Potential use of saliva infrared spectra and machine learning for a minimally invasive screening test for congenital syphilis in infants

Abstract

Congenital syphilis is a global public health issue, and its diagnostic complexity poses a challenge to early treatment. Fourier Transform Infrared Spectroscopy (FTIR) is a promising technological tool that facilitates the detection and diagnosis of various diseases by providing information on the biochemical composition of biofluids, including saliva. However, the potential use of FTIR in congenital syphilis has not yet been studied. This study aimed to explore the development of a method for the diagnosis of congenital syphilis using saliva FTIR spectra and machine learning algorithms in infants aged 0 to 12 months. First, the potential of FTIR for analyzing infant saliva was evaluated. Spectral analysis revealed subtle differences in vibrational modes between the test and control groups. Complementary analyses, such as Principal Component Analysis (PCA) and Leave-One-Out Cross-Validation (LOOCV), were used to assess the variance among samples, which enabled efficient discrimination and highlighted the relevance of the observed variance between groups. When applying Quadratic Standard Normal Variate preprocessing with LOOCV, the model achieved 90% accuracy, 100% sensitivity, and 80% specificity. Therefore, the method demonstrated potential as a screening test for congenital syphilis. The study’s limitations include a reduced sample size and the reliance on the data upsampling approach.

Introduction

Syphilis, caused by the bacterium Treponema pallidum, is one of the most prevalent sexually transmitted diseases worldwide¹. In situations where an infected pregnant woman does not receive immediate and effective treatment, there is a risk of vertical transmission of the pathogen to the fetus or newborn, leading to congenital syphilis². Despite the World Health Organization (WHO)’s commitment nearly two decades ago to eliminate the infection, the global burden of congenital syphilis has persisted³. In 2022, the WHO projected that there were 700,000 cases of congenital syphilis worldwide. Maternal syphilis was responsible for approximately 150,000 early fetal losses and stillbirths, and 70,000 neonatal deaths⁴. Challenges in documenting cases and the complexity of early diagnosis likely result in the underreporting of these estimates².

In survivors, congenital syphilis can have catastrophic effects on child development, including hearing loss, hydrocephalus, brain calcifications, and congenital malformations resulting from injury to the central nervous system. However, neurological manifestations may represent only the most severe end of a broad spectrum of damage caused by the disease. Newborns may be asymptomatic⁵, and in cases where clinical symptoms are present, they are often subtle or non-specific, which makes early diagnosis challenging⁶.

Currently, there are no available diagnostic tests specifically targeting congenital syphilis⁴. Its diagnosis requires a combination of clinical, epidemiological, and laboratory evaluations. Signs and symptoms, if present, are identified through a thorough physical examination, along with the maternal history of treatment, and follow-up. Serial non-treponemal tests, in comparison to maternal results, and additional tests—including cerebrospinal fluid analysis and radiology—also need to be considered². In the absence of obvious signs and symptoms, complementary laboratory tests using blood samples become crucial for determining the final diagnosis⁶. However, the first-choice test for syphilis screening, the Venereal Disease Research Laboratory (VDRL) test, has low specificity, requiring a second confirmatory test when results are positive. This makes the diagnostic process more challenging and costly for public health systems, particularly in resource-limited settings^2,3, and highlights the urgent need for effective solutions to address this global problem.

There is growing interest in investigating specific immunological biomarkers for T. pallidum to contribute to a fast and reliable diagnosis of congenital syphilis⁷. Salivary diagnostics, in particular, have evolved as promising early screening tools. Saliva consists of a diverse range of biomolecules (including water, electrolytes, lipids, carbohydrates, and proteins) from the salivary glands, cervical gingival fluid, nasal and bronchial secretions, as well as food debris and microorganisms, which may serve as potential biomarkers for various diseases^8,9. Saliva’s advantages over other body fluids include its non-invasive nature, ease of collection, low storage cost, and potential for large-scale use. This makes it an effective approach for rapid and reliable diagnostics¹⁰.

In congenital syphilis, T. pallidum spirochetes are only present in saliva when there are lesions in the mouth⁶, which poses a problem for conventional laboratory testing. In fact, there are no reports in the literature regarding the use of saliva for diagnosing congenital syphilis or identifying syphilis markers in this biofluid. However, saliva biomarkers have been studied for other diseases, including malignant neoplasms¹¹, viral diseases like HIV, type 2 diabetes, risk of dental caries¹², Sjögren’s syndrome¹³, and Alzheimer’s disease^8,14. Despite this, research in this field is challenging due to the influence of multiple factors—such as natural conditions, human diseases, and exogenous causes—that affect the composition and flow of saliva⁸. Consequently, progress may be hindered by the lack of efficient methods and techniques to differentiate molecular biomarkers in saliva¹⁵.

One of the most promising methods for providing information on the molecular composition of biological fluids is the Fourier Transform Infrared Spectroscopy (FTIR), combined with machine learning models. This optical technology has proven to be an economical, rapid, and non-invasive means of identifying diseases in biological samples by effectively distinguishing between groups with and without certain conditions¹⁶. Studies have shown that FTIR can identify molecular changes associated with cancer, degenerative, hematological, endocrine, oral, and infectious diseases, among others^17–20. The application of FTIR to saliva samples has been explored for several diseases, including esophageal cancer¹⁹, COVID-19²¹, diabetes²², and periodontitis⁹. The accuracy of this technique with saliva has been reported to be nearly 90% or higher, which demonstrates its high precision and feasibility for identifying diseases. However, no studies have investigated the potential use of FTIR for the early diagnosis of congenital syphilis.

Furthermore, despite the potential of infrared spectroscopy to overcome the limitations of sample collection and storage in tissue biopsies and conventional biochemical tests, further research, standardization of techniques, and technological advancements are still needed for its accurate and effective application¹⁶. As a non-invasive technique requiring minimal sample volume, little to no sample preparation, and no expensive reagents, FTIR is emerging as a promising tool for point-of-care applications, with the potential to facilitate early diagnosis and clinical monitoring^9,15,23. Given the need for rapid, non-invasive, cost-effective diagnostic solutions for congenital syphilis that can be scaled for widespread use^3,4,6, exploring the application of FTIR for its diagnosis represents a valuable step in advancing potential solutions to these challenges.

In this proof-of-concept study, we explored the development of a method for the diagnosis of congenital syphilis in infants aged 0 to 12 months, using saliva FTIR spectra and machine learning. Specifically, we explored the accuracy of FTIR, combined with machine learning models, to identify biomarkers of congenital syphilis in the saliva of infants diagnosed with the disease, comparing them to a control group. The results offer pioneering insights and demonstrate promising potential for rapid, non-invasive, and cost-effective screening of congenital syphilis in this population.

Methods

Study design

This is an exploratory observational study with a prospective and cross-sectional design. The study was carried out under the ethical principles of the Declaration of Helsinki, and was approved by the Research Ethics Committee of the Federal University of Mato Grosso do Sul (protocol number 71599223.1.0000.0021).

Participants

Participants were recruited at the Maria Aparecida Pedrossian University Hospital (HUMAP), a reference center for infectious diseases in Mato Grosso do Sul, Brazil. Participants were selected through convenience sampling. Therefore, the sample size was determined by including all infants born and/or hospitalized for the treatment of congenital syphilis at HUMAP between August 2023 and July 2024. The guardians of the infants signed an informed consent form authorizing the infants’ participation in the study.

Initially, 100 infants (38.5 ± 2.3 weeks gestation) from both sexes (51.0% girls) were included and divided into two groups (or classes): congenital syphilis group (CS) and control group (CG). The inclusion criteria for the CS group were live-born infants, aged 0 to 12 months, clinically stable, born or admitted for treatment of congenital syphilis at HUMAP. The inclusion criteria for the CG group were live-born infants, aged 0 to 12 months, clinically stable, who were hospitalized for standard care after birth or admitted for treatment of common conditions other than congenital infection. Exclusion criteria for the CS group included infants who did not have a confirmed diagnosis of congenital syphilis or who were outside the established age range.

Procedures

Saliva collection

For the saliva collection, the infant’s mother was instructed not to feed the infant milk or any other food within the past 30 min, unless the infant was feeding on demand. First, a gentle massage was applied to the infant’s cheeks to stimulate salivation. A commercial rayon sterile swab was then used to firmly but gently rub the entire oral cavity. The swab was placed in a commercial plastic falcon tube, which was then transported and stored at − 20 °C. Each tube was labeled only with a number corresponding to the order of participation and group name.

Sample processing and FTIR analysis

Saliva samples were thawed at room temperature prior to FTIR analysis. To facilitate saliva extraction, 40 µL of 0.9% saline solution was added into the plastic tube, followed by manual compression. The 40 µL of saline was determined during preliminary tests as the optimal volume to promote saliva release without causing excessive dilution of salivary biomolecules. A drop of the resulting solution was then placed directly on the crystal of the Attenuated Total Reflectance (ATR) accessory of an Agilent Cary 630 spectrometer. FTIR measurements were taken in the range of 1800 to 1300 cm⁻¹, 8 cm⁻¹ resolution, and 64 scans, with deionized water used as the background²⁴.

Initially, spectra from samples with insufficient volume for adequate measurements were discarded. Saliva collection from newborns presented challenges due to the limited quantity of saliva secreted at this age. This led to discarding some samples before analysis due to insufficient volume.

First, FTIR spectra were normalized using the Standard Normal Variate (SNV) method to standardize the data²⁵. To reduce spectral noise, the Inverse Fourier Transform (IFFT) was applied with a cutoff frequency of 10 Hz. This frequency was sufficient to attenuate random noise while preserving relevant information from the vibrational bands. The IFFT enables manipulation of the spectrum in the frequency domain, allowing undesirable components to be eliminated before reconversion to the time domain²⁶.

After the preprocessing, 14 samples from the CS group and 13 samples from the CG group remained. Upsampling was performed to balance the number of samples per group. In this process, two random spectra from each group were selected and averaged to generate an additional spectrum. This procedure was repeated until each group had 20 spectra²⁷.

Principal Component Analysis (PCA) was applied to reduce the dimensionality of the data and eliminate redundancies while preserving as much variance as possible. PCA transforms the original data into new orthogonal variables (principal components), derived from the decomposition of the eigenvalues and eigenvectors of the covariance matrix²⁸ The loadings, which represent the contribution of the original variables to each component, were analyzed to identify functional groups that contributed the most to the separation between groups. A visual analysis of the first three principal components was performed to assess the potential for group separation.

For model training and validation, several classifiers were tested. The classifier selected for optimal performance was the Support Vector Machine (SVM) with quadratic optimization (Quadratic SVM), a machine learning technique focused on classification. The Quadratic SVM seeks to identify the hyperplane that best separates the groups by maximizing the margin between the closest points of each group, known as support vectors, thereby enhancing the model’s robustness for group discrimination²⁹.

Model validation was conducted using the leave-one-out cross-validation (LOOCV) technique, which is well-suited for small datasets. In LOOCV, each sample is used as a test case, while the remaining samples are used for training. This process is repeated N times (where N is the number of samples), maximizing the use of the available data³⁰. Finally, the performance of the model was assessed using a confusion matrix, which allowed the evaluation of accuracy and robustness of the developed protocol.

Results and discussion

A total of 27 saliva spectra were obtained, with 14 from the congenital syphilis group and 13 from the control group. The reduced sample size was primarily due to the difficulty in obtaining sufficient saliva volume from newborns, as infants typically do not secrete a large quantity of saliva before 2–3 months of age³¹. Even after attempts to dilute the samples with saline solution, some still showed an insufficient spectral signal for proper analysis and were discarded.

To balance the number of samples per group, an upsampling process was applied, which resulted in 20 samples for each group. For data standardization, the Standard Normal Variate (SNV) normalization was applied, followed by spectral smoothing IFFT with a cutoff frequency of 10 Hz. Additionally, the spectral range was restricted to 1800–1300 cm⁻¹ to reduce interference from milk molecules present in the samples due to infant feeding.

The FTIR average spectra for control and congenital syphilis, Fig. 1, exhibit the same vibrational bands within the spectral range analyzed for both groups. Subtle differences in the relative intensity of the minor bands are observed and suggest changes in the vibrational modes of the molecules in the sample. These changes may be associated with the presence of new molecules in the samples, such as biomarkers from the immune response, or they could indicate physicochemical changes in the environment, including variations in pH and concentration^19,20,32. While these characteristics may not be sufficient to determine the diagnosis, they can contribute to group differentiation in a multivariate analysis study³³.

Fig. 1.

Average FTIR spectra for congenital syphilis (red line) and control (black line) groups. The spectra were normalized by SNV and smoothed by IFFT with a cutoff at 10 Hz. Shaded regions indicate the standard deviation of the spectral data corresponding to each class.

The major bands around 1640 and 1550 cm-1 are usually assigned to Amide I and Amide II groups from protein molecules, such as IgG (immunoglobulins)^17,18,31. The small variations in intensity observed in Fig. 1 are likely attributable to concentration differences between congenital syphilis and control saliva samples, rather than representing a diagnostic hallmark, as they exhibit similar relative intensities. The band at 1640 cm⁻¹ corresponds to C = C and C = O stretching vibrations, while the band at 1550 cm⁻¹ is associated with N–O stretching and N–H bending, primarily attributed to IgG proteins, which play a significant role in the immune response to syphilis, despite the spectral similarities observed in these bands³⁴.

The four minor bands observed in the range of 1500 to 1300 cm⁻¹, along with the 1750 cm⁻¹ band in the congenital syphilis group, appear to shift toward higher energy and exhibit subtle changes in relative intensity when compared to the FTIR spectra of the control group. The 1750 cm⁻¹ and 1400 cm⁻¹ band is attributed to C=O stretching, commonly associated with the Amide I group found in proteins, fatty acids, phospholipids, and lipids. The 1450 cm⁻¹ band is assigned to C-H scissoring bending, and COO⁻ symmetric stretching, usually related to Amide II in amino acids, proteins, and lipids. The small shoulder around 1375 cm⁻¹ can be assigned to COO⁻ symmetric stretching, CH₃ symmetric bending from amino acids and proteins, such as IgG, IgA, and IgM. Finally, the 1330 cm⁻¹ band can be attributed to C-N stretching, and O–H bending from Amide II groups^35,36. Some of these bands (1450, 1400, and 1375 cm⁻¹) may also be associated with cytokines, which are important in the immune response to syphilis³⁷.

In summary, the average FTIR spectra from both groups exhibit clear differences, as anticipated. FTIR spectroscopy provides detailed molecular vibrational information that reflects the composition of the samples. The immune response to syphilis is notably complex as it involves both innate and adaptive mechanisms that introduce a wide array of new molecules into the biological system. This has been demonstrated using saliva samples, which, despite their highly diluted nature, still contain a diverse variety of molecules^38,39. Syphilis significantly changes the molecular composition of saliva, as evidenced in the literature by the presence and quantification of T. pallidum DNA and associated proteins^40,41.

To emphasize the FTIR differences between the two groups, the spectral data from each sample were summarized using principal component analysis (PCA). The PC1 × PC2 × PC3 score plot, shown in Fig. 2, captures 88.8% of the data variance and illustrates the variance among individual samples, revealing distinct clusters. The separation between these clusters highlights the differences between the groups.

Fig. 2.

Score plot of the first three principal components (PCs) for congenital syphilis (blue circles) and control (red circles) samples. The distance between the points reflects spectral variance—greater separation indicates higher dissimilarity between the spectra.

Another notable observation involves group variance: the control group exhibits greater dispersion across the 3D plot, likely reflecting the high variability in innate characteristics among individuals. While these individuals are not infected with T. pallidum, they may present other immune conditions. Conversely, the T. pallidum-infected group (congenital syphilis) forms a more cohesive and compact cluster, indicating lower variance compared to the control group.

To further investigate the correlation between the original FTIR spectra and the clustering separation, the loading plots for the first three principal components (PCs), shown in Fig. 3, highlight the key wavenumbers contributing to data variance. The loading plots highlight the contribution (weight) of each wavenumber (spectral variable) to the PCs used for dimensionality reduction in the overall dataset. These contributions can be either positive or negative across the spectral range. A detailed examination of the loading plots reveals that the seven vibrational bands identified in the FTIR spectra contribute significantly to the variance captured by the first three PCs. The highest variance is associated with the Amide I bands at 1640 and 1550 cm⁻¹. In addition, minor bands play a more prominent role in PC2 and PC3, with the 1700 cm⁻¹ band showing a particularly strong contribution to PC3. Collectively, these contributions are crucial for clustering the samples and differentiating the groups, as demonstrated in Fig. 2.

Fig. 3.

Loading plots of the first three principal components (PCs) for the full dataset, highlighting the most relevant vibrational bands contributing to spectral variance. The dotted line represents the original FTIR spectrum.

In addition to the separation trend observed between the groups in Fig. 2, the clusters do not exhibit a clear boundary, with some samples overlapping in the same region. To achieve more accurate sample classification, we applied machine learning algorithms based on SVM models. The SVM identifies a separating hyperplane between the two groups; based on the distribution in the score plot, we selected a quadratic function to define this boundary. The performance of the Quadratic SVM was evaluated using the LOOCV method. A grid search was conducted to optimize the number of PCs and evaluate various SVM classification options, including linear, quadratic, cubic, and Gaussian kernels⁴². Other classifiers, such as K-Nearest Neighbors (KNN), Linear and Quadratic Discriminant Analysis, and Decision Tree, were also tested; however, they demonstrated inferior performance. The configuration yielding the best classification performance during LOOCV was the quadratic SVM model using the first five PCs (Supplementary Table S1). Increasing the number of PCs as latent variables in the model resulted in a decline in accuracy, which suggests overfitting. The first five PCs were used as input variables for the SVM, which collectively accounted for 95.83% of the cumulative variance in the data. The model’s performance is summarized in the confusion matrix presented in Fig. 4.

Fig. 4.

Confusion matrix for training and validation using Leave-One-Out Cross-Validation (LOOCV) with the Quadratic SVM classifier. Correct classifications are shown along the diagonal (green), while misclassifications appear off-diagonal (red).

The results demonstrate 90% accuracy, 100% sensitivity, and 80% specificity, with a consistent error rate primarily due to false-positive diagnoses. This highlights the potential of this method as a screening tool. The approach proved robust by generating no false-negative diagnoses for syphilis and ensuring reliable identification of positive cases. Although four control group samples were misclassified as positive for syphilis, this is not a major concern. In a clinical setting, false-positive results can be further evaluated using reference techniques, whereas negative diagnoses provide a high degree of confidence. Therefore, this method shows promise as a rapid screening tool for syphilis with minimally invasive sample collection, aiding in the early identification of positive cases that warrant further clinical investigation.

Importantly, an additional exploratory analysis was performed using only the original 13 spectra from each class, without applying upsampling. The result was very similar, with the model misclassifying only two samples (controls incorrectly classified as syphilis-positive), which resulted in an accuracy of approximately 92%. This was expected as the upsampling process was performed randomly among the samples and reinforces the robustness of the observed spectral differences.

We acknowledge that the potential inclusion of conditions other than congenital syphilis in the control group contributed to increasing the sample variability. Future studies should avoid that by including fully healthy infants as controls. However, the Quadratic SVM model still reached high performance, which suggests that the presence of other conditions in the control group did not significantly affect the model’s ability to distinguish between the groups.

It is important to stress that, as a proof-of-concept study, our findings require further exploration in larger-scale experiments. Validation and improvement of the achievements of this study will be necessary. Additionally, we acknowledge the small sample size and data upsampling process as important limitations in this study that should be avoided in a future fully validated model. Future research should focus on expanding the sample size through multi-center studies for better generalizability, enhancing preprocessing techniques to improve signal quality, and testing alternative machine learning models to optimize classification performance. To minimize sample loss, future approaches should use more effective saliva stimulation techniques or alternative collection strategies for newborns. Although leave-one-out cross-validation was employed to minimize bias, the lack of a truly independent test set prevents full assessment of the model’s generalizability. These factors highlight that the results should be interpreted as an initial proof-of-concept rather than definitive clinical validation. Future studies with larger sample sizes should adopt independent validation sets or stratified k-fold cross-validation to enhance model robustness and prevent overfitting. Furthermore, independent validation with external datasets and exploring specific biomarkers via complementary methods like proteomics could strengthen the findings. Longitudinal studies should assess the stability of spectral biomarkers, while direct comparisons with current gold-standard diagnostic tests will establish clinical relevance. Developing user-friendly tools for real-time analysis, conducting cost–benefit analyses, and exploring FTIR’s potential for multi-disease screening will further support its integration into clinical practice. Further refinement of spectral preprocessing and machine learning strategies—such as avoiding upsampling, exploring ensemble or deep learning models, and developing biomarker-specific approaches—are also crucial steps to strengthen the clinical potential of saliva FTIR spectroscopy for congenital syphilis screening.

When compared to current diagnostic practices for congenital syphilis, which rely on a combination of clinical signs, serological tests (e.g., VDRL, RPR, and treponemal tests), and maternal history^2,4,6, our FTIR-based approach offers notable advantages: it is minimally invasive, rapid, and cost-effective. However, it currently lacks the diagnostic specificity and large-scale validation needed to replace traditional methods. Conventional serological tests, despite their limitations (e.g., false positives, need for confirmatory tests), remain the gold standard due to their extensive validation in clinical practice.

Conclusions

Differences in spectral patterns between the saliva samples of infants with congenital syphilis and controls suggest potential for group separation, despite the absence of evident specific markers. The complementary PCA technique corroborated the initial visual analysis and highlighted the variance between the samples. The loadings further emphasized the main vibrational bands contributing to the separation, which reinforced the observed differences between the groups. In the protocol validation, the results showed an accuracy of 90%, with a consistent error rate primarily in false-positive diagnoses. This ensured high reliability in identifying negative cases. A summary of the analysis and accuracy is illustrated in Fig. 5.

Fig. 5.

Illustrative summary of the study procedures, analysis, and accuracy result.

It can be concluded that the proposed protocol can be reliable based on the tests performed and may serve as a potential model for the rapid and cost-effective diagnosis of congenital syphilis. Further investigation in larger-scale studies, including validation and refinement of the current results, will be essential to support the clinical translation of this approach.

Author contributions

D.C.D: Sample acquisition; D.C.D., C.C.: Data acquisition; C.C and B.M.: Methodology, data analysis, validation; D.C.D, C.C, B.M., and D.S.M.: Writing and revision; D.A.S.-M.: Conceptualization, methodology, supervision.

Funding

Coordination for the Improvement of Higher Education Personnel, Brazil (CAPES). Federal University of Mato Grosso do Sul (UFMS/MEC – Brazil) [grant code 001]. CAPES/PROCAD [grand code 88881.516322/2020-01]. National Council for Scientific and Technological Development—CNPq [grant code 310376/2023-8].

Data availability statement

Data will be shared upon request, which must be made by contacting the corresponding author (D.S.M.).

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-05144-4.