Expected in healthy children: about 2.3%, roughly 73 children. Both flagged by both; this model only flagged here and missed by the other method; other only flagged by the other method alone.
| Method | Children | Both | This model only | Other only | Flagged above +2 |
|---|
What extra weight does to the expected diameter:
Three example children at a fixed height and age, made heavier or lighter (WHO BMI-for-age z from −2 to +4). Each line is the expected (median) diameter under that method, as a percentage of its value at BMIz 0. Two independent models, this one and Lauzon-Schnittka, are compared with the PHN BSA0.5 index. This model is held flat beyond the build range of the PHN children (dashed).