The analysis of complex Raman spectra from biological samples has traditionally relied on conventional chemometric-based methods, the performance of which has been further improved by artificial intelligence (AI). The accurate identification of bacteria is critical to preventing sepsis, even in advanced clinical settings. Among several methods, Raman scattering has shown great promise in overcoming the limitations of conventional approaches. Despite this promise, unresolved challenges remain in the optimization of SERS and interpretation of AI algorithms. In this study, we used colloidal Au and Ag nanoparticles (NPs) to obtain reproducible surface-enhanced Raman scattering (SERS) spectra of bacteria. We investigated the effects of ligands on plasmonic NPs and wavelength dependence in SERS-based bacterial identification. The analysis was conducted within the biological fingerprint region of 500^–1300 cm^–1. Using a deep neural network model, the targeted SERS approach with mannose-modified AuNPs under 532 nm excitation achieved a high classification accuracy (96.1%) for 14 bacterial species. In addition, we propose a framework designed to explain how AI algorithms accurately interpret Raman spectra. The normalized positive values of Shapley additive explanations (npSHAP) were utilized to identify the top five peaks as an AI-selected spectral barcode. The spectral signatures identified by the proposed framework are presented in a manner that is both intuitive and straightforward, facilitating a clear and precise interpretation of the bacterial classification process.