Timely and accurate primary tumor diagnosis is critical, and misdiagnoses and delays may cause undue health and economic burden. To predict primary tumor types based on genomics data from a de-identified US nationwide clinico-genomic database (CGDB), the XGBoost-based Clinico-Genomic Machine Learning Model (XC-GeM) was developed to predict 13 primary tumor types based on data from 12,060 patients in the CGDB, derived from routine clinical comprehensive genomic profiling (CGP) testing and chart-confirmed electronic health records (EHRs). The SHapley Additive exPlanations method was used to interpret model predictions. XC-GeM reached an outstanding area under the curve (AUC) of 0.965 and Matthew's correlation coefficient (MCC) of 0.742 in the holdout validation dataset. In the independent validation cohort of 955 patients, XC-GeM reached 0.954 AUC and 0.733 MCC and made correct predictions in 77% of non-small cell lung cancer (NSCLC), 86% of colorectal cancer, and 84% of breast cancer patients. Top predictors for the overall model (e.g. tumor mutational burden (TMB), gender, and KRAS alteration), and for specific tumor types (e.g., TMB and EGFR alteration for NSCLC) were supported by published studies. XC-GeM also achieved an excellent AUC of 0.880 and positive MCC of 0.540 in 507 patients with missing primary diagnosis. XC-GeM is the first algorithm to predict primary tumor type using US nationwide data from routine CGP testing and chart-confirmed EHRs, showing promising performance. It may enhance the accuracy and efficiency of cancer diagnoses, enabling more timely treatment choices and potentially leading to better outcomes.
Keywords: Cancer diagnosis; Clinico-genomic database; Machine learning; Predictive model; Primary tumor type; Real-world data.
© 2023 The Authors.