Palmar hyperlinearity has been identified in 68%-85% of individuals with heterozygous filaggrin gene (FLG) mutations and up to 100% of individuals with homozygous FLG mutations. However, the definition of palmar hyperlinearity used in studies has varied. We aimed to standardize prediction of FLG status from palmar images using machine learning. Palmar images and FLG genotyping were available for 342 individuals with atopic dermatitis from the Tower Hamlets Eczema Assessment study. There were 97 individuals (28%) with FLG mutations (10 homozygous or compound heterozygous and 87 heterozygous) and 245 wild type individuals (72%). Images were cropped to create palmar and thenar eminence datasets. For feature extraction, separate models were created for histogram of oriented gradients (HOG) and Haralick features. Classification with k-nearest neighbour, linear support vector machine (SVM), and radial-basis function (rbf) SVM was performed and results reported for the best classifier using ten-fold cross-validation with a 90:10 split. For the thenar eminence dataset, mean accuracy was 75.9% (standard deviation (SD) 8.9) for the Haralick model and 75.6% (SD 6.5) for the HOG model, both with linear SVM classification. For the palmar dataset, mean accuracy was 77.6% (SD 7.0) for the Haralick model and 72.0% (SD 7.0) for the HOG model, using linear SVM and SVM-rbf, respectively. When the 10 homozygous/compound heterozygous were held out from training, all images were correctly predicted as having a FLG mutation with 3-nearest neighbour classification. FLG status can be predicted with 72-78% accuracy from palmar images. Advantages of our approach over clinical prediction are speed, with hundreds of photos rapidly scored without needing the time of clinical experts, and avoidance of interobserver variation.

ESDR054 - Machine learning prediction of filaggrin mutation status from palmar images: a proof-of-concept study

50th Annual Meeting of the European Society for Dermatological Research. VIRTUAL MEETING