AI-based bone age assessment versus forensic expert panel: agreement, sex bias, and ethical implications

Forensic age estimation is frequently requested when chronological age is unknown or contested. While automated bone age systems have been validated in clinical populations, their performance against committee-based forensic reference standards-and their compatibility with the AGFAD-compliant Mindestalter (minimum chronological age) framework-has not been systematically addressed in cases of unknown chronological age.

In this retrospective single-centre study, we investigated the technical agreement between BoneXpert and an institutional Forensic Expert Panel (FEP) -the legally recognised reference standard in Turkish forensic practice-rather than validating BoneXpert as a stand-alone forensic tool. Hand-wrist radiographs of 128 cases (44 males, 84 females; 2016-2024) were reassessed using the Greulich-Pyle atlas (FEP) and BoneXpert. Agreement and bias were analysed using pairedt-tests, intraclass correlation coefficients, and Bland-Altman plots. In addition, a complementary minimum-age threshold analysis was performed within the Mindestalter framework using the minimum chronological ages reported by Tisè et al. RESULTS: Excellent agreement was observed (overall ICC = 0.956; males 0.987; females 0.907). A small but significant bias (-0.24 years) indicated underestimation by BoneXpert, driven by female cases (-0.34 years; no significant bias in males). Under the Mindestalter framework, BoneXpert never classified a case as proven above a legal threshold when the FEP did not. Concordance was 93.8% (κ = 0.834) at the 14-year threshold and 80.5% (κ = 0.604) at the 15-year threshold. At the 18-year threshold, neither method could prove that the threshold had been reached using hand-wrist GP-based assessment alone.

BoneXpert provides consistent and ethically conservative results but should be used only as a decision-support tool within a multidisciplinary expert-led workflow, in line with AGFAD recommendations.