An Extracellular Vesicle Protein‐Based Machine Learning Framework for Early Detection of Oesophageal Squamous Cell Carcinoma: A Multicentre, Prospective Study

作者：Yu Wang, Shan Chao Xing, Yue Wu, Ning Xue, Pei-Min Chen, Run-Xian Jin, Yi Xu, Ming-Fang Ji, Yu-Hui Peng, Yuan-Tao Liu, Lina Chen, Meng Wu, Zi-Ying Jiang, Shang-Hang Xie, Yi-Ling Luo, Biao Zhang, Xin-Yuan Ou, Huang Qp, Boyu Tian, Li Ling, Cao Sh, Wanli Liu, Mu‐Sheng Zeng, Qian Zhong · 发表于：Journal of Extracellular Vesicles · 年份：2026 · DOI：10.1002/jev2.70246 · 被引用次数：2 · 研究领域：Extracellular vesicles in disease、Esophageal Cancer Research and Treatment、Ferroptosis and cancer prognosis

Early detection of oesophageal squamous cell carcinoma (ESCC) is critical for improving survival, yet current screening is hampered by the lack of effective, non-invasive methods. Here, we developed and prospectively validated an extracellular vesicle (EV) protein-based blood test for the preclinical detection of ESCC. We first engineered BarFlare, a high-sensitivity platform for serum EV protein analysis, and identified a novel biomarker panel that includes EV-associated squamous cell carcinoma antigen (SCC) and matrix metalloproteinase-13 (MMP13). These biomarkers were integrated with clinical factors into an interpretable multi-criteria decision-making classification fusion (MCF) machine-learning framework. The MCF model was trained and validated in prospective, multicentre diagnostic cohorts (n = 1018), and its preclinical detection capability was assessed in a prospective, population-based longitudinal cohort. The MCF framework accurately distinguished patients with ESCC from healthy controls in a test set (AUC, 0.987) and two external validation cohorts (AUCs, 0.926 and 0.960), including those with early-stage disease (AUCs, 0.901-0.980). Critically, in the longitudinal cohort, the framework identified individuals who would later develop ESCC from their baseline blood samples with a median lead time of 34.9 months (range, 0.4-72.5) before clinical diagnosis (AUC, 0.864; sensitivity, 73.3%; specificity, 82.2%). The risk score of the model correlated with time to diagnosi...