Deep learning-based cross-device standardization of surface-enhanced Raman spectroscopy for enhanced bacterial recognition

Abstract

Surface-enhanced Raman spectroscopy (SERS) is a powerful, label-free technique for pathogen detection; however, its broader adoption in clinical diagnostics is hindered by inconsistent spectral quality across portable and laboratory-grade instruments, limited cross-device reproducibility, and the poor generalizability of existing machine learning approaches. These limitations restrict reliable and rapid pathogen identification at the point of care. To address this gap, we collected SERS spectra from analytes spread on silver nanorod (AgNR) substrates using four portable Raman systems (Enwave, Tec5, First Defender, and Rapid ID) and one laboratory-grade reference device (Renishaw). The dataset included 20 analyte classes representing clinically relevant bacterial signatures and reference compounds. We propose a deep learning framework comprising: (1) SERS-D2DNet, a one-dimensional sequence-to-sequence neural network that transforms spectra from portable devices into high-fidelity laboratory-grade equivalents, and (2) SuperRaman, a lightweight super-operational neural network (Super-ONN) for efficient multiclass bacterial classification. Primary and ablation studies confirm the complementary role of domain transformation and classification, demonstrating improved feature separability and reduced misclassification rates. Quantitative results show that SERS-D2DNet reduced mean absolute error to 0.01 and increased R² to over 98 % across devices, while SuperRaman achieved up to 100 % classification accuracy post-transformation. Compared to existing approaches, SERS-D2DNet delivered the lowest MAE (0.024 to 0.034), while SuperRaman surpassed state-of-the-art classifiers. The combined framework requires only 6.6 million parameters, a compact 9 MB footprint, and a 3.27 ms inference time, making it well-suited for portable deployment. This study establishes a scalable, real-time solution for rapid sepsis detection and pathogen identification, bridging the performance gap between portable and laboratory-grade SERS systems.