Real-world outcomes of palliative anti–programmed cell death protein 1 monotherapy in patients aged ≥75 years with advanced melanoma

The CNN achieved an AUC of 0.938 (95%CI 0.925-0.952)for differentiating malignant from benign lesions, and 0.891 (95%CI 0.862-0.920)for distinguishing melanoma from nevi.At the lowest predefined threshold, sensitivity was 94.9% for malignant lesions and 87.2% for melanoma, with specificities of 71.4% and 73.1%, respectively.Basal cell carcinomas showed the highest mean DEXI score (7.99 [SD 2.09]) and top-class classification accuracy (83.9%).Melanomas scored significantly higher than nevi (5.12 vs 2.27; p <0.001) but melanoma was assigned as the top predicted class in only 47.9% of cases.Receiver Operating Characteristic curve for discrimination of malignant from benign lesions.Conclusions: This large-scale external validation study sets a benchmark for how AI-based tools should be evaluated prior to widespread clinical integration.While the CNN demonstrated meaningful diagnostic accuracy in a real-world, high-risk cohort, relevant limitations were identified, supporting the cautious integration of such algorithms as adjunctive decision support rather than as standalone tools for melanoma detection.