PII-126 - DEVELOPMENT OF A SKIN DRYNESS PREDICTION MODEL BASED ON FACIAL IMAGES USING DEEP LEARNING: A PROSPECTIVE COHORT STUDY.

H. Yang¹, M. Lee¹, N. Han¹; ¹jeju National University, jeju-si, Republic of korea.

Background: Skin aging increases water loss of the skin (xerosis), leading to dermatoses such as itching, ulcers, dermatitis, and fungal infections. Although measuring skin hydration is important for the detection of primary or secondary xerosis caused by medication such as isotretinoin and diuretics, existing measurements may not be clinically applicable due to the lack of accuracy and precision. This study aimed to develop a deep-learning model to predict skin dryness using facial images.

Methods: This study enrolled 300 participants (ages 18 to 65) in a single group. Demographic data and the Baumann’s Skin Type Indicator score were collected. Facial images were captured using the Visia-CR and the hydration/dryness value was measured on both cheeks by corneometry (Corneometer® CM825). Skin types were classified as ‘Dry’ or ‘Non-dry (or hydrated)’ based on the 40 A.U. threshold of Corneometer measurements. Images were augmented by rotating and flipping and preprocessed using the CLAHE method. Various models including Vision Transformer, EfficientNet_V2, and others, and the best-performing model was selected. Grid search was used to optimize hyperparameters. and model performance was evaluated using accuracy, loss function, recall, specificity, and F1-score.

Results: : Among 300 participants, 217 were female and 83 were male, with ages evenly distributed. Skin dryness value and Baumann Skin Type Indicator scores showed a significant negative correlation. (r = - 0.16, p = 0.004). A total of 3,648 hydrated images and 1,152 dryness images were used for model training and testing. The Vision Transformer model performed best, with a loss of 95.46% and an accuracy of 78.28%, which improved to 80.08% after hyperparameter tuning. The precision, recall, specificity, and F1-score of the final model were 1.22%, 96.11%, 28.57%, and 0.8804, respectively.

Conclusion: Our model for the prediction of skin dryness showed good accuracy compared to previously developed models, which offer a meaningful method for predicting skin conditions using a deep learning model. This model has high clinical applicability for early detection of skin diseases and drug side effects.