From Diagnostics to Survival Prediction: A Systematic Review of Artificial Intelligence Applications in Merkel Cell Carcinoma
Jonathan Mokhtar, Noora Jatan, Éli Blais, Victoria Zecchin Ferrara, Moshe Assouline, Olivier Mathieu, Alexandre G. LellouchBackground
Merkel cell carcinoma (MCC) is a rare, aggressive neuroendocrine skin cancer with substantial mortality and limited prognostic granularity under current staging systems. Artificial intelligence (AI) has shown promise across oncology, but no synthesis has yet evaluated its application specifically to MCC across the clinical continuum, from detection to survival prediction.
Methods
This systematic review followed PRISMA and Cochrane guidelines and was prospectively registered with PROSPERO. Four databases (PubMed, Embase, Scopus, and CENTRAL) were searched from inception through June 18, 2026. Eligible studies evaluated AI, machine learning, or computational models using human MCC data. Given anticipated heterogeneity in predictive targets and metrics, a narrative synthesis was performed.
Results
Of 950 records identified, seven studies published between 2021 and 2026 met inclusion criteria. Four addressed prognosis or survival, two addressed diagnosis or early detection, and one addressed treatment‐receipt prediction. The best‐performing survival model achieved an AUROC of 0.89 (internal) and 0.81 (external validation), with a concordance index of 0.93. Other analyses suggested improved model fit or additional prognostic information beyond conventional AJCC staging, although direct comparisons of discrimination were not consistently available. A smartphone‐based diagnostic tool achieved an AUC of 0.80 for general malignant‐lesion triage, although sensitivity fell from 82.5% to 67.6% under a real‐world, STARD‐compliant worst‐case analysis; importantly, the cohort included only one MCC lesion, which both the CNN and teledermatology assessment classified as low risk, precluding meaningful MCC‐specific performance estimation. A blood‐based repertoire‐sequencing model reported near‐perfect discrimination but lacked independent MCC validation. Overall risk of bias was high, with the “analysis” domain most frequently downgraded; no study integrating multimodal data was identified.
Conclusions
AI applications in MCC remain early stage but demonstrate meaningful potential to refine prognostic stratification beyond conventional staging, although MCC‐specific diagnostic evidence remains severely limited. Broader external validation and multimodal data integration represent the field’s most pressing needs before clinical translation is warranted.