Deep learning reveals prognostic information beyond AJCC stage in routine histopathology of early-stage melanoma: the population-based D-ESMEL study

A deep learning model using routine whole-slide histopathology from early-stage melanoma predicts distant metastatic risk with added prognostic value beyond AJCC stage, achieving a best-fold balanced accuracy of 0.88.

• Why it matters: Many patients who later develop distant metastasis are first diagnosed at stage I/II, so better prognostic tools are needed to identify who may benefit from (neo)adjuvant therapy.
• What they did: Using the population-based D-ESMEL study, the workflow trained on a discovery set of n=328 with 10-fold cross-validation and was validated in a nested case-control cohort of >5,500 patients (n=266 stage-matched controls).
• The result: In validation, deep learning metastatic risk was an independent prognostic factor (adjusted OR: 1.97; p<0.01), and combining predictions with AJCC stage increased C-index to 0.87 versus 0.80 for AJCC stage alone (p<0.001).

AI summary of the abstract · may contain errors · not medical advice · read the paper.