Estimating human age using machine learning on panoramic radiographs for multi-regional brazilian patients

Abstract

This study addresses the challenge of developing machine learning models for age estima- tion based on panoramic radiographs of patients from different regions of Brazil. Using two

geographically diverse datasets — one from UFPE (Northeast) and another from Unicamp (Southeast) — we investigated the limitations of artificial intelligence models in predicting age across distinct regional contexts. We designed an experimental protocol to evaluate the behavior of machine learning models in various scenarios. In the first experiment, the model

trained exclusively on UFPE data showed limitations when tested on Unicamp patients, re- sulting in a mean absolute error (MAE) of 3.10 years on the UFPE dataset and 4.97 years on

the Unicamp dataset, highlighting challenges in generalization.In the second experiment, fine- tuning approaches were explored, which, while improving the model’s performance on regional

data, did not completely eliminate biases. In the third experiment, training the model from scratch on a mixed dataset achieved the best balance between accuracy and generalization,

with an MAE of 3.25 years for UFPE and 3.69 years for Unicamp, indicating greater robust- ness compared to previous approaches. The fourth experiment introduced data augmentation

techniques to enhance the model’s robustness against outliers and extreme cases. Despite marginal improvements, high-magnitude errors persisted, suggesting the need for additional

strategies, such as more advanced data augmentation techniques and more complex archi- tectures. The results of this study reinforce the importance of diverse datasets and rigorous

experimental protocols to address regional variability and distinct demographic characteristics. The model trained on a mixed dataset proved to be the most effective approach, emphasizing that integrating diverse populations is crucial to improving model generalization. Thus, the study provides concrete evidence for the development of more robust systems capable of being reliably applied in clinical and forensic scenarios.