Photo-Based Measurement Agreement for Infant Head Proportion: 3,349 Photographs

Aggregate agreement statistics between an automatic photo measurement of infant
head proportion and human-reviewed reference values, across 3,349 photographs
spanning January 2024 to July 2026.

Two proportions are read from the traced outline of a head in a single
photograph taken from directly above: the width-to-length ratio (conventionally
the cephalic index) and the difference between two diagonals of the outline. Each
is compared against a reference value for the same photograph, reviewed by an
experienced team.

In the headline window the automatic measurement agreed with the reference on a
three-way read of the ratio 85.9% of the time, with a mean absolute error of 2.46
points and a median of 0.56. It recorded 80.5% of the left-right differences the
reference recorded as clearly visible. 12.0% of photographs were off by more than
five points, most often where the traced outline picked up bedding, hair or
clothing along with the head.

The release includes a second window of 910 photographs excluded from the
headline figures, together with the combined totals, so that the effect of the
exclusion can be judged rather than taken on trust. It also reports the
performance of an automatic outline-quality check, including the share of large
deviations that check cannot see.

This is a convenience sample. It is not balanced by age and is not a random
sample of any population; the distribution describes the photographs measured,
not the prevalence of any proportion among children. This is a measurement
agreement study, not a clinical validation: it compares two measurements of a
photograph and says nothing about health, development, or what any value means
for an individual child.

No photographs or per-photograph records are included. The file contains three
rows of summary statistics and nothing else.

Column definitions, the reasoning behind the excluded window, and the limits of
what the data supports are in README.md.