利用血浆细胞外囊泡（EV）拉曼光谱的AI驱动液体活检平台的临床可行性研究

Background: While liquid biopsy holds significant potential for cancer detection, its accuracy in the early stages of cancer remains a critical challenge. Here, we present a novel AI and nanotechnology-driven liquid biopsy platform for analyzing Raman signal patterns of plasma extracellular vesicles (EV). Beyond conventional bioassays and omics methods, our method identifies comprehensive differences in the molecular fingerprints of plasma EV. In a multi-center and multi-ethnic feasibility study involving both retrospective and prospective cohorts, we demonstrate that this approach enables high-sensitivity multi-cancer early detection (MCED) across five cancer types, particularly in their early stages. Furthermore, we outline a two-phase study design and strategy for both algorithm development and rigorous validation. Methods: This method consists of three major stages: EV isolation, signal detection, and AI interpretation. To ensure uniform performance and reproducibility within clinical laboratories, automated EV extraction and detection system was developed and utilized. We enrolled 447 participants, including 128 healthy controls and 319 patients across five different cancer types. The collected samples were partitioned using temporal cutoffs specific to each retrospective and prospective period; earlier samples were used for development, while subsequent samples served as the validation set. All sample testing was conducted by an independent third-party central lab. Results: Our system achieved an AUROC of 86.6%, a sensitivity of 56.4% at a specificity of 98.0%. The robust performance was confirmed across multiple clinical sites. Notably, the performance at early stages (stage I and II) yielded the AUROC for individual cancer types were 84.7% (lung), 91.8% (breast), 79.7% (colorectal), 91.5% (pancreas), and 83.3% (ovary), underscoring its feasibility as a robust prescreening tool for early-stage cancers. Furthermore, we integrated a secondary algorithm to identify the suspected tissue of origin (TOO) for cancer-positive samples, which can guide clinical follow-up and targeted diagnostic procedures. Our method demonstrates that liquid biopsy for prescreening early-stage cancer can be performed with high accuracy, with minimal turnaround time and hands-on intervention. Conclusions: In conclusion, these results validate the clinical utility of our liquid biopsy system as a high-sensitivity tool for early-stage cancer screening across multiple cancers. By precisely recognizing cancer-related EV changes using AI and nanotechnology, this platform facilitates rapid detection at the initial stages of cancer progression, offering a powerful advantage for improving clinical outcomes.