Photo Credit: Istock.com/metamorworks

The following is a summary of “Artificial Intelligence Predicts Fitzpatrick Skin Type, Pigmentation, Redness, and Wrinkle Severity From Color Photographs of the Face,” published in the March 2025 issue of Journal of Cosmetic Dermatology by Draelos et al. 

High demand has led non-dermatologists to perform skin assessments and laser treatments, increasing complications, while machine learning may enhance accuracy.  

Researchers conducted a retrospective study to design a high-performing machine learning model that simultaneously predicted Fitzpatrick skin type, hyperpigmentation, redness, and wrinkle severity.  

They created the Skin Analysis dataset using 3,662 images, each labeled by a dermatologist across 5 skin scales. Machine learning models were trained and evaluated across 15 different configurations, incorporating 3 neural network architectures and 2 loss functions.  

The results showed that the top-performing model utilized the EfficientNet-V2M architecture with a custom cross-entropy loss function. This model achieved a mean test set accuracy of 85.41% ± 9.86 and a mean test set AUROC of 0.8306 ± 0.09599. Performance trends indicated higher accuracy at the extremes of the scales, suggesting increased clinical ambiguity in the mid-range values.  

Investigators concluded that machine learning models successfully predicted multiple skin characteristics concurrently from facial color photographs, suggesting the potential to aid non-dermatologists in patient skin evaluation for improved treatment planning in the future. 

Source: onlinelibrary.wiley.com/doi/full/10.1111/jocd.70050