Machine learning analysis of pretreatment skin biopsies predicts nonresponse to dupilumab in patients with eczematous dermatitis

Abstract

While dupilumab has revolutionized the treatment of atopic dermatitis (AD), a subset of patients may fail to respond or worsen after dupilumab initiation. Using a retrospective cohort of 53 dupilumab responders and 17 nonresponders, we developed a logistic regression classifier to predict nonresponse using 7 cytokine staining and histological features derived from pretreatment biopsies. Our model demonstrated an accuracy of 95.7%, a sensitivity of 88.2%, a specificity of 98.1% and a PPV of 93.8% for predicting nonresponse using leave-one-out cross-validation, underscoring treatment-relevant immunological heterogeneity in eczema and demonstrating the potential of using machine learning and tissue biomarkers to predict dupilumab nonresponse.