External Validation and Interpretability Analysis of a Deep Learning Model for Forensic Age Estimation Across Brazilian and Romanian Populations

Objectives To address critical reliability gaps in forensic AI, this study investigated the generalizability of three CNN architectures for age estimation. The primary goal of the study was to assess whether a model trained on a Brazilian population performs accurately on an independent Romanian dataset across different demographics and imaging systems. Furthermore, we evaluated the best-performing model’s medical reliability through hierarchical Grad-CAM and latent feature embedding analysis to verify biological consistency. Methods A dataset of 10,036 Brazilian orthopanoramic radiographs (OPG) aged 2 to 96 years was used for training and internal testing. External validation was conducted on 150 independent Romanian OPGs collected from three clinical sources in Cluj-Napoca. For interpretability, 512-dimensional feature vectors were extracted and projected via PCA, PHATE and diffusion maps to analyse the model’s internal aging logic. Grad-CAM was applied to representative age-group samples to identify the interpretability of the model across different layers. Results Three CNNs were compared for this study: VGG-16, InceptionV4, and ResNet-18. ResNet-18 was the superior architecture, achieving an internal test Mean Absolute Error (MAE) of 3.61 years, which increased to 5.8 years during external validation. Grad-CAM analysis revealed that while early layers prioritized relevant landmarks like dental follicles and mandibular thickness, final decision layers were susceptible to “shortcuts” such as orientation markers and image borders. Latent space analysis using PHATE visualization suggests that the learned representations followed a chronological trajectory broadly consistent with age progression. Conclusions External validation reveals that AI models for forensic age estimation are highly vulnerable to population shifts and hardware-driven distribution changes. While latent space manifold projection via PHATE suggests that CNN can successfully capture and map structures of biological aging patterns, the Grad-CAM analysis indicated that, although anatomically relevant structures were activated, some deeper representations also involved non-anatomical regions, suggesting a potential susceptibility to shortcut learning. This highlights the critical importance of independent external validation as a foundational quality control metric for supporting the assessment of reliability, transparency, and potential forensic applicability.