Novel Deep-Learning Unsupervised Domain Adaptation Method for Mitigating Batch, Strain, and Instrument Variations to Enhance Raman Spectroscopy-Based Bacterial Pathogen Identification

The escalating threat of antibiotic resistance demands a rapid and accurate pathogen identification. While Raman spectroscopy combined with deep learning supports fast detection, model robustness remains compromised by domain shifts arising from instrument heterogeneity, batch variability, and strain diversity. To address this, we propose the Raman Spectral Classification Discrepancy Model (RSCDM), a novel domain adaptation framework that dynamically identifies target samples far from the source domain feature distribution via output discrepancies between task-specific classifiers. It adversarially aligns features of different domains to bridge the spectral variability. Unlike traditional methods relying on fixed-domain assumptions, RSCDM leverages classifier output differences to adaptively detect out-of-distribution samples, enabling task-driven domain alignment. Experimental results demonstrate that RSCDM enhances classification accuracy for seven bacterial species across batches and strains from 81.6% to 95.4% using a commercial spectrometer and from 77.5% to 91.3% for six clinical isolates (excluded from training) obtained with a home-built spectrometer. Fine-tuning the pretrained model on reference strains' spectra acquired by both commercial and home-built spectrometers further boosts the identification accuracy of clinical isolates to 99.3%, validating robustness to instrument diversity. Together, these findings address critical challenges in spectral domain adaptation and support translation of single-cell Raman spectroscopy into clinically relevant settings. The results demonstrate that our method improved robustness to batch/strain/instrument variability and reduced identification time from days to minutes under the high-bacterial-load conditions.