Comparative Accuracy of a Commercial Multimodal Large Language Model Versus Clinicians in Dental Age and Sex Estimation from Panoramic Radiographs: A Benchmark Study
Background/Objectives: Routine dental imaging captures cumulative dentoalveolar and periodontal changes, offering a non-invasive window into both chronological aging and physiological senescence. While task-specific deep learning models have been widely explored in dental radiology, the zero-shot diagnostic performance of commercial Multimodal Large Language Models (MLLMs) in adult populations remains underinvestigated. The aim of this study was to benchmark the zero-shot diagnostic accuracy and repeatability of a commercial MLLM against human clinicians of varying experience levels in estimating adult dental age and biological sex from panoramic radiographs (OPGs), using documented chronological age and biological sex as the ground truth. Methods: From 216 initially screened records, this retrospective single-center study included 127 standardized, anonymized OPGs from patients (74 men, 53 women; mean chronological age 55.5 ± 15.8 years, range 18–90 years) treated at the Clinical Department of Cranio-Maxillofacial Surgery, Military Institute of Medicine—National Research Institute in Warsaw. Each radiograph was independently evaluated across three trials by Gemini 3.1 Pro (Google, Mountain View, CA, USA) and once by two clinicians—a maxillofacial specialist (>10 years of experience) and a resident (<3 years of experience)—all blinded to clinical metadata. Results: Gemini 3.1 Pro demonstrated excellent repeatability across the three trials (ICC(A,1) = 0.926, ICC(A,3) = 0.974; p = 0.156) and moderate agreement with chronological age (ICC(A,1) = 0.604), but systematically overestimated age by a mean of 6.23 years. Mean absolute error (MAE) ranged from 12.91 to 13.03 years across the AI trials. The specialist achieved a significantly lower MAE (10.39 years) than all AI assessments (adjusted p ≤ 0.003), whereas the resident (MAE = 11.19 years) did not differ significantly from the AI (adjusted p = 0.282). In biological sex classification the specialist (66.1%) significantly outperformed all three AI trials (41.7–49.6%), while the resident (62.2%) significantly outperformed the first two AI trials (41.7% and 43.3%), but did not differ significantly from the third trial (49.6%). Across all runs, the AI exhibited significant directional asymmetry (p < 0.001), frequently misclassifying male patients as female. Conclusions: While Gemini 3.1 Pro demonstrated high internal repeatability and age-estimation accuracy comparable to a junior resident, its wide limits of agreement, systematic age overestimation, and poor sex-classification accuracy preclude its autonomous use in clinical or forensic workflows. Future studies incorporating healthy control cohorts and validated clinical indices are required to determine whether the observed age overestimation reflects inherent algorithmic miscalibration or sensitivity to cumulative dentoalveolar pathology in hospitalized patients.
Authors
- Dariusz Jurkiewicz (ORCID: https://orcid.org/0000-0003-3729-2679)
- Maria Sobol (ORCID: https://orcid.org/0000-0002-5105-0687)
- Aldona Chloupek (ORCID: https://orcid.org/0000-0003-2439-7199)
- Kevin Du
- Dominika Zawadka-Modras (ORCID: https://orcid.org/0009-0009-5406-7888)
Institutions
- Medical University of Warsaw (PL)
- Wojskowy Instytut Medyczny (PL)
Publication Details
- Journal
- Journal of Clinical Medicine
- Published
- 2026-11-20
- DOI
- https://doi.org/10.3390/jcm15229123
- Primary Topic
- Forensic Anthropology and Bioarchaeology Studies
- Type
- article
- Field-Weighted Citation Impact
- 0.00