Artificial Intelligence Based Skin Analysis Models for Predicting Visual Grades and Device Measured Physiological...

This study aimed to develop and validate an artificial intelligence (AI)-based skin assessment framework capable of predicting both dermatologist-assigned visual grades and device-derived physiological measurements from facial images. In addition, we evaluated the impact of image-acquisition modality (DSLR, tablet, or smartphone) on model performance.

A total of 1,099 Korean participants aged 14-69 years were enrolled. High-resolution facial images were obtained from seven standardized angles using a DSLR camera, and a subset of additional images was collected using a tablet and participant-owned smartphones. Five board-certified dermatologists graded eight facial signs, and physiological parameters were measured using non-invasive skin assessment devices. CoAtNet-4 served as the backbone architecture for both classification (visual grading) and regression (physiological prediction). Model performance was assessed using mean absolute error and correlation analyses.

When evaluated using DSLR images, the model achieved a mean exact-grade accuracy of 51.4%, with an average of 93.3% of predictions falling within ± 1 grade across facial signs. High correlations were observed for wrinkles, pigmentation, pores, and sagging, whereas lip dryness demonstrated comparatively lower correlations. For physiological metrics, strong correlations were observed for pigmentation spot count, cheek pore visibility, and wrinkle severity, whereas hydration and elasticity showed moderate correlations. Performance on mobile-device images remained high and showed strong agreement with DSLR-based predictions, although a noticeable decline in pigmentation-related accuracy was observed.

The proposed AI framework reliably approximates dermatologist visual grading and multiple device-based physiological measurements, offering a comprehensive, image-driven approach to skin aging assessment. However, model performance varies across facial attributes and remains sensitive to image quality, emphasizing the need for domain adaptation and image enhancement strategies to ensure robust application in consumer environments.