Explainable Ensemble Learning for Human Age Prediction from a Novel Clinical Full-Body Thermal Imaging Dataset

Background: Chronological age estimation from medical and biomedical images has been investigated using facial photographs, neuroimaging, and physiological biomarkers; however, the potential of full-body infrared thermography remains insufficiently explored. This study aimed to develop an interpretable machine-learning framework for chronological age prediction from full-body thermal images and to identify age-associated thermal features across anatomically defined body regions. Methods: A clinically acquired infrared thermal imaging dataset comprising 2263 subjects aged 19–85 years was analyzed. Thermal information was extracted from eleven anatomically defined regions of interest using statistical and gradient-based descriptors. Kendall’s Tau-based feature selection was applied using the training data, and multiple regression models were evaluated individually and within a stacking ensemble framework. Model performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), and Pearson correlation coefficient (PCC). LIME was used to examine feature contributions, while predicted-versus-chronological age plots were examined to characterize prediction patterns and age-dependent errors. Results: In 10-fold cross-validation, the stacking ensemble achieved a mean MAE of 6.46 ± 0.60 years on the full cohort (19–85 years) and 6.13 ± 0.31 years on the age-range-restricted cohort (22–67 years). Five repeated runs on the restricted cohort yielded a mean MAE of 5.77 ± 0.07 years. Explainability analysis identified both temperature- and gradient-based features from multiple anatomical regions as influential predictors. Conclusions: Full-body infrared thermography shows promise for chronological age prediction, with stacking regression demonstrating competitive performance. LIME identified influential regional thermal features. Further validation is needed to assess generalizability and clinical relevance.