We previously predicted filaggrin gene (FLG) loss-of-function mutation status from palmar images using textural features (histogram of oriented gradients (HOG) and Haralick). In this study, we used a convolutional neural network (CNN) to predict FLG status using a larger cohort of images. Images and FLG status were available from the Tower Hamlets Eczema Assessment study for individuals of Bangladeshi origin aged ≤30 years with atopic dermatitis. We utilised images cropped to two regions of interest (thenar eminence and palm). One image was used per patient and an 80:10:10 split used for training, validation, and testing. We used a pre-trained CNN (EfficientNetB0). The classification task was binary (FLG loss-of-function mutation present vs absent). Model training was performed over 30 epochs. The keras library in Python 3.9.0 was used. We included 531 images in each dataset. For thenar images, area under curve (AUC) was 87.8%, accuracy was 84.9%, and the positive predictive value (PPV) was 87.5%. For palmar images, AUC was 68.4%, accuracy was 66.0%, and the PPV was 65.9%. CNN accuracy was higher for thenar images compared to palmar images. This likely reflects a higher signal-to-noise ratio in thenar images, with previous studies having identified hyperlinearity patterns more strongly associated with FLG loss-of-function mutations at the thenar eminence, including a cross-hatch and diamond pattern. CNN accuracy for thenar images was also higher compared to previous accuracy achieved using textural feature extraction and classification (84.9% vs 73.3%). Limitations of our study were dataset size and lack of external validation. However, to minimise the risk of over-fitting we used a relatively simple CNN and performed textural feature extraction previously as a baseline for machine learning performance. To facilitate external validation we have made our code available.

ESDR245 - Deep learning prediction of filaggrin mutation status from palmar images

The scientific meeting of the European Society for Dermatological Research (ESDR) is held yearly, usually in September. With an international attendance of more than 1000 skin scientists from around the globe, the four-day event includes short oral presentations from peer-reviewed abstracts, keynote lectures from international experts, sponsored symposia, satellite programs, educational events as well as an important focus on posters, networking and building collaborations amongst scientists. For more information about this year’s meeting visit our website.