Computational tools for data-driven personalised medicine for Atopic Dermatitis
Abstract
Atopic dermatitis (AD) is a complex disease for which personalised medicine is of high relevance given the considerable variation in the clinical phenotypes and responses to treatments among patients. We have recently developed computational tools that facilitate a data-driven approach towards patient-centered care for AD by an “assess”, “predict” and “act” strategy. The first tool, EczemaNet, is a computer vision pipeline using deep learning to assess AD severity from photographic images. EczemaNet could accurately predict the intensity of seven disease signs for each image, after training on 1393 images from 310 AD children. Automatic evaluation of AD severity by EczemaNet could help patients to monitor their AD conditions more easily at home without attending clinics. The second tool, EczemaPred, is a computational package to develop statistical machine learning predictive models for any AD severity scores. EczemaPred uses longitudinal data to train Bayesian state-space models that predict each severity item constituting severity scores, and can deal with missing or irregular measurements, or data from a small population. We developed personalised predictive models for two patient-oriented AD severity scores, daily PO-SCORAD and weekly POEM, using EczemaPred and three different longitudinal datasets from more than 600 patients in total. Predicting temporal evolution of AD severity scores at an individual level, under different treatment regimens, could inform the design of personalised treatment strategies that the patients can act on. Our analysis found little evidence that additional factors (such as biomarkers and environmental factors) improve the prediction performance significantly.
