Surface enhanced Raman scattering (SERS) has great potential in the early diagnosis of diseases by detecting the changes of volatile biomarkers in exhaled breath, because of its high sensitivity, rich chemical molecular fingerprint information, and immunity to humidity. However, two factors limit the application of SERS: 1) specific enrichment of trace target molecules in SERS hotspots; and 2) stability and reproducibility of SERS signals in multiinterference environments. In order to accurately detect biomarkers in the complex exhaled breath and to eliminate the interference of other components, hollow ZIF-8 wrapped on the gold superparticle with a yolk-shell structure is proposed as a SERS substrate. Similar to the solid ZIF layer, the hollow ZIF-8 layer is also enriched with gas molecules, and the enriched molecules reacts with functional molecules on the surface of the superparticle, generating a strong response signal. The difference is that the hollow ZIF layer can effectively exclude interfering molecules that are not bound to the modified molecules, and the detection limit is 5 times lower than the detection limit of core-shell structure substances. A mask-type sensor is prepared, and the obtained spectra are modeled by PC-LDA to determine the probability of illness in the actual population.