Intelligent SERS sensing strategy based on ag NP fractal networks and gradient kernel size CNN for high-precision sweat monitoring

Surface-enhanced Raman scattering (SERS) is a promising tool for sweat analysis, but signal noise, nonlinear responses, and spectral overlap limit its quantitative accuracy in complex samples. In this work, an intelligent SERS sensing strategy was developed by combining an Ag NP fractal substrate with a Gradient Kernel Size Convolutional Neural Network (GKS-CNN). The self-assembled Ag NP substrate provided abundant electromagnetic hotspots, enabling sensitive detection of lactate (LA) and glucose (Glu) from 10 −4 to 10 1  mM, with limits of detection of 108 nM and 224 nM, respectively. PCA results showed that the preprocessed SERS spectra contained distinguishable concentration-related information. Compared with linear regression, traditional machine learning methods, and standard CNNs, GKS-CNN achieved better prediction performance by extracting multi-scale features from nonlinear and overlapping spectra. For pure-component samples, the R 2 values of LA and Glu reached 0.99 and 0.98, respectively; for the joint dataset containing pure and mixed samples, the R 2 values were 0.98 and 0.97, respectively. In real sweat analysis, sample-level splitting was used to reduce data leakage, and the model classified four exercise-related states with an independent test accuracy of 95.00%. The average accuracies of 20 repeated random splits and 5-fold cross-validation were 94.50% and 93.50%, respectively. These results suggest that the proposed SERS-GKS-CNN strategy has potential for non-invasive sweat metabolite analysis and preliminary physiological-state recognition.