Impact of age partitioning on classification discordance in pediatric ferritin reference intervals
Dien Minh Tran, Dung Kim Le, Thao Thu Ngoc, Trang Thi Thu Tran, Ronda F. Greaves, Tze Ping Loh, Mai Thi Chi TranAbstract
Objectives
To evaluate how alternative age-partitioning strategies affect result classification using discrete pediatric ferritin reference intervals (RIs).
Methods
In a prospective cohort of 3,322 apparently healthy individuals aged birth to <19 years, serum ferritin was measured on the Cobas Pro platform. A common preprocessing pipeline, including Horn’s algorithm, was applied once before all comparisons. Five age-partitioning schemes were evaluated: a data-driven classification and regression tree (CART)-based scheme (Scheme 1, reference), a physiology-informed high-resolution scheme, a CLSI-informed scheme, and two progressively broader pragmatic schemes. Nonparametric RIs were derived, and classification performance was assessed using apparent and out-of-bag (OOB) flagging rates, age-resolved analyses, discordance relative to Scheme 1, and boundary-focused and Lahti’s evaluations.
Results
Overall flagging rates were similar across schemes (5.06–5.33 % apparent; 5.23–5.80 % OOB). However, substantial age-specific differences were observed. Discordance relative to Scheme 1 was lowest for Scheme 3 (1.99 %) and similarly low for Scheme 2 (2.26 %), but increased for Schemes 4 and 5 (3.76–4.64 %). Boundary discordance reached 15.0 % at 6 months in broader schemes and reached 11.84 % at adolescent transitions. Age-resolved analyses showed marked heterogeneity during infancy, with overall flagging rates varying up to threefold between schemes within adjacent monthly bins. Lahti’s evaluation confirmed non-interchangeability of adjacent partitions, particularly in early infancy.
Conclusions
Despite similar overall flagging rates, marked age-specific classification differences occur, particularly in early infancy and around key developmental boundaries. Broader partitions may obscure physiological variation and increase classification discordance. Age-resolved evaluation should complement statistical criteria when defining pediatric RIs.