Exosomes emerges as indicators of the tumor microenvironment, yet their predictive utility for immunotherapy responses is limited by the insufficient sensitivity and specificity of currently available assays. Here, a multi-level engineering strategy is presented that enables accurate exosome-based prediction of immunotherapy responses by integrating systematic aptamer ligand tailoring, ultrasensitive ellipsometry-based sensing, and clinically relevant tumor-proximal fluid sampling. Aptamers specifically targeting PD-L1 are identified through systematic evolution of ligands by exponential enrichment (SELEX), followed by truncation and computational sequence refinement to enhance binding specificity. The optimized aptamer sequence (Tr-Apt13) is validated from molecular interaction analyses to in vitro assays, demonstrating superior target binding efficacy over conventional antibodies, attributed to dense surface immobilization and multivalent binding capability. When incorporated into an ellipsometry-based dual-prism solution-immersed silicon sensor, Tr-Apt13 enabled ultrasensitive detection of PD-L1-expressing exosomes with a detection limit of ≈9.8 particles mL^-1, exhibiting high linearity and stability. Clinical validation using bronchoalveolar lavage fluid (BALF) from lung cancer patients revealed precise discrimination between responders and non-responders at exosome concentrations as low as 10³ particles mL^-1, outperforming serum-based analysis, conventional ELISA, and tissue immunohistochemistry. Collectively, this multi-level strategy offers a new perspective on exosome-based diagnostics, enabling precise prediction of immunotherapy efficacy in lung cancer patients.