Successful control of atopic dermatitis (AD) symptoms is challenging due to considerable variation in responses to treatments between patients. Personalised treatment strategies may be more beneficial to individual patients rather than a “one-size-fits-all” approach to therapy. Choice of appropriate treatments for each patient is facilitated by better prognoses of AD severity dynamics. However, it is difficult to reliably identify and predict treatment effects at an individual level, due to the fluctuating dynamics of AD symptoms.
We applied a machine learning approach to analyse and learn dynamic patterns of AD severity scores for each patient and developed a Bayesian machine learning model for personalised prediction of daily AD severity scores. While most machine learning models lack interpretability and can be unsuitable for clinical use, our model adopted the structure of a previously published mechanistic model of AD pathogenesis, making it biologically interpretable. The model was developed using the already published longitudinal data of the daily recordings of AD severity scores and treatment used by 59 AD children over 6 months, and externally validated using a similar dataset of 334 AD children followed for 16 weeks.
We validated and confirmed the general applicability of the model. The model predictions of the daily evolution AD severity score were 60% more accurate than chance-level forecasts. The model was able to capture diverse patterns of severity trajectories while dealing with partially missing data. We also derived patient-dependent parameters that describe the short-term persistence of AD symptoms and dose-independent responsiveness to topical steroids, calcineurin inhibitors and step-up treatments.
The method proposed in this study is applicable to design complex yet interpretable machine learning models for clinical use. By predicting whether a chosen treatment is consistently effective and whether the disease is adequately controlled, for each patient, the method could help suggest personalised treatment strategies.

A Bayesian machine learning model to identify patient-specific dynamic responses to eczema treatments