Latent spaces of generative models for forensic age estimation

• DOI: 10.1007/s00194-025-00745-9
• Publication Date: 2025-03-03T15:40:31Z
• Authors: Anastasia Chernys...; Christopher Gundler; Alexander Johanne...; Eilin Jopp-van Well; Axel Heinemann; Benjamin Ondruschka
• Keywords: Forensic psychiatry; Generative model

Abstract

Background Machine learning may significantly support forensic medicine, particularly in age estimation through medical imaging soon. This technology offers great potential for supporting decisions especially when age documentation is missing or disputed.

Objective This study investigates the potential of generative models for forensic age estimation. The focus is on addressing the challenges of interpretability and generalizability commonly faced by traditional discriminative models.

Methods We applied a family of generative models to postmortem computed tomography (PMCT) scans, focusing on the ossification of the medial clavicular epiphysis for age prediction. The latent space representations from these models were analyzed for their ability to predict age accurately and interpretably across different datasets.

Results While the methods did not perform as well as discriminative state of the art approaches using German Working Group for Forensic Age Diagnostics (AGFAD) guidelines, the variational autoencoders were able to learn a meaningful latent space. The age of the participants could even be visualized within a two-dimensional projection. Additionally, the re-use of the learned space led to high performance on a smaller dataset collected from a forensic center.

Conclusion The consideration of “soft factors”, such as explainability in addition to absolute performance remains crucial for bringing machine learning methods into forensic practice. Depending on the set-up, generative models might be attractive for assessing the reasoning within models and sharing information between datasets.