Atopic dermatitis (AD) is an inflammatory skin condition that relies largely on subjective evaluation of clinical signs and symptoms for diagnosis and severity assessment. Using multivariate data, we attempted to construct prediction models that can diagnose the disease and assess its severity. We combined data from 28 mild-moderate AD patients and 20 healthy controls (HC) to create random forest models for classification (AD vs. HC) and regression analysis to predict symptom severities. The classification model outperformed the random permutation model significantly (area under the curve: 0.85 ± 0.10 vs. 0.50 ± 0.15; balanced accuracy: 0.81 ± 0.15 vs. 0.50 ± 0.15). Correlation analysis revealed a significant positive correlation between measured and predicted total SCORing Atopic Dermatitis score (SCORAD; r = 0.43), objective SCORAD (r = 0.53), eczema area and severity index scores (r = 0.58, each p < 0.001), but not between measured and predicted itch ratings (r = 0.21, p = 0.18). We developed and tested multivariate prediction models and identified important features using a variety of serum biomarkers, implying that discovering the deep-branching relationships between clinical measurements and serum measurements in mild-moderate AD patients may be possible using a multivariate machine learning method. We also suggest future methods for utilizing machine learning algorithms to enhance drug target selection, diagnosis, prognosis, and customized treatment in AD.
Read full abstract