Deep learning-enhanced SERS biosensing platform for the intelligent differential diagnosis of ankylosing spondylitis...

Ankylosing spondylitis (AS) and osteoarthritis (OA) are two common rheumatic diseases. Although their underlying mechanisms differ significantly-the former being an autoimmune inflammatory condition and the latter a degenerative disorder-they exhibit substantial overlap in early clinical symptoms. Conventional serological tests lack sufficient specificity, leading to a high risk of misdiagnosis. This study developed a highly sensitive surface-enhanced Raman scattering (SERS) biosensing platform based on Ag NPs loaded onto Ti₃C₂T_x MXene for the early diagnosis of AS and OA. High-density, uniform growth of Ag NPs was achieved on MXene surfaces via in situ self-reduction of AgNO₃. This composite system fully leverages MXene's unique metalloid electronic structure to enable efficient interfacial charge-transfer pathways. This facilitates strong synergistic coupling between MXene-induced chemical enhancement (CM) and electromagnetic enhancement (EM) arising from localized plasmon resonance in Ag NPs, significantly boosting the substrate's detection sensitivity. Experiments demonstrated that the composite substrate exhibited optimal SERS activity at an AgNO₃ concentration of 0.1 M. This substrate enabled ultrasensitive detection of the probe molecule (Rhodamine 6G, R6G) with a detection limit as low as 10^-8 M. By comparing SERS spectra from healthy controls (HS), AS, and OA subjects, this study revealed significant spectral differences reflecting specific alterations in serum antioxidant, lipid, and amino acid levels. To achieve intelligent classification and diagnosis of HS, AS, and OA, this study integrated SERS spectroscopy with deep learning algorithms. Three deep learning architectures-Artificial Neural Network (ANN), One-Dimensional Convolutional Neural Network (CNN), and Residual Network (ResNet)-were constructed, and their diagnostic performance was compared. Quantitative evaluation demonstrated that the ANN model established a strong classification benchmark, whereas the ResNet model performed poorly due to overfitting. The CNN model demonstrated clear superiority across all evaluation metrics, achieving 91.36% accuracy and an AUC of 0.9896, effectively capturing subtle spectral fingerprint differences. This platform integrates SERS technology with deep learning, offering a novel tool with significant clinical potential for the rapid, non-invasive, and precise differential diagnosis of AS and OA.