Atopic dermatitis (AD) is the most common chronic inflammatory skin disease seriously affecting the quality of life of patients. Reliable and accurate evaluation methods are necessary for early diagnosis and effective AD treatment. Thus, this study used machine learning (ML) to explore a novel diagnostic and therapeutic effect evaluation model for AD. Firstly, candidate model genes were screened from an integrated data set of four AD-related microarray datasets: GSE133477, GSE32924, GSE58558, and GSE107361, using Robust Rank Aggregation (RRA), and protein-protein interaction network (PPI). Next, three recognized models (REC) and three AD-associated gene models (AAG) established with LASSO, Logistic linear regression (LR), and random forest (RF) were developed and tested separately using GSE130588 and GSE99802 datasets. The results revealed that REC model of LASSO (model genes including IL7R, KRT16, CCL2, CD53, CCL18 and CCL22), REC model of LR(including IL7R, KRT16, CCL18) and AAG model of LR (including MX1, CCNB1, SERPINB13, ADAM19, CEP55, VMP1, TTC39A, and FCHSD1) accurately classified AD lesions and non-lesions based on the good AUCs (LASSO (REC):0.8761, and LR (REC and AAG):0.8302 in GSE130588; LASSO (REC): 0.7761, and LR (REC and AAG):0.8783 in GSE99802). In Dupilumab, Crisaborole, and fezakinumab-treated samples, the LASSO (REC) and LR(AAG) models were positively correlated with SCORD (Pearson correlation coefficients of 0.55 and 0.69, respectively) and negatively correlated with the treatment length. In addition, the two models also accurately predicted the infiltration of immune cells in the skin lesions and non-lesions. Therefore, the ML-based predictive model provides a new approach to predicting AD diagnosis and the therapeutic effect of AD treatment options.