Development of a SERS-based diagnostic tool for infectious vaginitis via intelligent analysis of vaginal fluid spectra.

BACKGROUND

Vaginal infections, including bacterial vaginosis (BV), vulvovaginal candidiasis (VVC), and trichomoniasis (TV), are common gynecologic disorders in reproductive age women. Accurate differential diagnosis remains challenging due to overlapping symptoms and limitations of conventional microscopy and biochemical assays. Surface-enhanced Raman spectroscopy (SERS) provides label-free molecular fingerprints of complex biological fluids and may support automated classification when combined with machine learning. The problem addressed here is the lack of a label-free analytical strategy for patient-level discrimination of healthy controls (HC), BV, VVC, and TV using vaginal fluid samples.

RESULTS

SERS spectra were acquired from vaginal fluid samples using citrate-reduced silver nanoparticles (AgNPs). A total of 170 participants were enrolled, with one vaginal secretion sample collected from each participant. Among these samples, 135 samples (HC = 40, BV = 40, VVC = 40, TV = 15) were used for model construction and 35 independent samples (HC = 10, BV = 10, VVC = 10, TV = 5) were used for independent blind validation. Although the average SERS spectra of HC, BV, and VVC showed substantial overlap, spectral deconvolution revealed subtle class-related features. Thirteen machine-learning and deep-learning models were evaluated using a patient-wise strategy. The one-dimensional convolutional neural network (1DCNN) achieved the most balanced performance, with an internal hold-out patient-level accuracy of 80.0% and a macro-AUC of 0.92. Independent blind validation correctly classified 31 of 35 participants, corresponding to an external patient-level accuracy of 88.6%.

SIGNIFICANCE: This study establishes a patient-wise SERS and machine-learning framework for four-class discrimination of HC, BV, VVC, and TV. Its novelty lies in combining vaginal-secretion SERS fingerprints, patient-level data splitting, majority-vote classification, and independent blind validation. This label-free and automated framework may provide an objective complement to conventional microscopic and biochemical testing in future diagnostic workflows.