Novel Tool to Predict Primary Tumor Location in the Setting of Metastatic Cancer With Unknown Primary.

PURPOSE: This study aimed to develop and validate a tool to predict the primary tumor site in patients with metastatic cancer of unknown primary using demographic and clinical variables, including imaging-derived sites of metastasis, without pathology or laboratory data. METHODS: In this population-based study, data were extracted from the SEER database. Patients diagnosed between 2010 and 2020 with a known primary cancer site, metastatic staging, and metastasis in the bone, liver, brain, and lungs were included. Patients were randomly split into training (70%) and/or testing (30%) sets. A total of 27 primary tumor sites were included. A gradient boosting model was trained to predict the primary tumor site, with random forest and classification tree models evaluated as comparators. Random guess and frequency-weighted guessing models were used for baseline comparison. RESULTS: A total of 689,715 patients were included from 9,208,295 queried. The most common primary cancer sites were non-small cell lung cancer (NSCLC) (n = 211,338), intestines (n = 64,187), and pancreas (n = 60,310). The baseline models achieved Top-5 accuracies of 0.179 for a random guess and 0.476 for a frequency-weighted guess. The gradient boosting model achieved Top-1, Top-3, and Top-5 accuracies of 0.40 ± 0.001, 0.680 ± 0.001, and 0.790 ± 0.001, respectively, across 100 Monte-Carlo replicates. Top-5 accuracy was highest for primary NSCLC (0.990 ± 0.001) and intestinal cancer (0.900 ± 0.003). Our model was most influenced by sex (gain = 0.27) and metastasis to bone (gain = 0.20). A web app was constructed: primarycancerpredictor.netlify.app. CONCLUSION: This study developed and evaluated a gradient boosting machine learning model and a user-friendly web tool to aid in the diagnosis of cancer of unknown primary.