Diagnostic Accuracy of Histopathologic Features Suggestive of Dermatophytosis: A Retrospective Cohort Selection Study...

Histopathologic features have been proposed as clues to potential dermatophyte infection. The inter-observer reproducibility and diagnostic accuracy of these features has not been rigorously studied.

Four blinded assessors at different levels of experience and training evaluated a cohort of 97 hematoxylin-eosin slides for 12 clues which might indicate dermatophytosis. Interobserver concordance, diagnostic accuracy metrics, and confidence intervals were calculated. Machine learning was used to develop predictive models. Model performance was compared with five-fold cross-validation.

Interobserver agreement between two board-certified pathologists was moderate to substantial for most clues (nine predictors with kappa values of 0.40-0.64). In contrast, agreement was only slight for the sandwich sign and compact red corneum (kappa ≤ 0.05). Among all features, the presence of structures suspicious for dermatophytes was the only significant predictor of PAS positivity on multiple logistic regression (balanced accuracy = 0.88, p < 0.001). In machine learning models with 5-fold cross-validation, logistic regression using only the presence of suspected fungal hyphae as a predictor outperformed more complex approaches, including random forest, k-nearest neighbors, latent factor analysis, and support vector machines. Using penalized maximum likelihood, the estimated probability of a positive PAS result was 0.93 when possible hyphae were identified and 0.03 when they were not.

Possible fungal hyphae in the stratum corneum on H&E-stained sections is the most reliable histopathologic clue to dermatophytosis.