Table 1_Explainable deep learning reveals a stagewise temporal deterioration from skin tone to contour to texture in Chinese women’s skin aging.pdf

A hybrid multimodal deep learning classification model was developed by integrating facial image data with quantitative facial aging indicators. Five-fold stratified cross-validation was performed to evaluate model performance, with Top-2 accuracy used as the primary evaluation metric. SHAP analysis was further applied to assess the contribution weights of different facial aging indicators to model predictions.

Based on the fitted quantitative facial aging indicators, a six-class facial aging classification model was established. The model demonstrated stable convergence and achieved a highest validation Top-2 accuracy of 85.0%. The confusion matrix showed that classification errors occurred primarily between adjacent aging stages.

SHAP analysis revealed a progressive aging pattern characterized by early changes in skin color, middle-stage changes in facial contour, and later deterioration in skin texture. This study demonstrates that integrating deep learning with SHAP analysis can effectively classify facial aging stages in Chinese women while providing interpretable insights into the key features driving aging-related predictions.</p>