Anthropometric PHV-prediction models versus skeletal age in elite Chinese U12–U13 male soccer players: a cross-sectional study

Abstract

Assessing biological maturity in youth soccer helps reduce maturity-related selection bias. Anthropometric peak height velocity (PHV) prediction models provide non-invasive, repeatable tools for maturity monitoring, but their performance relative to skeletal-age (SA) indicators remains insufficiently examined in Chinese youth athletes. We evaluated the association, discrimination, and age-group performance of PHV-derived indicators (MO and APHV) from three models against SA in Chinese elite U12–U13 male soccer players. In this cross-sectional study, 313 elite male players (123 U12, 190 U13) underwent anthropometry and left hand-wrist radiography. SA was assessed with the CHN05 standard by two blinded experts. MO and APHV were computed with the Mirwald [1], Moore [2], and Fransen [3] models. Correlation, one-way ANOVA, ordinal logistic regression, and ROC analyses compared raw PHV-derived indicators with SA-derived maturity status. SA inter-rater reliability was excellent (U12 ICC = 0.967; U13 ICC = 0.992). SO was negatively correlated with APHV ( r = − 0.797 to − 0.651) and SA was positively correlated with MO ( r = 0.573 to 0.716; all p < 0.001); these associations remained significant after adjustment for chronological age. All models differentiated SA-based maturity groups (all p < 0.001; η² = 0.22 to 0.53), with APHV models yielding higher Nagelkerke pseudo-R² than MO models. ROC discrimination was acceptable-to-excellent (AUC = 0.757 to 0.964, higher for APHV than MO). Individual error was larger in U13 than U12 (MAE 0.45 to 0.70 years). Bland–Altman limits of agreement were wide ( ≈ ± 1.1 to ± 1.8 years) with significant proportional bias. The PHV-derived indicators showed meaningful association and discrimination relative to SA and can supplement squad-wide monitoring. At the group level the three models were broadly comparable, with none uniformly superior and APHV outperforming MO. At the individual level, accuracy was limited: the limits of agreement exceeded the ± 1-year maturity-category band and showed proportional bias. Model performance differed by purpose and age group: the Moore model produced the narrowest range and largest individual-level error and suited group ranking, whereas Fransen was closest to SA in U12 and Mirwald in U13. These findings support their supplementary use in this population and the need for future longitudinal calibration.