Wearable plasmonic SERS patch for quantitative sweat metabolite profiling via dual-track spectral inference

Wearable sweat sensing offers a non-invasive route to monitor human metabolism, but quantitative analysis is limited by strong endogenous backgrounds, spectral overlap, and mechanical instability of the skin. Here, we present a wearable plasmonic surface-enhanced Raman scattering (SERS) patch that enables robust, multiplex sweat analysis under realistic physiological conditions. Gold nanoparticles are generated in situ within a hydrated agarose matrix, forming a diffusion-accessible and mechanically stable plasmonic network compatible with on-skin wear. To resolve low-abundance, intake-related metabolites masked by dominant endogenous signals, we introduce a dual-track deep-learning framework. High-abundance sweat components are quantified directly from mixed spectra, while trace targets are recovered through an unmixing-assisted residual pathway. Using paraxanthine, a primary metabolite of caffeine, we demonstrate quantitative detection across large concentration disparities despite severe spectral congestion. A controlled on-body study combining caffeine intake with exercise-induced sweating reveals subject-specific paraxanthine trajectories with delayed peaks consistent with known pharmacokinetics, while endogenous metabolites show broader exercise-driven trends. Together, this platform advances wearable optical sensing from qualitative detection toward quantitative, personalized metabolic monitoring in complex biofluids.