Machine learning-enabled label-free SERS for microbial sensing: Toward robust, generalizable, and deployable workflows.

Label-free surface-enhanced Raman spectroscopy (SERS) combined with machine learning (ML) provides a rapid, reagent-light approach to microbial detection, identification, and phenotyping. Translation remains limited, however, by variability in sample preparation, substrates, acquisition conditions, preprocessing, and validation design. We conducted a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-guided, multi-database review of 129 eligible experimental studies, published predominantly between 2021 and 2026, spanning bacterial, viral, fungal, other microbial, and cross-domain applications. The literature was analyzed as an end-to-end workflow comprising target and matrix definition, sample handling, SERS platform design, spectral acquisition and preprocessing, computational modeling and interpretation, and validation. We synthesized trends in substrates, capture and enrichment, chemometrics, classical ML, deep learning, generative modeling, and explainable artificial intelligence, together with emerging applications in antimicrobial resistance, multiplexing, co-infection, and variant or genotype discrimination. A corpus-level audit showed that controlled-development datasets accounted for 68.2% of studies, 62.0% included no more than 25 independent biological units, and large spectral datasets frequently reflected intensive within-sample replication. Moreover, 68.2% lacked biological-unit-aware predictive validation, whereas only 6.2% performed external or major-domain validation. These findings indicate that spectral volume and model complexity alone are insufficient measures of evidential strength. Progress toward deployable microbial-SERS-ML requires biologically independent evaluation, leakage-safe preprocessing, reproducible substrates and sampling workflows, calibration across instruments and matrices, and prospective multi-batch, multi-site validation.