Large-scale clinical validation of a blood-based, multi-cancer, early detection test across different sample types, platforms, and populations.

Abstract

10521 Background: Many established cancer screening or diagnostic methods often face challenges in low- and middle-income countries due to high cost, complexity, and reliance on extensive medical infrastructure. OncoSeek, a multi-cancer early detection (MCED) test developed with a panel of protein tumor markers (PTMs), is both affordable (reagent cost ~$20) and accessible, requiring only a blood draw. We evaluated its performance of multi-cancer detection and diagnosis in large-scale clinical studies. Methods: 15,122 participants (3,029 cancer vs 12,093 non-cancer) were divided into one training and six validation cohorts according to the different sites in three countries (Brazil, China, and United States). One tube of blood (plasma or serum) from each participant was collected and quantified using a panel of 7 PTMs (AFP, CA125, CA15-3, CA19-9, CA72-4, CEA, and CYFRA 21-1) through four common immunoassay platforms (Roche, Abbott, Luminex, and ELISA) in both retrospective and prospective settings. OncoSeek, utilizing artificial intelligence (AI) algorithm, was developed to differentiate cancer cases from non-cancer cases based on 7 PTM concentrations and clinical information including age and sex. It also predicted the potential affected tissue of origin (TOO). Furthermore, the fifth validation cohort, comprising 1849 patients (1031 cancer vs 818 non-cancer), was leveraged to broaden OncoSeek's application for cancer diagnosis. This cohort specifically targeted symptomatic patients, requiring further confirmation through biopsy or surgery. Results: The conventional clinical method, using a single threshold for each PTM, lead to accumulate the false positive rate with the growing number of PTMs. However, OncoSeek, empowered by AI, significantly reduced the false positive rate, elevating specificity from 54.3% to 93.0% and achieving an overall sensitivity of 51.7% in the training cohort. Performance remained robust (58.4% sensitivity and 92.0% specificity) across all seven cohorts, with area under the curve (AUC) values ranging from 0.744 to 0.912. The overall accuracy of TOO prediction was 65.4%. In the fifth cohort with symptomatic patients, OncoSeek achieved a 0.845 AUC for cancer diagnosis at 73.1% sensitivity and 90.6% specificity. Conclusions: OncoSeek significantly outperforms the conventional clinical method, showcasing its robust performance across various races, sample types, and platforms. The extensive retrospective assessment of OncoSeek in a symptomatic population demonstrates the feasibility of this MECD test in aiding clinicians for decision-making. Its accuracy of TOO facilitates the diagnostic workup.