Dermoscopy of intraepidermal keratinocytic neoplasms in solid organ transplant recipients: a single-center prospective real-world study supported by machine learning.

BACKGROUND

Actinic keratoses (AKs) and Bowen disease (BD) are precursors of cutaneous squamous cell carcinoma (cSCC) and are particularly relevant in solid organ transplant recipients (SOTRs) because of chronic immunosuppression. Early, accurate diagnosis is essential. Dermatoscopy is the most widely used and validated noninvasive diagnostic tool, although most studies have involved immunocompetent patients.

OBJECTIVES

To evaluate the diagnostic accuracy of dermatoscopy in differentiating AK from BD in SOTRs, assess established dermoscopic criteria in this population, and compare the performance of expert observers, logistic regression, and machine learning models.

METHODS

We conducted a prospective observational study of 38 lesions clinically classified as AK or BD in SOTRs. Each lesion was assessed dermatoscopically by 3 blinded experts using 13 predefined criteria and subsequently biopsied for histopathological confirmation. Analyses included logistic regression, interobserver agreement, diagnostic performance, and evaluation of random forest models using SHapley Additive exPlanations (SHAP) values.

RESULTS

We evaluated 38 lesions in 36 patients; 22 were histopathologically diagnosed as BD (58%) and 16 as AK (42%). Dermoscopic sensitivity and specificity for BD were 61% and 79%, respectively (positive predictive value, 81%; negative predictive value, 58%). For AK, sensitivity was 79% and specificity was 61%. Logistic regression identified the strawberry pattern, thick white-yellow scales, and glomerular vessels as significantly associated with BD. SHAP analysis supported these associations but also identified differences between human and model-based feature interpretation.

CONCLUSIONS

Established dermoscopic criteria may underestimate lesion severity in SOTRs, particularly for BD. Some features traditionally associated with AK, including the strawberry pattern and rosettes, were more indicative of BD in this cohort. Diagnostic algorithms require further validation in immunosuppressed populations, and explainable machine learning may provide complementary diagnostic insights.