Estimating Transepidermal Water Loss from Smartphone Facial Images Using Self-Supervised Features and a Mixture Density Network

<jats:p>Skin conditions fluctuate daily, and the subtle changes are difficult to detect. Transepidermal water loss (TEWL), a key indicator of skin barrier function, is effective for monitoring these changes; however, its measurement requires specialized and costly equipment, making daily implementation impractical. This study aimed to develop and evaluate a machine-learning model for estimating TEWL from facial images captured using a standard smartphone. Facial images were acquired from 74 participants using smartphones, and TEWL was measured from 15 distinct facial regions using a dedicated device. Image patches corresponding to these measurement sites were extracted to serve as training data. We developed a method combining a ResNet-50-based self-supervised feature extractor with a mixture density network. In a per-participant cross-validation, the model achieved a correlation coefficient of r &gt; 0.5 for 27 (38%) and r &gt; 0.3 for 48 (68%) participants. Our model significantly outperformed conventional methods, including linear, Ridge, and Lasso regression. These findings demonstrate that the proposed method is a highly effective and accessible approach for estimating TEWL from smartphone images.</jats:p>