Abstract
A subset of cancers presents with unclear or potentially incorrect primary histopathologic diagnoses, including cancers of unknown primary (CUP). We aimed to develop and validate an artificial intelligence (AI) tool, GPSai, that predicts tumor tissue of origin in CUP and flags potential misdiagnoses for additional workup during routine molecular testing. The GPSai model was trained on whole exome and whole transcriptome data from 201,612 cases submitted for tumor profiling at Caris Life Sciences. Retrospective (N=21,549) and prospective (N=76,271) validations were performed. The clinical impact was evaluated over eight months of live testing and through physician surveys. GPSai demonstrated 95.0% accuracy in non-CUP cases and reported on tumor tissue of origin in 84.0% of CUP and 96.3% of non-CUP cases. During the initial eight months of implementation, GPSai changed the diagnosis on 704 patients (0.88% of all profiled cases), which were supported by orthogonal evidence including imaging, immunohistochemistry, mutational signatures, hallmark fusions, or viral reads. Diagnosis changes prompted changes in targeted-therapy eligibility based on Level 1 clinical evidence in 86.1% of cases (n=606/704). A majority (89.7%; n=87/97) of physician responses indicated acceptance of the GPSai results and 53.6% (n=52/97) of responses stated that the results prompted a change in treatment plan. GPSai accurately identifies tumor tissue of origin and has the potential for clinical impact in a small but meaningful subset of patients with CUP or pathologically ambiguous tumors. Our results support the integration of this AI tool into routine molecular testing to improve diagnostic accuracy and guide subsequent therapeutic decisions.