Bioactive Gold Plasmonic Interfaces Meet Artificial Intelligence: Advancing Saliva‐Based Surface‐Enhanced Raman Spectroscopy Diagnostics

ABSTRACT

Bioactive gold plasmonic nanostructures are enabling non‐invasive saliva diagnostics by amplifying weak Raman signals and converting complex biofluid chemistry into information‐rich spectral fingerprints. This Review examines saliva‐based Raman and surface‐enhanced Raman spectroscopy (SERS) integrated with artificial intelligence (AI) from a bioactive‐materials perspective. Saliva is positioned as a cross‐disease liquid biopsy relevant to oncological, infectious, inflammatory, metabolic, neurodegenerative, and pulmonary disorders. Particular emphasis is placed on how gold nanostructure geometry, substrate architecture, surface functionalization, antifouling strategies, affinity capture, and protein‐corona formation govern molecular access to electromagnetic hot spots, spectral reproducibility, and computational reliability. Case studies spanning oncology, infectious, inflammatory, and pulmonary disorders illustrate how substrate engineering and dataset construction jointly determine diagnostic performance and clinical portability. Because human studies combining saliva, gold‐enabled SERS, and AI remain limited, explicit inclusion criteria distinguish direct salivary evidence from contextual studies using related biospecimens. Clinically viable saliva‐based SERS diagnostics require coordinated co‐design of bioactive plasmonic interfaces and computational pipelines, together with standardized pre‐analytics, reproducible substrates, rigorous patient‐level validation, shared datasets, and independent multicenter evaluation.

1. Introduction

The clinical need for earlier and less burdensome diagnosis extends well beyond a single disease class. Oncology requires stage‐sensitive detection and longitudinal surveillance; infectious diseases require rapid identification of transmissible cases; inflammatory disorders require objective assessment of activity and endotype; metabolic diseases require repeatable monitoring; and neurodegenerative diseases require biomarkers that may precede overt functional decline. These shared needs motivate diagnostic platforms that can interrogate accessible biofluids repeatedly while preserving molecular breadth [1, 2, 3, 4, 5].

Liquid biopsy addresses this need by sampling disease‐associated cells, nucleic acids, extracellular vesicles, proteins, metabolites, hormones, and immune mediators from accessible fluids. Although blood remains the dominant matrix, saliva, urine, and breath‐derived specimens can reduce procedural burden and support decentralized or serial testing [2, 4, 6].

Saliva is especially attractive because collection is painless, repeatable, and compatible with community and point‐of‐care workflows. It also contains a biologically diverse mixture of proteins, nucleic acids, metabolites, cytokines, hormones, microbial products, and extracellular vesicles that reflects both local oral biology and systemic physiology. At the same time, variable flow rate, circadian rhythm, hydration, diet, oral inflammation, and sample handling can alter its measured composition, making analytical control essential [1, 2, 7, 8].

Disease‐associated salivary signatures have now been investigated in oral and systemic cancers, viral infections, inflammatory airway disease, metabolic dysfunction, and neurodegeneration. Representative examples include saliva SERS or Raman classification in oral cancer, COVID‐19, childhood asthma, and Parkinson's disease. In this Review, the term surface‐enhanced Raman spectroscopy coupled with artificial intelligence (SERS‐AI) denotes workflows in which a SERS‐derived spectrum is interpreted by a multivariate, machine‐learning, or deep‐learning model rather than by a single peak threshold (Figure 1) [5, 9, 10, 11].

FIGURE 1.

End‐to‐end concept for a saliva SERS‐AI diagnostic platform. Saliva is applied to a gold‐based plasmonic interface, where nano‐bio interactions determine which molecules reach electromagnetic hot spots and therefore shape the measured spectrum. A prespecified computational pipeline then converts the amplified fingerprint into a diagnostic output.

Raman spectroscopy provides a label‐free vibrational fingerprint of molecular composition, but conventional Raman scattering is weak and can be obscured by fluorescence. Surface‐enhanced Raman spectroscopy addresses this limitation by concentrating electromagnetic fields near plasmonic nanostructures and by positioning analytes close to those enhanced fields. Consequently, the observed signal depends not only on the sample but also on surface chemistry, nanogap distribution, molecular orientation, and adsorption kinetics [12, 13, 14, 15].

Artificial intelligence becomes valuable because saliva spectra are high‐dimensional and contain distributed biochemical information. Classical machine learning can integrate selected bands or principal components, whereas deep learning can derive hierarchical representations directly from spectra. These benefits are conditional: patient‐level data partitioning, leakage‐resistant preprocessing, transparent validation, and evaluation across substrate batches and instruments are necessary to prevent models from learning technical artifacts [16, 17, 18, 19].

Gold occupies a central position in this materials‐computation interface because its optical response is tunable, its surface is chemically stable, and thiol, polymer, antibody, and aptamer chemistries can be used to control biomolecular capture and antifouling behavior. These properties support reproducible device architectures ranging from colloids and paper substrates to microfluidic and hybrid platforms [20, 21, 22, 23].

Within this broad diagnostic landscape, oncology remains an instructive application because it has generated comparatively mature saliva‐SERS case studies and clinically meaningful classification tasks. It is therefore treated here as one evidence‐rich domain rather than as the sole rationale for the technology. Infectious, inflammatory, metabolic, neurodegenerative, and pulmonary applications are introduced first and revisited where the strength of direct saliva evidence permits [2, 9, 10, 24].

Accordingly, this Review links three layers: the biological and pre‐analytical constraints of saliva; the materials science of gold‐enabled SERS and nano‐bio interface engineering; and the computational practices needed to convert spectral fingerprints into credible clinical outputs. Particular attention is given to the difference between direct human saliva + Au‐SERS + AI evidence and adjacent studies using non‐gold substrates, conventional Raman spectroscopy, breath, sputum, or respiratory droplets.

The Review first describes saliva as a liquid‐biopsy medium and then examines Raman enhancement, gold nanostructure design, and interface‐controlled sampling. It next evaluates classical and deep‐learning workflows before synthesizing explicitly selected case studies. The final sections address reproducibility, pre‐analytics, validation, regulation, manufacturability, and external evaluation.

By positioning saliva SERS‐AI as a cross‐disease, materials‐led, computation‐enabled diagnostic technology, this Review aims to distinguish demonstrated capability from plausible but unvalidated opportunity and to identify the design controls most likely to enable reproducible clinical translation.

2. Saliva as a Liquid Biopsy Medium

2.1. Composition and Diagnostic Relevance of Saliva

Saliva is an exocrine secretion produced by major and minor salivary glands and is composed of approximately 99% water, with the remaining fraction containing proteins, inorganic ions, and diverse small organic and nitrogenous molecules [8]. Although proteins represent only a minor mass fraction, saliva contains a broad and information‐rich proteome that supports diagnostic applications. Classic and abundant components include enzymes and host‐defense proteins such as amylase, peroxidase, lipase, lysozyme, lactoferrin, kallikrein, and cystatins, together with hormones and growth factors [8]. Proteome‐wide analyses have identified thousands of salivary proteins, with a large proportion belonging to major families including α‐amylases, proline‐rich proteins (PRPs), mucins, histatins, statherin, and related peptide derivatives [25]. These macromolecules coexist with nucleic acids and extracellular vesicles, enabling saliva to carry multiple classes of disease‐relevant information beyond conventional protein biomarkers [2, 26, 27, 28].

From a physiological standpoint, saliva production is dominated by the submandibular, parotid, and sublingual glands, and daily secretion volume is commonly in the 0.5–1.5 L range in healthy adults [29]. Saliva also contains inorganic ions such as sodium, potassium, calcium, magnesium, chloride, and carbonate, which influence viscosity, buffering capacity, and electrostatic screening, all of which can shape biomolecule adsorption and analytical readouts [8]. Because of its direct contact with the oral mucosa and upper aerodigestive tract, saliva can reflect local pathophysiology, including oral premalignancy and oral squamous cell carcinoma [7, 28, 30]. At the same time, saliva can also capture systemic signals via transudation, glandular filtration, and vesicle‐mediated transport, enabling diagnostic exploration beyond oral disease, including oncology, metabolic, and infectious conditions [1, 3, 31]. This “local‐plus‐systemic” feature is a key reason saliva has attracted sustained interest as a non‐invasive liquid biopsy medium [2, 4, 32].

2.2. Advantages and Limitations Relative to Other Biofluids

Saliva offers practical advantages that align well with screening and longitudinal monitoring. Collection is non‐invasive and typically painless, enabling repeat sampling with low cost and minimal requirements for specialized personnel or infrastructure [2, 7, 26]. Handling risk is also generally lower than blood‐based sampling, and saliva contains innate factors that help limit certain transmission risks, contributing to a low likelihood of oral transmission for some pathogens under typical conditions [33, 34]. These properties support scalability for population‐level programs and frequent monitoring strategies, especially when paired with rapid analytical platforms.

Table 1 summarizes these comparative trade‐offs across saliva, blood, and urine. In this framework, saliva stands out for its breadth of biomarker classes, including proteins, cfDNA, miRNA, and exosomes, coupled with low‐biohazard, low‐cost oral collection [1, 4, 7, 35]. On the other hand, blood remains the most information‐dense liquid biopsy matrix in many contexts because of higher biomarker loads and established workflows for circulating analytes, but this comes with invasiveness, patient burden, and higher operational costs [36]. Urine is similarly non‐invasive and convenient, but composition can vary substantially with hydration, renal handling, and collection timing, which can complicate standardization for screening applications [6, 36].

TABLE 1.

Comparative Overview of Saliva‐Based Biofluids for Liquid Biopsy Applications.

Biofluid	Key biomarkers	Collection method	Advantages	Limitations	Refs.
Saliva	Proteins, cfDNA, miRNA, exosomes	Non‐invasive (oral collection)	Painless, low‐cost, low biohazard	Lower analyte concentration	[1, 4, 7]
Blood	CTCs, cfDNA	Invasive (venipuncture)	High biomarker load	Pain, risk, expensive	[36, 37]
Urine	Metabolites, cfDNA	Non‐invasive	Simple collection	Variable composition	[6, 36]

Abbreviations: CfDNA, cell‐free DNA, short DNA fragments released into biofluids from apoptotic or necrotic cells, including tumor‐derived DNA in cancer settings; miRNA, microRNA, small non‐coding RNA molecules that regulate gene expression and are stable in biofluids; CTCs, circulating tumor cells, intact cancer cells shed from primary or metastatic tumors into the bloodstream.

The main limitations of saliva are rooted in analytical sensitivity and matrix complexity. Many disease‐associated analytes occur at lower concentrations than in blood, and salivary measurements can be influenced by flow rate, circadian effects, hydration, oral inflammation, and the presence of mucins as well as other macromolecules that affect viscosity and adsorption behavior [2, 3, 29, 38]. These constraints are directly relevant for Raman and SERS workflows because molecular accessibility at the sensing interface and signal‐to‐noise performance can be modulated by ionic strength, protein corona formation, and heterogeneous biomolecule mixtures [2, 25]. Consequently, saliva‐based liquid biopsy platforms benefit from careful pre‐analytics and from detection modalities capable of resolving low‐abundance molecular signatures within a complex background, which motivates the integration of plasmonic enhancement and computational decoding in subsequent sections of this Review.

3. Raman Spectroscopy Enhanced by Gold Nanostructures

3.1. Fundamentals of Raman Spectroscopy and SERS Enhancement

Raman spectroscopy (RS) is a vibrational spectroscopic technique that probes molecular structure by measuring the inelastic scattering of monochromatic light. When incident photons interact with chemical bonds, a small fraction is scattered with an energy shift that corresponds to specific vibrational modes, generating a molecular “fingerprint” spectrum that can reflect proteins, nucleic acids, lipids, and metabolites in complex biofluids [39]. From an instrumentation standpoint, RS integrates: (a) a stable excitation source, (b) optics and spectral filters to reject elastically scattered light, (c) a spectrograph/monochromator to disperse Raman‐shifted photons, and (d) sensitive detectors, supported by increasingly capable analysis pipelines [39]. This combination is particularly attractive for liquid biopsy concepts because the Raman fingerprint is label‐free and multiplexed by nature, capturing many biochemical features simultaneously rather than tracking a single marker.

Despite these advantages, conventional RS is often limited in bioanalysis by weak Raman cross‐sections and the presence of background fluorescence, which can obscure subtle disease‐associated signals in saliva. A common solution is surface‐enhanced Raman spectroscopy (SERS), which amplifies Raman signals by placing analytes in close proximity to plasmonic nanostructures. Mechanistically, SERS enhancement is typically explained by two contributions. First, electromagnetic enhancement arises when localized surface plasmon resonances generate intense near‐fields at sharp features, nanogaps, and interparticle junctions, creating “hot spots” that can boost Raman scattering by many orders of magnitude [13, 15]. Second, chemical enhancement is associated with short‐range charge‐transfer interactions between the adsorbate and the metal surface, which can modify molecular polarizability and further increase Raman intensity [14, 23]. In practice, clinically relevant SERS performance depends on how reliably a substrate creates hot spots, how stably biomolecules approach those hot spots, and how reproducibly spectra can be acquired across batches, operators, and instruments.

The feasibility of RS and SERS for saliva‐related oncology is supported by translational studies that pair spectroscopy with clinically meaningful classification tasks. For example, fused Raman analysis across blood and saliva has been reported to deliver high diagnostic accuracy in head and neck cancer, illustrating that saliva can carry useful spectral information when appropriately analyzed [40]. Complementary work has also explored Raman fingerprints in cancer‐associated tissues and related biospecimens, reinforcing the broader oncology relevance of vibrational spectroscopy [37, 41]. These studies motivate the materials question that becomes central for saliva. How do we engineer plasmonic interfaces that can consistently extract disease‐informative spectral signatures from a dilute, protein‐rich, and variable biofluid.

3.2. Gold Nanostructures as Bioactive Plasmonic Materials

Gold (Au) occupies a privileged position in biomedical plasmonics because it combines tunable optical resonances with chemical stability and strong compatibility with biofunctionalization chemistries. From a materials perspective, Au supports localized plasmon modes that can be tuned across the visible and near‐infrared by controlling morphology, aspect ratio, and interparticle coupling, enabling designers to match excitation wavelengths and optimize near‐field distributions for specific sensing geometries [20, 21, 22, 42]. Compared with more reactive plasmonic metals, Au offers surface robustness in complex biological matrices and a mature toolbox for forming self‐assembled monolayers, thiol linkages, and polymer coatings that stabilize performance under real‐world sampling conditions [14, 23].

Importantly, describing Au as “bioactive” in this context does not mean intrinsic pharmacology. It means the material surface can be engineered to actively control nano‐bio interactions, including biomolecule adsorption, target capture, and spectral proximity to hot spots. This is precisely the scenario in saliva diagnostics. Salivary biomolecules, including proteins, miRNA, cfDNA, and exosome‐associated cargo, must interact with the plasmonic surface in a controlled way to generate a stable, information‐rich SERS fingerprint. The schematic in Figure 2 highlights this logic. Saliva droplets deliver diverse targets to an Au SERS substrate, where engineered nano‐bio interfaces determine which biomolecules are preferentially presented to the hot spots. The resulting enhanced Raman fingerprint is then suitable for AI‐based clustering and classification.

FIGURE 2.

Schematic illustration of saliva‐based biomarker sensing enabled by bioactive gold‐plasmonic SERS substrates for non‐invasive disease screening. Salivary biomolecules, including DNA, miRNA, exosomes, and proteins, interact with engineered plasmonic nano–bio interfaces, where localized surface plasmon resonances amplify molecular Raman signatures. The resulting spectra are analyzed using machine‐learning algorithms to differentiate healthy and cancer‐associated profiles. The t‐SNE plot illustrates AI‐derived clustering of spectral features, demonstrating material‐enabled discrimination of disease‐specific salivary biosignatures.

A plasmonic interface should ultimately be judged by the biochemical information and reproducibility of the spectrum it produces. A recent AuNP plasmonic droplet assay measured saliva from 114 participants spanning normal, premalignant, and malignant oral conditions and reported class‐dependent changes in aromatic amino‐acid, amide, collagen, lipid, and nucleic‐acid‐associated bands. Figure 3 therefore replaces the former generic sensing schematic with representative clinical saliva SERS spectra, while Box 1 summarizes selected assignments needed to interpret the spectral differences [10].

FIGURE 3.

Representative averaged saliva SERS spectra from normal (a), premalignant (b), and malignant (c) oral‐condition groups, together with an overlaid comparison (d). The spectra illustrate how a gold‐nanoparticle plasmonic droplet assay captures disease‐associated changes across the salivary biochemical fingerprint. Key assignments are summarized in Table 2. Adapted from (Puravankara et al., 2026) under the Creative Commons Attribution 4.0 license.

Table 3 provides a practical snapshot of how Au nanostructure geometry maps to SERS function in biofluid analysis, spanning tip‐enhanced architectures (nanostars), mode‐tunable rods, hybrid Au@Ag constructs, and composite platforms such as Au‐graphene hybrids. The value of this table is that it connects morphology to enhancement mechanism and then to the biofunctionalization strategies required for real samples.

TABLE 3.

Gold Nanostructures Used in SERS for Biofluid Analysis.

Gold nanostructure type	Morphology/Size	Enhancement mechanism	Biofunctionalization	Example biomarkers/Targets	Reported EF or accuracy	Refs.
Au Nanostars	Branched (50–100 nm)	Tip‐localized LSPR hot spots	PEG/thiol/aptamer	Oral cancer proteins	EF ∼101 2	[14, 22, 43, 44]
Au@Ag Core–Shells	Core 40 nm, shell 5 nm	Hybrid plasmonic coupling	Thiolated DNA	Exosomal miRNAs	94% classification accuracy	[23, 24, 45]
Au Nanorods	Aspect ratio 3–5	Longitudinal LSPR tunability	Antibody conjugation	Cytokines/IL‐6	EF ∼1010	[20, 42, 46]
Au‐Graphene Hybrids	2D–3D composite	Charge‐transfer + EM enhancement	π–π stacking/linker	Saliva metabolites	Enhanced reproducibility	[21, 22, 23]

Abbreviations and notes: SERS, surface‐enhanced Raman scattering; EF, enhancement factor (values depend strongly on calculation method and measurement conditions; reported here for within‐study context only and not directly comparable across studies) values depend strongly on method; reported for within‐study context only); LSPR, localized surface plasmon resonance; EM, electromagnetic; PEG, polyethylene glycol; miRNA, microRNA; IL‐6, interleukin‐6. Hot spots denote nanoscale regions of intensified electromagnetic fields. Plasmonic coupling refers to field amplification arising from interacting metallic nanostructures. Charge‐transfer enhancement indicates chemical Raman enhancement via electron exchange between analyte and metal surface. π–π stacking describes non‐covalent adsorption of aromatic molecules onto graphene surfaces.

3.3. Substrate Engineering and Nano‐Bio Interfaces for Saliva

Saliva is an analytically challenging matrix because it is dilute in many disease‐associated analytes yet rich in macromolecules that readily adsorb to surfaces. A recurring issue is the formation of a protein corona, where abundant proteins and glycoproteins compete for binding sites and can alter both target capture and the local electromagnetic environment. This can be beneficial if it stabilizes adsorption and improves reproducibility, or detrimental if it masks low‐abundance markers and adds spectral background. Therefore, SERS translation in saliva is as much an interface problem as it is a plasmonics problem [14, 23].

Surface functionalization becomes the control knob that converts a “reactive” metal surface into a selective sensing interface. PEGylation is widely used to reduce nonspecific adsorption and stabilize colloids [44]. Thiol chemistry enables robust anchoring of ligands and reporters on Au, and aptamers or antibodies introduce molecular recognition that can enrich biomarkers such as cytokines, exosomal targets, or cancer‐associated proteins [23]. These strategies are not merely additive. Their combined effect determines whether analytes are positioned within a few nanometers of hot spots, which is the distance regime where SERS becomes diagnostically powerful [13, 15].

Device architectures further shape performance in saliva. Paper‐based substrates and microfluidic platforms are attractive because they can standardize sample handling and mitigate matrix effects through controlled flow, filtration, or preconcentration, while remaining compatible with low‐cost manufacturing. In a clinically oriented example, portable saliva SERS platforms have been reported for detecting oral pre‐malignant lesions and carcinomas, indicating toward point‐of‐care workflows where substrate design, sampling, and analytics are tightly coupled [43]. At the same time, earlier saliva SERS studies have highlighted the risk of poor signal‐to‐noise ratios when exogenous chemicals or native saliva components interfere with spectral readout, motivating practical pretreatments such as conditioning steps, protein cleanup, or controlled drying protocols [12, 41]. More broadly, optical biofluid diagnostics beyond SERS, such as optical coherence tomography‐based approaches for biochemical sensing, reinforce the shared challenge of extracting stable signals from heterogeneous biological environments [47].

Within this landscape, Table 3 is most useful when read as a design map rather than a catalog. Nanostars can create intense tip‐localized hot spots, rods offer longitudinal plasmon tunability, Au@Ag structures can provide hybrid coupling, and Au‐graphene composites can introduce charge‐transfer contributions and improve reproducibility through more structured adsorption environments. The key message, consistent with Figure 2, is that substrate morphology and surface chemistry must be co‐designed with saliva's matrix constraints.

3.4. Why Gold Deserves Explicit Emphasis in Clinical SERS?

Gold merits explicit emphasis not because it is the only plasmonic option, but because it aligns unusually well with the full translational chain. First, Au has strong regulatory familiarity and a long history of biomedical use in coatings and nanomaterial contexts, which reduces perceived risk when moving from proof‐of‐concept studies to clinical validation and manufacturing [20, 22, 42]. Second, Au‐based substrates are workflow compatible. They can be integrated into paper strips, microfluidic cartridges, and benchtop or portable Raman readers, enabling sample‐to‐answer pipelines that fit clinical time constraints and operator realities [23, 43]. Third, manufacturability and reproducibility are more tractable with Au than with many alternatives, because Au supports stable surface chemistries and scalable fabrication routes spanning colloidal synthesis, templated growth, and printed or coated architectures [21, 23].

However, gold's translational advantage is only realized when reproducibility is treated as a primary design requirement. Hot spot variability, batch‐to‐batch drift, and uncontrolled adsorption are the most common reasons SERS performance fails to generalize outside a single laboratory. This is why a clinically credible Au‐SERS platform must report substrate morphology distributions, enhancement factor acceptance windows, and standardized saliva handling steps, and should explicitly address nano–bio interface control rather than treating it as a secondary detail [14, 23]. In other words, Au supports translation, but only when gold's surface is engineered as a biomaterials interface, not merely as a plasmonic amplifier.

3.5. Design Principles Governing Nano‐Bio Interactions in Saliva Au‐SERS Platforms

Although saliva‐based SERS diagnostics are frequently described as analytical or computational workflows, clinical performance is primarily dictated by the material‐biology interface rather than the downstream algorithm. The Raman spectrum reflects molecular sampling occurring within localized electromagnetic hot spots surrounding plasmonic nanostructures, and therefore the measurement corresponds to the fraction of biomolecules effectively presented to the interface. Diagnostic reproducibility thus depends on whether the substrate consistently interrogates comparable biochemical populations across samples rather than on maximizing absolute enhancement. From this perspective, Au–SERS saliva diagnostics should be interpreted as an interfacial bioengineering challenge rather than purely a spectroscopic optimization problem [13, 14, 15, 23, 48].

3.5.1. Interaction Distance Defines the Measured Biology

SERS enhancement decreases rapidly with increasing distance from the plasmonic surface and is effectively confined to molecules located within only a few nanometers of the nanostructure. As a consequence, the recorded spectrum does not represent bulk saliva composition but rather the subset of biomolecules capable of reaching and remaining within this interaction volume. In unmodified substrates, abundant salivary proteins and glycoproteins dominate adsorption and define the detected fingerprint, meaning the measurement primarily reflects the surface‐bound molecular layer instead of low‐abundance disease markers. Therefore, the central design objective is controlling molecular proximity and accessibility rather than simply maximizing enhancement factor [2, 13, 14, 15, 23, 25, 46].

Design principle 1: Diagnostic specificity is governed by controlled molecular positioning at the plasmonic interface, not by bulk biomarker abundance.

3.5.2. The Protein Corona Is Not Noise but the Primary Signal Gatekeeper

Upon exposure to saliva, nanoparticle surfaces rapidly acquire a dynamic protein corona that determines molecular access to hot spots and therefore shapes the measurable biochemical information. Rather than eliminating this layer, clinically reliable platforms stabilize a reproducible interfacial composition to minimize spectral variability. Antifouling and interface‐regulating coatings, including PEGylated or zwitterionic chemistries, reduce stochastic adsorption and convert a random interface into a controlled sampling environment. Consequently, substrates with moderate enhancement but stable chemistry often provide more reliable classification performance than highly enhancing but chemically unstable surfaces [14, 23, 44, 46, 49].

Design principle 2: Reproducible corona composition outweighs maximal electromagnetic enhancement for diagnostic robustness.

3.5.3. Selectivity Emerges From Competitive Adsorption, Not Only Affinity Ligands

Affinity‐based strategies such as aptamers or antibodies enrich target species, but in complex biofluids they mainly bias competitive adsorption rather than isolate a single molecular entity. The resulting Raman signature therefore remains multiplexed and represents a biochemical microenvironment instead of a single biomarker peak. Functionalization operates probabilistically by increasing the likelihood that disease‐related molecular ensembles occupy hot spots across measurements rather than producing absolute molecular exclusivity [23, 48, 50].

Design principle 3: Functionalization tunes adsorption probability distributions rather than producing single‐analyte detection.

3.5.4. Substrate Morphology Determines Biological Variability Sampling

Hot spots represent discrete nanoscale sampling volumes whose spatial distribution controls how biological heterogeneity is captured. Highly heterogeneous substrates sample molecular populations unevenly, increasing spectral variance and potentially exaggerating dataset separability. More uniform architectures average biochemical variability and improve generalization across patients and acquisition conditions. Accordingly, clinically relevant SERS requires statistical control of hot‐spot distribution rather than maximization of peak intensity alone [14, 15, 23].

Design principle 4: Homogeneous field distribution improves biological representativeness, whereas extreme hot‐spot localization amplifies sampling bias.

3.5.5. AI Functions as a Translation Layer for Material‐Encoded Information

Machine learning does not generate diagnostic information but interprets patterns [51] encoded by the nano‐bio interface. When interfacial chemistry is unstable, models learn substrate or batch artifacts; when it is controlled, models capture biological signatures that generalize across cohorts. Consequently, computational robustness is downstream of materials engineering, and algorithmic performance reflects interface reproducibility [9, 16, 24, 52, 53].

Design principle 5: AI reliability is a consequence of interface reproducibility.

3.5.6. A Unified Materials‐Data Framework for Clinical Translation

Collectively, a clinically deployable saliva SERS platform should prioritize controlled interaction distance, stabilized corona composition, probabilistic molecular enrichment, uniform hot‐spot statistics, and dataset variability aligned with material reproducibility. Under these conditions, SERS functions not only as a high‐sensitivity spectroscopic technique but as a standardized biochemical sampling platform, enabling computational models to operate on reproducible biological representations rather than measurement artifacts [14, 16, 23, 48].

4. Artificial Intelligence for Raman and SERS Spectral Interpretation

The spectral fingerprints generated at bioactive plasmonic nano‐bio interfaces are intrinsically high‐dimensional representations of interfacial molecular sampling. Converting these material‐encoded biochemical signatures into clinically actionable outputs requires computational frameworks capable of extracting structured patterns from complex spectra. Artificial intelligence (AI) has therefore emerged as a critical enabler in data‐rich experimental sciences, where machine learning (ML) supports automated pattern recognition and predictive modeling across complex multidimensional datasets [54, 55]. Rather than serving solely as a computational accelerator, ML provides a framework for extracting physically meaningful information from measurements whose dimensionality exceeds direct human interpretation [51]. In Raman and surface‐enhanced Raman spectroscopy (SERS), this capability is particularly relevant because spectral signatures originate from heterogeneous molecular ensembles influenced by both biochemical composition and nano‐interface interactions [16].

ML algorithms iteratively learn relationships between spectral features and target outcomes by analyzing labeled or unlabeled datasets and subsequently generalizing these relationships to new measurements [56, 57]. This approach significantly reduces manual preprocessing and enables rapid interpretation of large spectral libraries, accelerating experimental workflows and facilitating real‐time sensing applications. Consequently, machine learning has become integral to Raman analysis across biomedical and chemical applications, where it enables classification, regression, and anomaly detection within complex spectra [16, 58].

From a sensing perspective, the role of ML extends beyond data analysis to functional integration within intelligent diagnostic platforms. Automated signal processing coupled with adaptive learning algorithms allows SERS devices to operate as decision‐support systems capable of handling biological variability and measurement noise [17]. These capabilities are particularly important in biofluid analysis, where spectral features arise from overlapping molecular contributions and cannot be interpreted through single‐peak analysis alone [59].

Machine learning approaches applied to SERS are broadly categorized as supervised or unsupervised learning paradigms depending on whether labeled reference outputs are available [60]. Unsupervised learning identifies intrinsic structure and clustering within spectral datasets, whereas supervised learning predicts defined biochemical or diagnostic outcomes from previously annotated spectra [59]. Widely used supervised methods include Partial Least Squares Discriminant Analysis (PLS‐DA), Support Vector Machines (SVM), gradient boosting algorithms, and nearest‐neighbour classifiers, which have demonstrated effectiveness in biological classification and correlation analysis [58].

Importantly, the performance of conventional ML models strongly depends on the quality and stability of spectral feature representation, often requiring careful preprocessing and feature engineering. This dependence highlights the close relationship between material design and computational performance: reproducible nano‐bio interfaces generate consistent spectral patterns that are learnable by algorithms, whereas unstable interfaces produce dataset‐specific artifacts that reduce generalizability [16, 17].

When material variability is not explicitly controlled, substrate‐dependent spectral fluctuations can become inadvertently correlated with diagnostic labels, effectively introducing a form of label leakage into the training process. For example, differences in nanoparticle batch, hot‐spot density, drying conditions, or acquisition settings may co‐vary with clinical grouping during data collection. In such cases, models may learn enhancement patterns, interface chemistry signatures, or preprocessing artifacts rather than true disease‐associated biochemical changes. This can inflate apparent performance during internal validation while leading to rapid degradation under external testing, cross‐site deployment, or substrate replacement. Therefore, rigorous experimental design that decouples biological variation from material and acquisition variability is essential to ensure that predictive performance reflects clinically meaningful signal rather than interface‐driven confounding.

4.1. Classical Machine‐Learning Pipelines for SERS Data

SERS measurements from bioactive plasmonic substrates convert complex saliva chemistry into high‐dimensional spectral fingerprints. Classical machine learning combines prespecified preprocessing with feature‐engineered models, but every learned preprocessing step must be fitted within the training data to avoid leakage. For clinical samples, spectra from the same participant must remain in the same partition; nested cross‐validation is preferred for model selection, and a locked external cohort is needed before claims of transportability (Figure 4) [16, 17, 18, 19, 57].

FIGURE 4.

Leakage‐aware patient‐level validation framework for saliva SERS–AI diagnostics. (A) Participants, rather than individual spectra, are assigned to development and external‐test cohorts, ensuring that all spectra from the same participant remain within a single partition. Development data can additionally be stratified by acquisition site or substrate batch, whereas the external cohort remains locked during model development. (B) Nested validation separates model optimization from performance estimation: baseline correction, normalization, feature selection, and hyperparameter tuning are performed exclusively within training folds of the inner loop, while the outer loop provides a patient‐level estimate of model performance. (C) After preprocessing and model parameters are frozen, one‐time evaluation on an independent external cohort provides discrimination metrics, uncertainty estimates, subgroup analyses, and wavenumber‐level explanations. This workflow minimizes information leakage and reduces the risk that apparent diagnostic performance reflects participant overlap, preprocessing leakage, substrate‐batch effects, or instrument‐specific artifacts rather than reproducible biochemical signatures. SERS, surface‐enhanced Raman spectroscopy; AI, artificial intelligence; AUC, area under the receiver operating characteristic curve.

Preprocessing is used to reduce non‐biological variance introduced by baseline drift, fluorescence background, cosmic rays, and intensity differences across runs, typically via baseline correction, denoising, normalization, and, where needed, spectral alignment [16, 56, 57, 58]. Feature engineering then compresses spectra into informative variables, for example principal component analysis (PCA) scores, band ratios, or selected peak intensities, which can stabilize performance when datasets are modest [17, 57].

Within this classical paradigm, support vector machines (SVM) remain popular for Raman and SERS classification because they often perform well in high‐dimensional spaces with limited samples, provided preprocessing is consistent [9, 61, 62]. Random forests (RF) provide an alternative that is comparatively robust to noisy features and can offer straightforward measures of feature importance, which can be helpful when translating models to clinicians and assay developers [53]. Partial Least Squares Discriminant Analysis (PLS‐DA) is widely used in spectroscopic diagnostics because it is interpretable and directly linked to latent‐variable decomposition, but it may underperform when class boundaries are strongly nonlinear [58]. Collectively, these methods form the “feature‐engineered” branch in Figure 4, and Table 4 can be positioned as a compact map of the dominant algorithm families and their typical trade‐offs.

TABLE 4.

AI and Machine Learning Algorithms Applied to Raman/SERS Data.

Algorithm	Learning type	Application/Dataset	Feature engineering	Accuracy/AUC	Key strength	Limitation	Refs.
SVM	Supervised	Saliva SERS (COVID‐19, oral CA)	PCA features	90%–95%	High precision with small data	Sensitive to noise	[52, 53, 62]
Random Forest (RF)	Supervised	Saliva SERS (COVID‐19)	Raw spectra	94%	Robust, interpretable	Requires feature tuning	[53]
PLS‐DA	Supervised	Serum/Saliva, cancer vs. control	Preprocessed bands	85%–90%	Simple, interpretable	Limited nonlinear capture	[63]
CNN/ResNet	Deep Learning	Exosome SERS (cancer)	None (end‐to‐end)	95%–98%	Learns hierarchical features	Data‐hungry	[45, 58]
Transfer Learning Models	Deep Learning	Cross‐device spectral sets	Fine‐tuning	>90%	Portable	Needs curated base model	[64]

Abbreviations and notes: SVM, support vector machine; RF, random forest; PLS‐DA, partial least squares discriminant analysis; CNN, convolutional neural network; ResNet, residual neural network; PCA, principal component analysis; AUC, area under the receiver operating characteristic curve. Supervised learning denotes training with labeled data. Deep Learning indicates a multilayer neural network method that automatically learns features from raw data. Transfer learning refers to fine‐tuning pretrained models on new spectral datasets. End‐to‐end learning indicates automatic feature extraction from raw spectra. Preprocessed bands represent selected Raman spectral regions after baseline correction and normalization.

4.2. Deep Learning for End‐To‐End Spectral Decoding

Deep learning shifts the burden from hand‐crafted features to learned representations [65]. Convolutional neural networks (CNNs) can ingest minimally processed spectra and learn hierarchical patterns that combine local peak structure with broader spectral context, improving resilience to subtle baseline or alignment variation when training data are adequate [16, 66]. Residual architectures (ResNet‐style) further support deeper models as well as stable optimization, and have been applied to SERS classification problems where inter‐sample spectral variation is complex [67].

For bioactive‐materials‐enabled assays, this is important because the substrate and nano‐bio interface strongly shape the distribution of observed spectra. When those materials‐driven effects vary across batches or sites, deep models can sometimes learn more transferable representations than rigid peak‐based rules, but only if the training set captures realistic variability [16, 17]. In cancer‐oriented SERS, deep learning has also been used to decode multiplexed spectral signatures in vesicle‐related or biomarker‐rich contexts, supporting the “end‐to‐end” branch depicted in Figure 3 [64]. The main constraint is data: deep models are more data‐hungry and can overfit if spectra are few or if confounders (instrument, operator, batch) correlate with labels [16, 17]. This makes dataset design, protocol standardization, and external validation essential to claim clinical readiness [68].

4.3. Multimodal, Transfer, and Federated Learning Strategies

Clinical translation requires models that generalize across instruments, substrate batches, and collection sites, because real‐world Raman and SERS signals reflect both biology and materials‐dependent measurement physics. Transfer learning is a practical route: a model trained on a large spectral collection can be fine‐tuned with a smaller dataset from a new device or site, reducing the need to rebuild models from scratch [16, 64]. This is directly relevant for saliva SERS, where differences in laser wavelength, spectrometer response, and substrate fabrication can shift intensities and baselines even when the underlying biology is similar.

Beyond transfer learning, multimodal learning can combine spectra with clinical metadata (age, symptoms, risk factors) to improve calibration and reduce spurious decisions, provided the fusion strategy is prespecified and validated [17]. Federated learning is often discussed as a way to train across hospitals without pooling raw patient data; therefore, it has been proposed as a privacy‐preserving framework for multi‐center model training. Its conceptual alignment with distributed spectroscopic datasets suggests a promising direction for future saliva‐based SERS studies [16, 17, 18]

5. Evidence Scope and Case Studies of Saliva‐Based Raman/SERS‐AI Diagnostics

Only a limited number of peer‐reviewed human studies currently combine saliva, SERS, and artificial intelligence or multivariate classification, and the subset that also uses a gold‐containing plasmonic substrate is smaller still. To avoid overstating maturity, case studies were selected when they: (i) analyzed human whole saliva, saliva supernatant, or a clearly defined salivary fraction; (ii) used Raman or SERS spectral acquisition; (iii) applied multivariate statistics, machine learning, or deep learning to a diagnostic classification task; (iv) reported sample size and quantitative performance; and (v) were peer‐reviewed original studies available. Direct saliva + Au‐SERS + AI studies are designated as core evidence. Non‐Au SERS and conventional Raman saliva studies are retained only as explicitly labeled contextual comparators. Breath, sputum, respiratory‐droplet, and facemask studies are discussed for translational relevance but are not counted as saliva + Au‐SERS + AI evidence [5, 9, 10, 11, 24, 43].

From a bioactive materials perspective, saliva case studies are particularly informative because they expose the end‐to‐end coupling between plasmonic architecture, biointerface interactions, and downstream machine‐learning robustness under real clinical variability [2, 4, 26]. Accordingly, we summarize representative Au‐SERS‐AI implementations spanning oncology, viral screening, and periodontal inflammation, focusing on what the substrate enables and what the model can reliably separate. Table 5 benchmarks reported performance and cohort sizes, while Figure 5 highlights two common translational workflows and their clinical sampling logic.

TABLE 5.

Benchmark of AI–SERS Platforms in Saliva Diagnostics (Case Studies).

Disease/Condition	Platform/Substrate	Analyte/Biomarker	AI model	Sample size (n)	Accuracy/Sensitivity/Specificity	Validation‐type	Refs.
Oral Cancer	Portable AuNP paper SERS	Saliva proteins	SVM	50	92%/90%/91%	Internal cross‐validation (single‐site cohort)	[43]
Lung Cancer (Early)	3D plasmonic hexaplex paper (Au@Ag)	Proteins/metabolites	Logistic Regression	75	91%/80%/—	Internal train–test split (single‐center cohort)	[24]
COPD	Aluminum‐foil enhanced Raman/SERS	Global saliva fingerprint	PCA‐HCA‐LDA	30	>98%/>98%/>98%	Contextual non‐Au pilot; internal leave‐one‐out validation	[5]
Childhood Asthma	Gold‐nanorod saliva SERS	Inflammatory, protein and lipid fingerprint	PCA‐LDA	44	84%/85%/82%	Core Au‐SERS chemometric study; cross‐sectional internal validation	[11]
COVID‐19	Label‐free Au SERS	Viral proteins	RF/SVM	175	95%/94%/89%	Internal cross‐validation (single‐site cohort)	[9]
Periodontitis	Raman spectroscopy (saliva)	Salivary biomarkers	SVM	30	90%/—/—	Internal validation (limited cohort)	[69]

Abbreviations and notes: SERS, surface‐enhanced Raman scattering; AuNP, gold nanoparticle; Au@Ag, gold‐silver core–shell nanostructure; SVM, support vector machine; RF, random forest. Accuracy, proportion of correctly classified samples; sensitivity, true positive rate; specificity, true negative rate. Label‐free indicates detection without external Raman reporter molecules. Portable denotes handheld or point‐of‐care Raman devices suitable for clinical screening.

FIGURE 5.

Representative case studies demonstrating the translational potential of bioactive plasmonic SERS platforms combined with artificial intelligence for saliva‐based diagnostics. (a) Label‐free detection of early‐stage lung cancer using a three‐dimensional plasmonic hexaplex paper (3D‐PHP) sensor integrated with a saliva collection tube. Salivary biomolecules interact with the engineered plasmonic substrate to generate enhanced Raman fingerprints, which are subsequently analyzed using machine‐learning algorithms for patient classification across cancer stages. (b) Non‐invasive discrimination of healthy, COVID‐19–infected, and recovered individuals using gold nanoparticle–assisted SERS of saliva coupled with AI‐driven spectral analysis, enabling accurate, label‐free classification based on disease‐associated Raman signatures. Panels adapted with permission from [67] and [72], respectively.

Note: Most reported performances derive from internal validation within single‐site cohorts; independent multi‐center external validation remains limited in current saliva‐based Au–SERS–AI studies.

5.1. Oncology‐Focused Saliva Diagnostics

5.1.1. Oral Cancer

Salivary diagnostics have long been investigated for oral malignancies due to the direct contact between saliva and diseased mucosa [7, 28]. Recent advances in portable AuNP‐based SERS substrates coupled with machine learning have strengthened this approach. Kaushik et al. demonstrated a portable gold nanoparticle paper SERS platform capable of detecting salivary protein signatures associated with oral pre‐malignant lesions and carcinomas [43]. Using a SVM classifier on spectra from 50 subjects, the system achieved 92% accuracy, 90% sensitivity, and 91% specificity, highlighting its suitability for point‐of‐care screening (Table 5). These findings align with broader efforts to deploy saliva‐based optical diagnostics in community and low‐resource clinical settings [40].

More recently, Puravankara et al. evaluated an AuNP plasmonic droplet assay in 114 participants (38 normal, 37 premalignant, and 39 malignant). Using a portable 785 nm Raman system and a Nu‐SVC classifier, the study reported 88.9% accuracy, sensitivity, and specificity for malignant vs. normal classification; normal vs. premalignant discrimination reached 82.4% accuracy, 66.7% sensitivity, and 100% specificity. The real saliva spectra and representative assignments are shown in Figure 2 and Table 2. Although the cohort is larger than many earlier demonstrations, the 75/25 single‐center split still requires independent replication [10].

TABLE 2.

Representative vibrational assignments for the saliva spectra shown in Figure 2. (Puravankara et al., 2026).

Raman shift (cm−1)	Representative assignment	Interpretive relevance
524	S‐S stretching, lysozyme	Protein/innate‐defense contribution
730	Tryptophan ring mode	Aromatic amino‐acid/protein contribution
852	Tyrosine ring breathing	Aromatic amino‐acid metabolism
1000	Phenylalanine symmetric ring breathing	Phe/Tyr ratio and protein metabolism
1258	Amide III, C‐N stretching	Protein backbone/secondary structure
1301–1305	CH2 twisting and wagging	Lipoprotein and lipid contribution
1358	CH3/CH2 wagging	Collagen and lipid‐associated contribution
1449	C‐H deformation/glycoprotein‐associated mode	Protein/lipid composite signal
1645	Amide I, C═O stretching	Protein secondary structure

5.1.2. Early‐Stage Lung Cancer

Beyond oral disease, saliva has proven informative for systemic malignancies such as lung cancer. Linh et al. developed a three‐dimensional plasmonic hexaplex paper (3D‐PHP) sensor composed of Au@Ag nanostructures integrated directly into a saliva collection tube [24]. This label‐free Au‐SERS platform captured composite spectral fingerprints from salivary proteins and metabolites, which were subsequently analyzed using logistic regression. In a cohort of 75 individuals, the model differentiated benign from malignant early‐stage lung cancer with 91% sensitivity and 80% specificity (Table 5). As illustrated in Figure 5a, this approach highlights how metabolite‐linked Raman features can encode disease‐relevant biochemical alterations, offering a non‐invasive complement to conventional imaging‐based screening strategies.

5.2. Infectious Disease Applications

5.2.1. COVID‐19 and Viral Screening

The COVID‐19 pandemic accelerated the development of rapid, non‐invasive diagnostic platforms, positioning saliva‐based Au‐SERS‐AI systems as compelling alternatives to nucleic‐acid testing. Karunakaran et al. reported a label‐free gold SERS saliva platform combined with AI classifiers to discriminate healthy, infected, and recovered individuals using Raman spectral fingerprints of viral and host‐response proteins. Using SVM and random forest (RF) models on saliva samples (n = 175), the system achieved up to 95% accuracy, 94% sensitivity, and 89% specificity, as summarized in Table 5.

Complementary work by Szymborski et al. confirmed the robustness of this approach, showing that RF classifiers delivered the best performance for saliva‐based SARS‐CoV‐2 detection, even with limited clinical metadata [53]. In parallel, Ember et al. demonstrated a portable Raman device with on‐board machine learning capable of real‐time COVID‐19 classification from liquid saliva samples [52]. Together, these studies, illustrated schematically in Figure 5b, establish saliva‐based Au‐SERS‐AI as a rapid, reagent‐free, and scalable strategy for viral screening and patient monitoring.

5.3. Inflammatory and Periodontal Diseases

5.3.1. Point‐of‐Care Relevance

Inflammatory oral diseases such as periodontitis represent a clinically important but often underdiagnosed category where rapid saliva‐based screening could significantly improve patient outcomes. Villalba‐Hernández et al. employed Au‐SERS combined with SVM analysis to detect salivary enzymatic biomarkers associated with periodontitis [69]. Despite a modest sample size (n = 30), the approach achieved 90% classification accuracy (Table 5), demonstrating proof‐of‐concept for chairside diagnostics. These findings reinforce the broader applicability of Au‐SERS‐AI platforms for inflammatory conditions, particularly where conventional diagnostics rely on subjective clinical assessment or invasive sampling.

5.4. Pulmonary Diseases and Respiratory Infections

5.4.1. Direct Saliva Evidence: COPD and Asthma

COPD provides direct proof that a global salivary vibrational fingerprint can encode chronic pulmonary disease. In a pilot study of 15 patients with COPD and 15 healthy controls, saliva deposited on aluminum foil for signal enhancement was analyzed by SERS and multivariate statistics. PCA and hierarchical clustering separated the groups, and an LDA model reported accuracy, sensitivity, and specificity above 98%. The result is promising but should be interpreted cautiously because the cohort was small, validation was internal, and the enhancement substrate was not gold; it is therefore contextual rather than core Au‐SERS‐AI evidence [5].

Asthma offers stronger alignment with the gold‐focused scope. Zamora‐Mendoza et al. mixed saliva with gold nanorods and acquired SERS spectra at 785 nm in 44 children (26 with asthma and 18 healthy controls). Spectral differences across 760–1750 cm⁻¹ were associated with nucleic acids, proteins, amides, lipids, triglycerides, and an IL‐8‐associated band near 1326 cm⁻¹. PCA‐LDA yielded 84% accuracy, 85% sensitivity, and 82% specificity. Because inhaled corticosteroids, oral inflammation, age, and current respiratory infection can alter saliva, future models should prespecify these covariates and validate patient‐level performance in independent pediatric cohorts [11].

5.4.2. Lung Cancer and Pulmonary Fibrosis

Among pulmonary indications, early lung cancer currently provides the clearest direct saliva + gold‐containing SERS + computational‐classification example. The Au@Ag three‐dimensional paper platform discussed in Section 5.1 analyzed composite protein and metabolite fingerprints in 75 individuals and reported 91% sensitivity and 80% specificity. Its integration with a saliva‐collection tube also demonstrates how sample handling and plasmonic readout can be co‐designed. Nevertheless, independent multi‐center comparison against benign inflammatory lung disease, smoking‐related disease, and oral pathology is still required [24].

Pulmonary fibrosis remains an evidence gap rather than an established saliva‐SERS application. Candidate salivary markers include extracellular‐matrix and epithelial‐injury products such as collagen fragments, hydroxyproline, fibronectin, elastin‐related products, KL‐6, and matrix metalloproteinases. A 2026 review proposed an alveolus‐saliva axis but emphasized that paired bronchoalveolar lavage and saliva studies in the same patients, standardized reference ranges, and harmonized sampling protocols are still lacking. These chemically diverse targets are compatible in principle with multiplexed affinity‐SERS or global‐fingerprint approaches, but this is a prospective design opportunity, not a demonstrated clinical saliva + Au‐SERS‐AI case study [70, 71].

5.4.3. Respiratory Infections

Direct saliva + Au‐SERS‐AI evidence for tuberculosis was not identified in the case‐study search. However, adjacent respiratory‐sampling platforms demonstrate relevant materials concepts. Chen et al. developed a facemask incorporating Ag@Au nanoflower arrays and a catalytic/plasmonic Au‐Ag sensor to detect the Mycobacterium tuberculosis ESAT‐6/CFP‐10 antigen complex in respiratory droplets or sputum. The clinical study reported 76.5% sensitivity and 100% specificity among 17 positive and 12 negative patients. Because the specimen was a droplet or sputum, and the output was not an AI‐derived salivary fingerprint, this study should inform capture architecture and portability without being counted as saliva‐SERS‐AI evidence [72].

Respiratory‐infection screening is best developed for SARS‐CoV‐2. Direct saliva studies include gold‐assisted label‐free SERS with SVM/RF classification, SERS‐chemometric comparisons of saliva and nasopharyngeal swabs, and conventional liquid‐saliva Raman systems with on‐board machine learning. Adjacent breath analysis further shows the potential of receptor‐functionalized SERS: a handheld breathalyzer evaluated 501 participants and achieved greater than 95% sensitivity and specificity in under five minutes. Breath volatile‐organic‐compound sensing is not saliva analysis, but it demonstrates how multi‐receptor chemistry, portable optics, and prespecified multivariate models can be integrated for respiratory screening [9, 52, 73].

Collectively, pulmonary applications span three evidence levels: direct core saliva + gold‐containing SERS + computational classification (asthma, lung cancer, and selected COVID‐19 studies); contextual direct‐saliva spectroscopy using non‐gold SERS or conventional Raman methods (COPD and additional COVID‐19 studies); and adjacent respiratory specimens (tuberculosis droplets/sputum and COVID‐19 breath). Future work should recruit differential‐diagnosis cohorts, separate participants at the patient level, control smoking and oral inflammatory status, report substrate batch effects, and include external testing before claiming disease specificity.

6. Practical Considerations for Clinical Translation

Collectively, the case studies reviewed indicate that saliva‐based Au‐SERS‐AI platforms can reach clinically meaningful diagnostic performance across diverse disease indications, despite heterogeneity in substrates, analytes, and computational pipelines [12, 43, 45, 52, 74]. However, successful clinical translation depends less on peak classification accuracy than on the reproducibility of plasmonic substrates, the control of saliva pre‐analytics and spectral quality, and the transparency and robustness of artificial intelligence workflows [16, 17, 19, 23, 74]. These factors represent the principal points at which laboratory demonstrations either mature into deployable technologies or fail to translate beyond proof‐of‐concept studies [1, 2, 23]

6.1. Substrate Reproducibility and Standardization

6.1.1. Nanoparticle Morphology

Plasmonic enhancement in Au‐based SERS substrates is inherently sensitive to nanoparticle morphology, size distribution, and interparticle spacing. Variations in nanostructure geometry alter local electromagnetic fields and can introduce substantial signal variability, even when the biological sample remains unchanged. For clinical translation, studies should therefore report nanoparticle morphology distributions using TEM or SEM, rather than representative images alone, and include quantitative descriptors of dispersion and aggregation state to support reproducibility and inter‐study comparison [14, 15, 20].

6.1.2. Enhancement Factor Reporting

In parallel, enhancement factor (EF) reporting must move toward standardized and transparent practices. Absolute or relative EF values should be reported alongside batch‐to‐batch variability and predefined acceptance windows, because EF instability can masquerade as biology‐driven spectral differences. Without this information, it remains difficult to distinguish true disease‐related signals from substrate‐driven artifacts, limiting the interpretability of downstream AI performance claims [14, 15, 21, 23, 74].

6.2. Saliva Pre‐Analytics and Spectral Quality Control

6.2.1. Collection Protocols

As previously mentioned, saliva is a dynamic biofluid whose secretion and composition vary with circadian rhythm, hydration status, and behavioral factors, while collection and handling steps can further shift measurable proteomic and metabolomic profiles. Consequently, inconsistent pre‐analytical handling can dominate spectral variability and obscure disease‐related signatures. Minimum reporting standards should include time‐of‐day, fasting status, mouth‐rinse protocol, and post‐collection processing steps such as centrifugation or filtration, ideally consolidated into a concise SOP to improve comparability. Where applicable, protein depletion or solid‐phase extraction methods should be specified, as they influence which biochemical fractions dominate the spectral signal [2, 8, 16, 19, 25, 29].

6.2.2. Internal Standards

Spectral quality control is equally critical. The inclusion of internal standards and explicit reporting of acquisition parameters support monitoring of laser stability, drift, and preprocessing impacts across measurements. Reporting laser wavelength and power, acquisition time, cosmic‐ray rejection thresholds, and baseline correction strategies is essential to evaluate data quality and to judge whether model performance is likely to generalize. These measures are particularly important in saliva‐based SERS, where weak signals from low‐abundance biomarkers must be separated from background fluctuations and technical noise [16, 19, 39].

6.3. AI Transparency, Validation, and Regulatory Alignment

6.3.1. Model Reporting

Artificial intelligence methods have substantially improved the interpretability of complex SERS datasets and mitigated challenges associated with spectral overlap, sample heterogeneity, and nonlinear patterns. However, for clinical relevance, AI models must be accompanied by transparent reporting of preprocessing, train‐validation‐test splits, performance metrics beyond accuracy, and clear statements of intended use, limitations, and failure modes. Such reporting practices are increasingly viewed as necessary to evaluate robustness and reduce overfitting in spectroscopy‐based diagnostics [16, 17, 19].

6.3.2. Calibration Transfer

Given the sensitivity of Raman and SERS spectra to instrument response, substrate properties, and acquisition conditions, calibration transfer and domain adaptation strategies are particularly relevant for any point‐of‐care pathway. Transfer learning and decision‐fusion approaches have been proposed and demonstrated in spectroscopic contexts to preserve performance across acquisition domains, supporting a shift away from narrowly optimized single‐site models. Preference should therefore be given to calibration‐free or transferable models designed to tolerate laser and instrument variability, rather than pipelines that require complete retraining with each device change [16, 19, 64].

6.3.3. Human Factors and Cost

Finally, clinical translation requires consideration of human factors and economic constraints, including ease of sample handling, workflow simplicity, operator training burden, and total cost of ownership. Recent saliva‐based demonstrations using portable or near point‐of‐care workflows emphasize that usability and operational feasibility are central to real‐world adoption, not merely analytical performance. In parallel, advances in paper‐based Au‐SERS substrates and simplified sensor architectures suggest plausible pathways toward lower‐cost deployment, reinforcing the importance of reporting per‐test cost and device economics in translational studies [23, 24, 43, 52, 75].

7. Challenges and Future Directions

Saliva‐based Au‐SERS‐AI diagnostics are advancing from elegant spectroscopy toward clinically oriented decision tools, but success will be determined by a small set of coupled material and computational constraints. The priorities summarized in Table 6 span signal physics, biointerfaces, data infrastructure, validation, and manufacturability [13, 16, 18, 23].

TABLE 6.

Key Challenges and Future Research Priorities.

Challenge	Impact	Potential Solution	Example Approach/Reference
Signal reproducibility (hot spots)	Variability in EF	Template‐grown Au substrates	[13, 14, 23]
Protein corona formation	Alters binding/spectra	PEGylation, zwitterionic ligands	[22, 44, 49, 74]
Small datasets for AI	Model overfitting	Transfer/federated learning	[16, 17, 64, 76]
Lack of standardization	Poor comparability	Shared saliva spectral database	[1, 2, 77]
Device cost	Limits adoption	Paper SERS + handheld Raman	[24, 43, 52, 78]

Abbreviations and notes: EF, enhancement factor in surface‐enhanced Raman scattering (SERS). Hot spots denote nanoscale regions of intensified electromagnetic fields responsible for signal amplification. Protein corona refers to the adsorption of biomolecules onto nanoparticle surfaces, altering spectral signatures. PEGylation indicates surface modification with polyethylene glycol to reduce nonspecific adsorption; zwitterionic ligands provide charge‐neutral antifouling coatings. Transfer learning and federated learning denote AI training strategies that improve performance with limited datasets. Standardization refers to harmonized acquisition protocols, substrates, and preprocessing across devices. Paper SERS indicates low‐cost cellulose‐based plasmonic substrates suitable for point‐of‐care testing.

7.1. Materials‐Level Challenges

7.1.1. Hot‐Spot Variability

A defining strength of SERS is also its core weakness. Signal amplification is dominated by highly localized electromagnetic hot spots, so small variations in nanogap geometry, aggregation state, or analyte placement can produce large run‐to‐run changes in enhancement factor and spectral appearance [13, 15]. This variability is especially consequential for label‐free biofluids, where competing constituents can stochastically occupy the most enhancing regions and distort what the model interprets as “disease signal.” Consistent with Table 6, one forward direction is to prioritize architectures engineered for reproducible hot‐spot statistics rather than maximal enhancement, including template‐guided or otherwise standardized Au substrates designed to reduce geometric uncertainty [13, 23]. This materials‐first emphasis is essential if adaptive diagnostic workflow is to remain stable under routine clinical handling.

7.1.2. Protein Corona Effects

In saliva, nanoparticle surfaces do not remain “clean.” Proteins and other biomolecules adsorb rapidly and can reshape surface chemistry, binding preferences, and the effective distance between analyte and plasmonic surface, shifting both intensity and fingerprint features. The practical result is that nominally identical substrates can behave differently across donors or collection conditions. Biointerface engineering is therefore not optional but central: surface chemistries that reduce nonspecific adsorption and stabilize analyte proximity can improve spectral consistency and enhance interpretability. As highlighted in Table 6, coatings such as PEGylation and other anti‐fouling ligand strategies are frequently proposed to mitigate corona‐driven spectral drift and improve operational robustness [44, 49, 79]. More broadly, future Au‐SERS platforms intended for saliva will benefit from being designed explicitly as bioactive nano‐bio interfaces, where the surface ligand layer is treated as a functional component of the sensing system rather than a post hoc stabilizer [18, 79].

7.2. Computational and Data Challenges

7.2.1. Dataset Scarcity and Generalization

Although AI can accelerate spectral interpretation, many reported Au–SERS studies remain limited by small cohorts, narrow inclusion criteria, and site‐specific acquisition conditions, all of which increase the risk of overfitting and inflated performance estimates. This is a central barrier to credibility for clinical deployment and is directly captured in Table 5 as “small datasets for AI.” Transfer learning and domain‐aware validation strategies have been positioned as practical routes to improve robustness when data are scarce, or devices differ [16, 64]. In parallel, federated learning is frequently discussed as a way to train across institutions without sharing raw patient data, offering an appealing alignment with clinical governance, provided the approach is framed carefully and evaluated under realistic cross‐site shifts [18, 76]. These strategies are most valuable when paired with transparent reporting practices and well‐defined intended use cases, as emphasized in ML guidance for biomedical audiences [17].

7.2.2. Explainability and Scientific Interpretability

For saliva‐based cancer diagnostics and broader disease screening, trust will depend not only on predictive accuracy but also on whether models provide mechanistic or at least spectroscopic plausibility. Explainability methods that highlight influential wavenumber regions can help connect model decisions to biochemical changes, reduce the likelihood that predictions hinge on acquisition artifacts, and facilitate clinical acceptance. Reviews focused on AI for Raman/SERS and data‐driven spectral discovery emphasize that interpretability and rigorous validation are becoming essential elements of credible deployment [16, 18, 19]. Importantly, the “bidirectional feedback loop” should be understood not only as an engineering concept but also as an interpretability opportunity: adaptive systems can be designed to flag uncertainty, request repeat sampling, or trigger confirmatory analysis pathways when spectral evidence is weak or inconsistent [16, 17].

7.3. Integration Opportunities

7.3.1. Microfluidics and Controlled Sample Handling

Integration with microfluidics offers a route to improve reproducibility by controlling pre‐concentration, separation, and delivery of analytes to plasmonic regions. For saliva, microfluidic handling can also reduce operator variability and enable standardized workflows compatible with point‐of‐care environments. These advantages align naturally with the framework, where stable input spectra are a prerequisite for meaningful AI refinement. Integration is also compatible with high‐throughput formats such as microwell arrays, enabling parallelized sampling while reducing the dependence on manual handling [80].

7.3.2. Edge AI for Rapid Decision Support

Portable or near‐patient systems benefit from moving inference closer to the device. Edge AI reduces latency and can enable near real‐time decision support, consistent with the vision of rapid turnaround emphasized in many AI‐enabled sensing frameworks [16, 75]. The most compelling near‐term direction is not purely “connected diagnostics,” but rather robust systems that can operate under constrained clinical conditions while documenting uncertainty and maintaining traceability of preprocessing and model decisions [17, 19].

7.3.3. Roll‐To‐Roll Gold Substrates and Cost Compression

Finally, manufacturability and cost are likely to determine whether saliva‐based Au‐SERS moves beyond specialist centers. Table 5 highlights device cost as a major adoption barrier, pointing toward paper‐based SERS and portable Raman‐like readouts as a pragmatic pathway. Recent work in portable, saliva‐facing diagnostic concepts reinforces that cost and workflow simplicity must be treated as design requirements rather than downstream considerations [78]. In this context, scalable fabrication approaches such as roll‐to‐roll Au substrates, coupled with simplified optics and automated analysis, could shift Au‐SERS‐AI from a laboratory technology toward community‐clinic deployment, aligning the materials platform with the public health promise of accessible early detection [23, 78].

8. Conclusions

Gold‐based plasmonic nanostructures have evolved from signal amplifiers into bioactive interfaces that govern molecular accessibility, protein corona composition, and hot‐spot statistics in saliva‐based Raman diagnostics. Surface‐enhanced Raman spectroscopy should therefore be viewed not only as an optical enhancement strategy, but as a nano‐bioengineering platform in which substrate geometry and surface chemistry determine which biochemical populations are reproducibly sampled.

Artificial intelligence serves as a translation layer that decodes material‐encoded spectral fingerprints into clinically interpretable outputs. When nano–bio interfaces are controlled and reproducible, models learn biology; when they are unstable, models learn artifacts. Clinical viability thus depends on coherent co‐design of plasmonic substrates, saliva pre‐analytics, and transparent computational pipelines.

With advances in gold substrate manufacturability, antifouling interface engineering, standardized reporting metrics, and transferable AI frameworks, Au‐SERS platforms are moving closer to practical deployment. Sustained emphasis on interface stability and dataset rigor will determine whether bioactive plasmonic nanostructures and artificial intelligence mature into a reproducible, point‐of‐care saliva diagnostic technology.

Author Contributions

Ajitesh Dhal: conceptualization, investigation, methodology, visualization, Writing – original draft. Ana Elena Aviña: conceptualization, methodology, investigation, writing – review & editing. Nguyen Le Thanh Hang: investigation, writing – review & editing. Peng Peiwen: supervision. Cheng‐Jen Chang: supervision. Tzu‐Sen Yang: conceptualization, supervision, writing – review & editing.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: smll75720‐sup‐0001‐SuppMat.docx.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.