Machine learning methods for determining skin age: A systematic review

Abstract

Aim: This systematic review explores how machine learning is used in determining skin aging, aiming to evaluate accuracy, limitations, and gaps in the current literature.

Materials and methods: OVID Embase, OVID Medline, IEEE Xplore, and ACM Digitial Library were searched from inception to March 16, 2024.

Results: A total of 1467 non-duplicate articles were screened, and 27 were ultimately included in the systematic review. The machine learning models exhibited a range of accuracies from a mean absolute error of 2.30-8.16 years. The most common approach was full facial image analysis, followed by non-image-based studies utilizing biomarkers such as the methylome and the proteome. The incorporation of dynamic facial expressions in the analysis was shown to improve the accuracy of age estimation, with a mean absolute error of 3.74. Confocal microscopy demonstrated potential for accurate skin aging estimation, with some studies achieving up to 85 % accuracy. Many studies were found with high PROBAST risk of bias scores, most commonly due to small sample sizes.

Conclusion: Future studies should aim for greater diversity in ethnicity and variables within datasets to improve generalizability.

Keywords: Artificial intelligence; Biomarker; Image analysis; Machine learning; Skin aging.