Cutaneous squamous cell carcinoma (cSCC) accounts for 20-50% of all skin cancers. While most cSCC patients have a good prognosis, some patients have an elevated risk of recurrence and metastasis such that cSCC mortality approaches that from melanoma. cSCC tumors located on the head and neck (H&N) make up approximately two-thirds of all cases. While most patients with H&N cSCC can be cured by conventional surgical excision or Mohs micrographic surgery, a small percentage of patients will experience regional or distant metastasis, both associated with poor outcomes. For H&N cSCC that displays high-risk clinical features, clinical decisions include the use of adjuvant therapy and/or sentinel lymph node biopsy with therapeutic neck dissection. Current staging systems are limited in their accuracy to identify truly high-risk patients to facilitate appropriate management. In this study, we sought to address this unmet clinical need by developing a gene expression signature to identify patients at high risk for recurrence and metastasis, with a subanalysis of patients with H&N lesions. Archival primary formalin-fixed paraffin-embedded cSCC tumor tissue, clinicopathologic information, and outcomes data were collected from 12 centers under an IRB-approved protocol. For test development, we selected 73 candidate genes based on literature and pathway analysis for a targeted qPCR approach and also performed global gene expression profiling by microarray on a subset of cases to maximize prognostic gene selection. Multiple machine learning methods were applied using gene expression data from a development cohort of 217 cases (130 H&N) with 25 recurrences (18 H&N). The targeted approach yielded 23 and 13 genes that were differentially expressed between recurrent and non-recurrent cases in the full cohort and H&N subset, respectively (p<0.05). Substantial overlap in genes altered in the full cohort and the H&N group was observed, particularly for regional or distant recurrences, indicating that signatures developed from the full cohort may also apply to the H&N subset. Preliminary modeling with cross-validation suggested that gene sets from the targeted approach could have improved accuracy metrics compared to clinicopathologic staging. Using the microarray data from 80 cSCC cases (52 H&N) with 29 recurrences (20 H&N), machine-learning identified a subset of 67 top-performing genes to predict metastasis. These genes will be assessed by qPCR and can be combined with gene sets from the targeted approach in a gene expression assay to be validated in an independent cohort. Overall, our preliminary results suggest that identification of a prognostic gene expression-based signature for cSCC is possible and, importantly, that these findings can be applied to patients with H&N cSCC to help guide appropriate management strategies based on accurate risk assessment.