Immuno-labeling SERS technology combined with interpretable 1D-CNN algorithm: sensitive identification of NSE in the serum of stroke patients

A double-antibody sandwich SERS immunoassay was developed for the highly specific capture of NSE in clinical serum. • The interpretable 1D-CNN model achieved an AUC of 0.9704, significantly outperforming DNN and RNN architectures. • SHAP and saliency maps deciphered the “black box”, identifying key Raman shifts at 607 and 1078 cm -1  as biomarkers. • A high-performance PSS@Cu-Ag@Ag SERS substrate with core-shell structure was fabricated using magnetron sputtering. • The periodic array provided dense “hot spots” with superior reproducibility, achieving RSD values below 6.4%. To address the issues of low imaging sensitivity in the clinical diagnosis of ischemic stroke (IS) and the cumbersome procedures associated with traditional Neuron-Specific Enolase (NSE) testing methods, this study proposes a rapid diagnostic approach which integrates Surface-enhanced Raman scattering (SERS) immunonanosensing technology with deep learning. Using magnetron sputtering, a periodic core–shell structured PSS@Cu-Ag@Ag active substrate was fabricated on a periodic microsphere array (PSS). The NSE biomarker in serum was specifically captured via a double-antibody sandwich immunoassay, and the spectra were subjected to deep feature extraction using a one-dimensional convolutional neural network (1D-CNN). Results showed that the model achieved an area under the curve average AUC of 0.9704 under five-fold cross-validation, and the SHapley Additive exPlanations (SHAP) interpreter successfully identified key differential Raman shift such as 607 cm −1 and 1078 cm −1 , enhancing the interpretability of the algorithm. This approach enables robust and transparent classification of clinically heterogeneous samples, providing a highly sensitive and intelligent medical solution with great potential for practical clinical implementation in the auxiliary diagnosis of IS. Download: Download high-res image (125KB) Download: Download full-size image