Microfluidic SPRi profiling of plasma extracellular vesicles enables eight-marker phenotyping for machine learning-assisted multi-cancer discrimination

Accurate discrimination of clinically related cancer types using circulating extracellular vesicles (EVs) remains challenging because plasma EVs are heterogeneous, low-abundance, and difficult to profile in a multiplexed and quantitative manner. Here, we developed a microfluidic surface plasmon resonance imaging (SPRi) platform for eight-marker EV protein phenotyping and machine learning-assisted multi-cancer discrimination. Heterogeneous EV populations were broadly captured through cholesterol–phospholipid membrane interactions, followed by protein-specific signal amplification using aptamer-functionalized gold nanoparticles targeting CD63, EGFR, EpCAM, HER2, ICAM-1, ITGB1, MUC1, and PD-L1. The platform achieved sensitive EV detection over 10³–10⁸ particles mL⁻¹, with a limit of detection below 10³ particles mL⁻¹, and showed favorable analytical reliability in specificity and spike-and-recovery assays. In 94 clinical plasma samples from healthy donors and patients with gastric cancer, colorectal cancer, or breast cancer, single-marker and averaged-response analyses revealed clear cancer-associated EV alterations but showed limited ability to resolve cancer type-specific differences. By integrating the eight-dimensional EV protein profiles, multiple machine learning models improved multi-cancer discrimination, with all six models achieving multiclass AUC values of 1.000 and five-fold cross-validation accuracies around 0.95 in the present cohort. This study demonstrates that microfluidic SPRi-based EV phenotyping can transform plasma EV surface protein signatures into model-recognizable discriminative patterns, providing a feasible liquid biopsy strategy for non-invasive multi-cancer discrimination.