Label-free saliva screening platform using M13 bacteriophage-based 3D plasmonic structures for MRONJ diagnosis

Abstract

Medication-related osteonecrosis of the jaw (MRONJ) is a severe complication associated with antiresorptive or antiangiogenic agents, often leading to pain, infection, and reduced quality of life. Current imaging-based diagnostics have limitations in detecting lesions smaller than 10 mm. In this study, we propose a label-free saliva screening approach for MRONJ diagnosis using a three-dimensional plasmonic structure based on M13 bacteriophage. Raman spectroscopy was employed to detect metabolite alterations in saliva, which are known to be associated with MRONJ. The M13 bacteriophage facilitates controlled interparticle gap of gold nanoparticles, thereby increasing hotspot density and enhancing Raman signal intensity. Data preprocessing was conducted on saliva Raman spectra collected from MRONJ patients and controls. To filter outliers, we computed Pearson correlation coefficients between each spectra and the group mean and excluded those with coefficients lower than 0.9. A total of 90 spectra were classified using an optimized multi-layer perceptron model, yielding a specificity of 84.6%, sensitivity of 100.0%, and an AUC of 0.92. This study demonstrates the potential of a saliva-based, non-invasive MRONJ screening strategy. Subsequent research should expand clinical datasets and investigate broader diagnostic applications.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-40342-8.

Introduction

Medication-related osteonecrosis of the jaw (MRONJ) is a multifactorial complication associated with the use of antiresorptive or antiangiogenic medications. In addition to the direct effects of these medications, local and systemic factors (e.g., chemotherapy, corticosteroid use, and comorbidities) increase the risk of MRONJ development^1–4. More than 20 years have passed since MRONJ was first reported in 2003; however, its exact pathophysiological mechanism remains unclear^2,5. MRONJ can lead to jawbone infections, chronic pain, tooth loss, and functional impairment, significantly affecting a patients’ quality of life⁶. Notably, the incidence of MRONJ tends to correlate with medication dosage, and patients with cancer receiving high dose antiresorptive or antiangiogenic therapy are at a particularly high risk. Studies have reported that up to 15% of patients with cancer receiving these medications may develop MRONJ because of a combination of patient-related factors and medication exposure, making them more vulnerable than the general population⁷.

The early detection and treatment of MRONJ are essential for improving patient outcomes Aggressive surgical treatments, such as marginal or segmental osteotomy, are prioritized in the intermediate to late stages of the disease, whereas conservative therapy is generally preferred in the early stages^3,8–11. Diagnosis of MRONJ requires a comprehensive evaluation of patient symptoms, clinical findings, and radiographic imaging³. However, there are no established guidelines for the optimal imaging modality to identify patients at high risk of clinical MRONJ, making accurate evaluation challenging¹². Particularly in MRONJ stage 0, where bone exposure is absent, radiographic diagnosis is feasible but may yield nonspecific findings, increasing the risk of false-positive diagnoses^1,13. In addition, the multidisciplinary nature of MRONJ management complicates screening and diagnosis. Although antiresorptive and antiangiogenic medications are primarily prescribed by physicians, MRONJ diagnosis and treatment are often managed by dentists, potentially leading to patient inconvenience and discontinuity of care.

Recent studies have suggested an association between metabolic changes in saliva and MRONJ^14,15. For instance, Thumbigere-Math et al. reported the differential expression of approximately 200 proteins in bisphosphonate-related osteonecrosis of the jaw (BRONJ) patients¹⁴. However, Yatsuoka et al. identified only a few clinically relevant biomarkers, including metalloproteinase-9 and desmoplakin, from a pool of 200 candidates^14,15. This discrepancy in biomarker identification is primarily attributed to the detection limits of mass spectrometry (MS) instruments. Sample purification and concentration steps are necessary for the MS-based analysis of biofluids such as serum and saliva, and variability in sample processing and analytical platforms makes it challenging to establish consistent biomarker profiles. Owing to these limitations, the development of a novel diagnostic approach with high reproducibility and minimal preprocessing is essential for saliva-based screening of MRONJ^16,17.

Surface-enhanced Raman scattering (SERS) has been proposed as an alternative to mass spectrometry, and it is a spectroscopic method capable of analyzing metabolites in biological samples without labeling. In disease diagnosis studies incorporating machine learning algorithms, SERS has been shown to effectively distinguish between control and patient groups by detecting multi-metabolite pattern changes^18,19. However, building a high-performance machine learning model for Raman spectral data requires the elimination of various noises, including thermal and electronic noise, through quantitative spectral analysis. Unlike blood, which is regulated by strong homeostatic mechanisms, the composition of saliva is primarily controlled by salivary gland secretory activity and is significantly influenced by physiological conditions such as dehydration, stress, diet, and hormonal fluctuations. These variations in salivary composition may introduce outliers into the dataset, potentially reducing model accuracy. Therefore, noise removal and outlier identification are critical for Raman spectral analysis of saliva-based MRONJ screening. In this study, we developed a data preprocessing algorithm by implementing noise reduction techniques and an outlier removal algorithm based on the Pearson correlation coefficient to enhance the reliability of saliva-based Raman spectroscopy for MRONJ screening.

Result and discussion

Three-dimensional plasmonic structure

In this study, the three-dimensional(3D) plasmonic structures for MRONJ screening was fabricated with metabolite ink, as illustrated in Fig. 1a. The metabolite ink was a mixture of 80 nm gold nanoparticles (Au NPs), M13 bacteriophage and saliva, which was deposited using the meniscus dragging deposition (MDD) technique. The MDD utilizes capillary forces and solution evaporation at the substrate surface to drag the meniscus, enabling the uniform deposition of nanoparticles and biomaterials from a colloidal solution onto the substrate, as shown in Fig. 1b. Figure 1c,d show the Raman spectra of the control group and the MRONJ patient group, respectively.

Fig. 1.

(a) Schematic of the preparation of metabolite ink by mixing M13 bacteriophage, gold nanoparticles (Au NPs), and saliva. (b) Fabrication of a 3D plasmonic structure on a SiO₂ substrate using the meniscus dragging deposition (MDD) technique. The SEM image shows the nanoscale morphology of the 3D plasmonic structure, along with a conceptual diagram illustrating Raman signal enhancement through the plasmonic structure. (c) Raman spectra obtained from saliva samples of the control group. (d) Raman spectra obtained from saliva samples of the medication-related osteonecrosis of the jaw (MRONJ) patient group.

In the 3D plasmonic structures developed in this study, Au NPs and M13 bacteriophage play crucial roles in signal enhancement and structural formation. Au NPs serve two primary roles. The first role is to maximize signal enhancement through localized surface plasmon resonance (LSPR), which induces strong electromagnetic fields at the plasmonic metal–dielectric interface. The intense electric field concentrated between adjacent Au NPs creates hot spots, where Raman scattering signals from target molecules are substantially amplified. To assess gap-dependent signal enhancement, finite-difference time-domain (FDTD) simulations were performed using 80 nm Au NPs with M13 bacteriophage (Figure S1). The simulations reveal pronounced electromagnetic enhancement at 20–40 nm gaps, with signal intensity progressively diminishing at larger gaps (60–80 nm). Importantly, electromagnetic enhancement persists even at 80 nm gaps, enabling sensitive detection of both small biomolecules (metabolites, genomes) and larger biomolecules (proteins in saliva)^18,19. The second role is to ensure structural stability and reproducibility through their exceptional chemical stability and oxidation resistance, maintaining consistent SERS performance over extended periods.

The M13 bacteriophage plays a critical role in optimizing the fabrication of the 3D plasmonic structure by preventing Au NPs aggregation. The functional groups of amino acids on the surface of the M13 bacteriophage are known to interact electrostatically with Au NPs, forming stable bonds. This interaction prevents the aggregation of Au NPs, thereby increasing the density of the hotspot regions²⁰. To quantitatively analyze the effect of M13 bacteriophage concentration on the interparticle spacing of Au NPs, SEM analysis, interparticle distance quantification, and Raman spectral measurements were performed on the fabricated substrates. Figure S2(a)–(h) show SEM images of the different 3D plasmonic structures fabricated under varying bacteriophage concentrations and saliva conditions. Analysis of the interparticle distance histogram and Raman spectra revealed distinct differences depending on the structure conditions. In the structure composed solely of Au NPs, excessive particle agglomeration prevented the formation of effective hot spots, resulting in no significant Raman signal (Figure S2(i) and (m)). Conversely, in the structure composed of Au NPs and saliva, regions with appropriate particle spacing for hot spots formation were simultaneously observed, along with regions where excessive particle agglomeration suppressed hot spots (Figure S2(j) and (n)). When M13 bacteriophage was added at a concentration of 1 mg/mL (Figure S2(k) and (o)), a uniform interparticle spacing was formed across the substrate, resulting in a stable, evenly distributed hot spots. Conversely, when the M13 bacteriophage concentration was increased to 10 mg/mL (Figure S2(l) and (p)), the interparticle spacing increased excessively, weakening plasmonic coupling and inhibiting effective hot spots. Figure S3 shows that the presence of saliva results in much more complex spectral features than when Au NPs and M13 phage are present alone.

The M13 bacteriophage (4E-type) used in this study expresses glutamic acid at the pVIII peptide, resulting in a higher zeta potential compared to the wild-type strain, as analyzed through dynamic light scattering (DLS) measurements (Table S1). Table shows that the 4E-type had a higher zeta potential than the wild-type. Additionally, glutamic acid contains carboxyl (-COOH) functional groups, which enhance binding efficiency with Au NPs²¹.

Raman spectroscopy and data preprocessing

Raman spectroscopy enables both quantitative and qualitative analysis of liquid biological samples such as saliva and serum with high reproducibility. Recently, the integration of Raman spectroscopy with machine learning has demonstrated potential applications in medical diagnostics, environmental monitoring, food quality inspection, and new material development^18,19,22,23 However, to train machine learning models using Raman spectra, preprocessing must be performed to improve the signal-to-noise ratio (SNR) and to reliably select high-quality data.

In this study, we implemented noise reduction processes to eliminate spikes caused by cosmic rays and pixel defects. Additionally, we removed random noise originating from laser intensity fluctuations, CCD thermal noise, and electronic noise. Figure 2a illustrates an example of spike noise removal, where impulsive artifacts originating from cosmic rays and detector defects are effectively suppressed while preserving the overall spectral profile. The black line shows the processed spectrum after spike removal, and the red line indicates the detected spike components. Figure 2b demonstrates the effect of spectral smoothing for random noise reduction, where high-frequency noise is substantially suppressed while the main Raman features are well preserved, as highlighted in the magnified region (1020–1380 cm⁻¹). Figure 2c shows the normalization results, where the spectra were divided by the peak intensity at 2500 cm⁻¹, a spectral range known to be relatively unaffected by complex metabolite variations in saliva²⁴.

Fig. 2.

(a) Spike removal process for Raman spectra. (b) Random noise removal process, including a comparison between the original and removed data. (c) Normalization of the Raman spectra based on the 2500 cm⁻¹ peak, which has little relation metabolite signals, to standardize the spectral intensities. (d) Flowchart of the data selection algorithm based on Pearson correlation. (e) Histogram of Pearson correlation coefficients for the control group, showing the removal of 18 outlier data points based on the correlation threshold of 0.90, with approximately 71.9% of the data retained after filtering. (f) Histogram of Pearson correlation coefficients for the MRONJ patient group, showing the removal of 3 outlier data points based on the correlation coefficients of 0.90, with approximately 93.8% of the data retained after filtering.

After data preprocessing, a data selection algorithm based on the Pearson correlation coefficient was applied to remove outlier saliva Ramans spectra due to diet, oral medication intake, etc. This outlier removal procedure was performed once on the entire dataset prior to data splitting, in order to exclude severely corrupted or non-representative spectra before model development. The Pearson correlation coefficient is a constant that quantifies the linear relationship between two variables. The calculation formula is given in Eq. (1), and Fig. 2d shows a flowchart of the data selection algorithm.

1

Here, R is the Pearson correlation coefficient, X_i denotes the intensity at the i-th Raman shift of an individual sample spectrum, and Y_i is the corresponding intensity of the mean spectrum calculated from either the control group or the MRONJ patient group. and represent the average intensities of each spectrum and the group mean spectrum²⁵.

To determine the optimal R value for the data selection process, we evaluated the area under the curve (AUC) of the MLP in the range of 0.80 to 0.95 with R values in increments of 0.05. Figure S4 presents the receiver operating characteristic (ROC) curves for different R values, showing the impact of the R value on model performance. R = 0.90, which resulted in the highest AUC of 0.859. Spectra with R value less than 0.9, compared to the mean spectrum of each group (control and MRONJ patient), were excluded from the machine learning dataset. Figure 2e illustrates that 18 data with a correlation coefficient below 0.90 were excluded from the control. Before removing outlier data, the machine learning classification achieved a specificity of 69.2% and a sensitivity of 80.0%, as presented in Figure S5.

Figure 3a shows a cone-beam computed tomography (CBCT) image of the MRONJ patient group. The study was performed by labeling MRONJ patient and control groups based on imaging diagnostics using clinical symptoms, medical history, and imaging modalities, including panoramic radiographs and CBCT. The clinical photograph and panoramic X-ray image of the MRONJ lesion are shown in Figure S6. The inset in Fig. 3b shows the measurement locations, which are arranged in a 3 × 3 array with a spacing of 750 μm. Figure 3b shows a 3D plot of the Raman spectra obtained from these positions, showing the signal consistency across different measurement points. The 1127 cm⁻¹ peak was selected for reproducibility evaluation because this peak is mainly associated with urea in saliva, which is a widely present and relatively uniformly distributed component in salivary samples. The Raman assignments and their corresponding molecular origins are summarized in Table S2. Figure 3c shows the Raman intensity at 1127 cm⁻¹ peak from nine detection points. The relative standard deviation (RSD) was 2.0% and this result demonstrates high reproducibility. In addition, three independent substrates were fabricated using the same saliva sample, and the substrate-to-substrate reproducibility was evaluated. As shown in Figure S7, the inter-substrate RSD was 2.7%, further confirming the excellent reproducibility of the proposed SERS platform. Figure 3d shows the mean Raman spectra and standard deviations of 64 controls and Fig. 3e shows 48 patients with MRONJ. Raman spectra from 48 MRONJ patients and 64 controls are shown in Figure S8. The control group RSD of the peak intensity at 1127 cm⁻¹ was 11.2% and the MRONJ patient group was 12.5%. These results were obtained before excluding the outlier data. After removing outlier data, the RSD of the control group was reduced by 10.2%, and that of the MRONJ patient group was reduced by 12.1%.

Fig. 3.

(a) Coronal and sagittal views of cone-beam computed tomography (CBCT) from a patient with MRONJ stage III, showing osteonecrotic lesions (white arrows). (b) Schematic of the nine-points Raman measurement process on the sample surface (inset), with a 3D plot displaying the Raman spectra obtained from the nine detection points. (c) Signal repeatability and sample uniformity were assessed based on Raman spectra obtained from nine detection points and the relative standard deviation (RSD) of the peak intensity at 1127 cm⁻¹ was 2.0%. (d, e) Mean Raman spectra and standard deviations of the control group and MRONJ patient group. (f) Comparison of train and test accuracy for different machine learning models (SVM, 1D-CNN, and MLP).

Development of a classification model for MRONJ screening

To enhance the classification accuracy of the Raman spectral data using machine learning, feature extraction and model optimization were performed. Although the Raman spectrum contains signals ranging from 100 to 3600 cm⁻¹, selecting a specific wavelength range is crucial for improving analytical efficiency and eliminating unnecessary noise. In this study, we utilized the Raman spectra range of 500 to 1750 cm⁻¹, as it encompasses characteristic peaks primarily associated with biomolecular vibrational modes of nucleic acids, proteins, and carbohydrates, which provide discriminative information for learning differences between MRONJ patient and control groups²⁴.

To improve the generalization performance of the machine learning model and to avoid data leakage between training and evaluation, the dataset was split at the patient level rather than at the individual spectrum level. All Raman spectra obtained from the same subject were assigned exclusively to a single subset. Specifically, 60.0% of the subjects were allocated to the training set, 20.0% to the validation set for hyperparameter tuning, and the remaining 20.0% to the independent test set for final performance evaluation. This patient-level partitioning strategy ensures that no spectra from the same individual appear in more than one subset, thereby preventing overly optimistic performance estimation and providing a more realistic assessment of the model’s clinical generalization capability.

Three machine learning models known for their strong performance in Raman spectra classification were tested: support vector machine (SVM), multi-layer perceptron (MLP), and 1D-convolutional neural network (1D-CNN). The classification accuracy results, as shown in Fig. 3f, the classification accuracy results were 57.7% for SVM, 72.4% for 1D-CNN, and 73.7% for MLP, indicating that the MLP achieved the highest diagnostic performance. Given its lower computational cost and higher accuracy compared to 1D-CNN, we selected MLP as the final machine learning model for MRONJ classification.

To prevent overfitting, we applied several regularization techniques to the MLP model. Batch normalization was used in the fully connected layers to normalize the mean and variance of the minibatches, thereby accelerating the training process. L2 regularization was implemented by adding the squared sum of the weight values to the loss function, thereby reducing model complexity and improving generalization. Additionally, dropout was applied to randomly deactivate a subset of neurons during training to prevent the model from over-relying on specific nodes and further enhance the generalization performance.

To further improve model accuracy, Bayesian optimization was employed for hyperparameter tuning. Bayesian optimization is a probabilistic approach that optimizes hyperparameters more efficiently than traditional grids or random searches. The hyperparameter search focused on tuning the number of layers in the fully connected layer, the number of nodes per layer, the regularization strength (λ) for L2 regularization, and the dropout rate. The optimal hyperparameters, as shown in Fig. 4a, were determined to be three layers with 1001–987-1013 nodes, respectively, an L2 regularization strength (λ) of 0.099, and a dropout rate of 0.23.

Fig. 4.

(a) Schematic of the multi-layer perceptron (MLP) architecture used for MRONJ classification based on Raman spectra. The preprocessing steps include spike elimination, random noise elimination, normalization, and data selection based on the Pearson correlation coefficient (R > 0.9). The dataset is divided into a training set (60.0%), validation set (20.0%), and test set (20.0%). (b) Receiver operating characteristic (ROC) curve for the MRONJ classifier, showing an area under the curve (AUC) of 0.92, indicating high classification performance. (c) Confusion matrix of the test set, demonstrating the classifier’s performance with a sensitivity of 100% (95% CI 88.0–100%) for MRONJ patient group and a specificity of 84.6% (95% CI 65.6–98.4%) for the control group.

Figure 4b shows the ROC curve of the model trained with the optimized hyperparameters. The AUC value was 0.92, indicating a high classification performance in distinguishing MRONJ patient group from the control groups. The confusion matrix in Fig. 4c shows a specificity of 84.6% (95% CI 65.6–98.4%) and a sensitivity of 100% (95% CI 88.0–100%), demonstrating that the model effectively differentiates MRONJ patients from healthy controls with high sensitivity. In addition, to assess the overall classification performance on the entire dataset including the test set, we performed a tenfold cross-validation. In this setting, the model achieved an average specificity of 75.5% and an average sensitivity of 89.0%. These results indicate that, due to the limited cohort size, the estimated performance metrics remain sensitive to small changes in the dataset, and that the inclusion or exclusion of a small number of samples can substantially affect the results. This highlights the need for larger-scale clinical validation to obtain more stable and statistically robust performance estimates in future studies.

Materials and methods

Study population

This study included patients diagnosed with MRONJ, who presented with jaw necrosis and associated pain and underwent surgery at the Department of Oral and Maxillofacial Surgery, Pusan National University Dental Hospital, between March 2023 and September 2023. MRONJ was diagnosed based on the American association of oral and maxillofacial surgeons (AAOMS) criteria. Preoperative and postoperative follow-up specimens were collected until February 2024. Patients with a history of cancer treatment or concurrent infections were excluded. The control group comprised healthy patients without concurrent inflammatory conditions or benign or malignant diseases of the oral cavity. Table S3 presents the demographics of the study population including age, sex, MRONJ stage, and implant status. This study was approved by the Institutional Review Board (IRB) of Pusan National University Dental Hospital (IRB No: 2022-10-014). All research procedures were conducted in accordance with relevant ethical guidelines and regulations. Written informed consent was obtained from all participants prior to sample collection.

Saliva sample collection and preparation

Unstimulated saliva samples were collected using Dry Saliva Collection Kits (NEST Biotechnology, China). Prior to sample collection, all participants were instructed to rinse their mouth with water to minimize transient oral contaminants. At least 500 μL of saliva was collected from each subject. After collection, the saliva samples were aliquoted into 100 μL portions. The aliquots were temporarily stored at 4 °C immediately after collection and then transferred to a − 80 °C deep freezer for long-term storage within 24 h. Before SERS measurements, frozen samples were thawed at 4 °C for approximately 3 h and then used immediately for experiments. To minimize degradation and compositional changes, each aliquot was subjected to only a single freeze–thaw cycle.

Plasmonic structures

In this study, we used 80 nm Au NPs (in 0.02 M sodium citrate, NanoComposix) at a concentration of 1 mg/ml for the 3D plasmonic structure. Centrifugation was performed to increase the concentration of the Au NPs solution to 40 mg/ml. The concentrated AuNP solution was then mixed with the 4E-type M13 bacteriophage solution (1 mg/ml) and saliva samples in a volume ratio of 5:2:3(Au NPs:M13 bacteriophage:saliva).

The plasmonic structure was fabricated on a 290 nm SiO₂/Si substrate (base substate). In this study, we used a thickness of 525 ± 15 μm of base substrate, a resistivity of 1.0–10.0 Ω·cm, and an orientation of < 100 >. For all experiments, the base substrate was cut into dimensions of 1.0 × 0.5 cm. The diced substrates were cleaned using an ultrasonic (SD-D200H, Sungdong) in 99% acetone, 99% ethanol, and 99% isopropyl alcohol for 10 min each. The cleaned substrates were dried by N₂ gas blower and treated to surface using an oxygen plasma (Covance, Femto Science).

After treating the surface, the substrates were mounted onto a one-axis motorized stage (DDS220/M, Thorlabs). A doctor blade (JP/SA-204, Hohsen) was positioned above the motorized stage, maintaining a 500 μm gap between the base substrate and the doctor blade. The mixed solution was dropped on between the base substrate and the doctor blade. The motorized stage was moved at a speed of 0.005 mm/s to enable the coating process. A dedicated software program was used to control the motorized stage for precise movement during the coating process.

Raman spectroscopy

Raman spectroscopy measurements were performed using a Raman spectrometer (NS200 custom, Nanoscope Systems) equipped with a 633 nm wavelength laser and an objective lens (CF Plan 20 × , Nikon). Spectral data were collected within the 100 cm⁻¹ to 3600 cm⁻¹ range, which includes the characteristic Raman signals of proteins, lipids, nucleic acids, and metabolites.

Finite-difference time-domain (FDTD) simulations

The electric field intensity and spatial distribution are critically dependent on multiple parameters, including laser wavelength, nanoparticle geometry, inter-particle spacing, surrounding dielectric environment, and spatial arrangement. To comprehensively elucidate the electromagnetic enhancement mechanism at practically relevant gap sizes, FDTD simulations were conducted at 20, 40, 60, and 80 nm inter-particle gaps. FDTD simulations employed a Total-Field/Scattered-Field (TFSF) source to rigorously separate incident and scattered wave contributions. Polarization states at 0° and 90° were independently evaluated to characterize the anisotropic optical response of the coupled nanostructure system. All simulations were performed at 633 nm wavelength to maintain consistency with experimental Raman measurements. The computational domain incorporated M13 bacteriophage (refractive index n = 1.50) modeled at 1 nm spatial resolution. Perfectly matched layer (PML) boundary conditions were applied to minimize numerical artifacts and reflections. Gold nanoparticles with 80 nm diameter were modeled as the primary coupling elements within the plasmonic structure.

FE-SEM

The morphological characteristics of the 3D plasmonic structure were analyzed by field-emission scanning electron microscopy (FE-SEM; JSM-7900F, Olympus). Additionally, an energy-dispersive X-ray spectroscopy (EDS) module (Ultim Max, Oxford) attached to an electron microscope was used to examine the elemental composition of the plasmonic substrate.

Dynamic light scattering analysis

The hydrodynamic size and zeta potential of the M13 bacteriophage solution were measured using a Zetasizer Nano ZS90 (Malvern Panalytical, UK). The measurements were conducted at 25 °C with a scattering angle of 90°. The data were analyzed using the Zetasizer software provided by Malvern Panalytical.

Data preprocessing and machine learning

Raw Raman spectra were preprocessed to improve the signal-to-noise ratio (SNR) and to ensure robust input data quality for machine learning analysis. First, spike noise caused by cosmic rays and pixel defects was removed using a three-sigma rule-based algorithm. Briefly, a smoothed spectrum was generated, and the noise component was estimated as the difference between the raw spectrum and the smoothed spectrum. Data points exceeding the 99.7% confidence interval (± 3σ) of the noise component were identified as spike artifacts and removed. To further suppress high-frequency random noise originating from laser intensity fluctuations, CCD thermal noise, and electronic noise, a Savitzky–Golay filter with a window size of 15 points was applied. Finally, intensity normalization was performed to compensate for signal intensity variations among samples. Each spectrum was normalized by dividing by the signal intensity at 2500 cm⁻¹, which is a spectral region relatively insensitive to biochemical variations in saliva and is commonly used as a reference baseline.

The machine learning experiments were conducted in a local computing environment equipped with an NVIDIA GeForce RTX 3080 GPU (12 GB GDDR6X), an AMD Ryzen 5 3600 CPU, and 48 GB of RAM. The software environment included Windows 10, Python 3.9.9, TensorFlow 2.18, CUDA 11.2, and cuDNN 8.1, enabling graphics processing unit (GPU) acceleration. All computations, model training, and evaluations were performed using a GPU-supported TensorFlow backend.

For Bayesian optimization-based hyperparameter tuning, the search space was defined by considering both the model complexity and the dataset characteristics, with the number of layers ranging from 2 to 5, the number of nodes per layer ranging from 256 to 1024, the L2 regularization strength (λ) ranging from 0.0001 to 0.01, and the dropout rate ranging from 0.1 to 0.5. To determine the optimal combination, five initial random samples were evaluated, followed by 30 search iterations. The model was trained using the Adam optimizer with an initial learning rate of 0.000005 and a batch size of 8. Binary cross-entropy was used as the loss function for the classification. To prevent overfitting, early stopping was applied based on the validation loss with a patience of 10 epochs. The model was trained for over 300 epochs.

Conclusions

In this study, we proposed a saliva-based 3D plasmonic structure combined with Raman spectral preprocessing and machine learning classification for noninvasive MRONJ screening. MRONJ is a multifactorial complication associated with the use of antiangiogenic or antiresorptive medications, and the 2022 AAOMS diagnostic guidelines rely on clinical examination and radiological diagnosis³. However, radiology-based methods often yield nonspecific findings in the early stages, making accurate diagnosis challenging. The multidisciplinary nature of MRONJ management further contributes to patient inconvenience and care discontinuity. Consequently, there is an increasing demand for new diagnostic technologies with higher sensitivity and accessibility.

In this study, we developed a high-accuracy classification model capable of distinguishing MRONJ patient group from control group by utilizing SERS to perform quantitative and qualitative chemical analyses of metabolites in saliva. The three-dimensional plasmonic structure was fabricated by uniformly coating on base substrate with metabolite ink, a mixture of 80 nm Au NPs, M13 bacteriophage (4E-type) and saliva, using the MDD technique.

Data preprocessing was performed on the acquired Raman spectra to enhance classification accuracy. Spike noise and random noise removal algorithms were applied. A data selection algorithm based on the Pearson correlation coefficient was employed to exclude outliers, thereby improving model classification performance.

Among the three machine learning models (1D-CNN, SVM and MLP) evaluated for MRONJ classification, the MLP demonstrated the best performance, achieving an accuracy of 73.7%. We further optimized the MLP hyperparameters using Bayesian optimization, and the final optimized model demonstrated a specificity of 84.6% (95% CI 65.6–98.4%), sensitivity of 100% (95% CI 88.0–100%), and AUC of 0.92, effectively classifying MRONJ patient group from control group.

Since the present cohort is dominated by stage 2 MRONJ cases (33 out of 48 patients), the reported diagnostic performance mainly reflects classification performance in intermediate-stage disease. In addition, the control and MRONJ groups were not fully matched in terms of age and sex distribution, which represents an important limitation of this study. Furthermore, the Pearson-correlation-based data selection step, while useful for removing severely corrupted spectra, may introduce selection bias and therefore should be applied with caution in future studies.

Despite these limitations, future studies using large-scale and demographically better-matched clinical cohorts will be necessary to further evaluate the generalization performance of the model and to validate its clinical applicability. The saliva-based, high-accuracy model with a 3D plasmonic structure developed in this study based on Raman spectral noise reduction, data selection, and machine learning optimization may be applicable not only to MRONJ classification but also to a wide range of saliva-based disease diagnostic applications.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, Y.K., J.-J.K., M.J., T.-Y.J. and J.-W.O.; Saliva sample collection, J.-J.K. and S.-H.S.; Formal analysis, S.H., Y.J., N.-Y. K. and G.-H. B.; Investigation, T.K., H.L. and Y.L.; Writing—original draft preparation, Y.K., J.-J.K., M.J., T.-Y.J. and J.-W.O.; Writing—review and editing, Y.K., J.-J.K., M.J., T.-Y.J. and J.-W.O; All authors reviewed the manuscript.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2024-00406152), This work was also supported by the 2024 Yeungnam University Research, This work was also supported by the Technology Development Project of the Ministry of SMEs and Startups in 2023 (RS-2023-00262093), and This work was also supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education(RS-2022-NR074932).

Data availability

The datasets analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.