Abstract: Cerebrospinal fluid (CSF) rhinorrhea often presents as clear nasal discharge, making it challenging to differentiate from normal secretions and delaying diagnosis. As CSF leakage provides a direct pathway for pathogen entry into the central nervous system, rapid and accurate detection is essential to prevent severe infections such as meningitis. This study introduces a machine learning (ML)-assisted surface-enhanced Raman scattering (SERS) diagnostic platform that reliably distinguishes CSF from nasal secretion samples. The core sensing element is an Au@Ag bimetallic nanopillar substrate, engineered to exploit synergistic plasmonic effects between gold and silver for maximal SERS enhancement while offering superior corrosion resistance. This high-performance substrate enables sensitive and reproducible detection of clinical specimens. To address spectral resolution and range inconsistencies among different Raman instruments, a cross-instrument spectral preprocessing algorithm was developed to standardize input spectra. Among the ML pipelines evaluated, the NearMiss-2 (NM2)-logistic regression (LR) model demonstrated the highest classification performance in both internal and external validations. Notably, when applied to spectra from a portable Raman spectrometer, the NM2-LR pipeline achieved a 0.95 true positive rate and a 1.00 true negative rate. This Au@Ag nanopillar-based ML-SERS platform provides a rapid, cost-effective, and portable solution for CSF rhinorrhea diagnosis, with significant potential for broader biomedical applications.

Key words: Cerebrospinal fluid rhinorrhea, Surface-enhanced raman scattering, Machine learning, Au@Ag bimetallic, Spectral preprocessing