Surface-enhanced Raman spectroscopy of serum exosomes coupled with support vector machine for diagnosis of Parkinson's disease.

Parkinson's disease (PD) diagnosis faces substantial challenges due to the lack of reliable biomarkers and the limitations of existing detection techniques. Exosomes, which carry biomolecular cargo reflective of disease pathology, are increasingly recognized as promising biomarkers because of their stable presence in biofluids and accessibility through minimally invasive methods. Surface-enhanced Raman spectroscopy (SERS) provides high sensitivity and rapid molecular fingerprinting, but its clinical translation is hindered by spectral variability and sample heterogeneity. To address these limitations, we developed a novel diagnostic approach by integrating serum exosome SERS with support vector machine (SVM) classification. Systematic evaluation of 27 distinct data preprocessing strategies confirmed that data preprocessing critically influences classification performance, and the optimized model successfully differentiated PD patients from normal controls (NC), achieving an accuracy of 0.85 (95% confidence interval [CI], 0.75-1.00) and an area under the receiver operating characteristic curve (AUC) of 0.85 (95% CI, 0.67-1.00), which was statistically significant as validated by permutation testing (p < 0.05). Comparative analysis with the diagnostic criteria of the Movement Disorder Society (MDS) demonstrated that our model outperforms several conventional MDS methods. Furthermore, this study revealed that the "coffee-ring" effect, which introduces sample heterogeneity during SERS measurements, substantially compromised reproducibility and predictive accuracy. Several post-processing strategies were implemented to mitigate the coffee-ring effect, and notably, one of these strategies achieved results comparable to those of the optimal model (AUC = 0.85, accuracy = 0.85). This demonstrates that such post-processing approaches can effectively suppress the influence of the "coffee-ring" effect. In addition, linear SVM feature importance mapping identified potential exosomal biomarkers, including proteins (S-S stretching, tryptophan), nucleic acids (adenine vibrations, CH₃/CH₂ twisting), lipids, and saccharides. Collectively, these findings highlight a promising strategy for clinical PD diagnosis by combining exosome-based SERS and machine learning, with biomarker identification and heterogeneity analysis further advancing diagnostic reliability and paving the way for practical clinical translation.