Identification of model-ranked candidate medical interventions for pediatric atopic dermatitis using Dempster-Shafer evidence theory

Atopic dermatitis (AD) is common in infants and children, but comparative evidence for ranking drug combinations at the individual-patient level remains limited. Pathological and diagnostic data from 1,736 pediatric patients with AD were collected. After major measured confounding factors were screened, 455 patients with relatively low EASI scores were retained. CLIQUE subspace clustering was used to identify groups defined by similar medication patterns, and Dempster-Shafer evidence theory (DSET) was used to construct mass functions under prespecified clinical constraints. EASI improvement per hospitalization day was treated as an exploratory constructed surrogate outcome. Nested fivefold cross-validation was the sole primary validation strategy; all preprocessing, parameter selection, clustering, mass-function construction, safety-weight application, and prediction were performed within the corresponding training folds. Proportional calibration and expert-panel agreement were treated only as secondary descriptive analyses. In the primary nested fivefold cross-validation analysis, the clinically constrained framework generated model-ranked candidate combinations for held-out patients and yielded a mean hospitalization-duration difference of 1.57 days relative to experience-based selection. The highest-ranked candidates included topical fluticasone propionate with complementary adjuncts. A secondary proportional calibration analysis suggested an average difference of approximately 1.68 days across the three leading candidates. These values are internally validated associations or descriptive estimates, not causal treatment effects. The clinically constrained CLIQUE-DSET framework generated interpretable candidate rankings and showed a modest internally validated association with shorter hospitalization duration than experience-based selection. Because the endpoint was an exploratory surrogate and the study was retrospective, single-center, and focused predominantly on mild cases, the findings are hypothesis-generating and do not support direct prescribing or causal claims.