A Simple Mouth Rinse and Machine Learning Reveal a Spit-Based Warning Sign of High Blood Sugar

A swish of fluid around the mouth, analyzed with a machine learning pipeline, may one day flag dangerously elevated blood sugar without a single needle prick. Researchers report in the journal Clinical Proteomics that a single protein recovered from oral rinse samples, called REG1A or Regenerating Islet-Derived Protein 1 Alpha, stands out as a non-invasive marker of elevated HbA1c, the long-term blood glucose measure clinicians rely on to diagnose and monitor type 2 diabetes. The study, drawn from the Dutch population cohort known as The Maastricht Study, suggests that the mouth may carry a readable chemical echo of what is happening in the pancreas and bloodstream, and that modern algorithms can tease that signal out of a complex soup of oral proteins.

The clinical logic behind the work is straightforward. HbA1c, or glycated hemoglobin, reflects the average blood glucose concentration over the preceding two to three months, because sugar molecules attach themselves to hemoglobin in red blood cells as they circulate. Measuring it requires a blood draw, and while that is routine in clinical settings, it is a barrier for large-scale screening, for point-of-care testing in resource-limited settings, and for repeated monitoring in epidemiological studies. Oral fluids have long tempted researchers as an alternative: they can be collected cheaply, safely and repeatedly, without trained phlebotomists or biohazard disposal. The challenge has always been sensitivity and specificity, because saliva and oral rinse fluid contain a diluted and highly variable mixture of proteins from saliva glands, gingival tissue, blood plasma leakage and the oral microbiome.

To find a usable signal in that mixture, the team, led by E. Stamatelou of the Academic Centre for Dentistry Amsterdam (ACTA) together with collaborators at Maastricht University, Philips, and Amsterdam UMC, designed a case-control study nested within The Maastricht Study, an extensively phenotyped population cohort. Participants were selected from the existing cohort and oral rinse samples were retrieved from the associated biobank, giving the researchers both the biological specimens and rich clinical data about each donor. In total, 176 participants were included, stratified into groups based on their blood HbA1c levels, allowing the researchers to ask a precise question: which proteins in the oral rinse differ between people with low and high HbA1c?

The proteomic readout relied on the Proximity Extension Assay, a technology commercialized by Olink that targets 92 cardiometabolic proteins simultaneously. In this assay, antibodies paired with DNA tags bind to their target proteins in the sample; when two antibodies bind the same protein, their DNA tags join and can be quantified by PCR-like readout, producing a normalized protein expression value for each target. This approach is well suited to small sample volumes and low-abundance analytes, which is exactly the situation in oral rinse fluid, where proteins of systemic origin are present at far lower concentrations than in blood plasma.

The machine learning design was deliberately conservative. The researchers applied a two-threshold strategy: first they split participants at the population median HbA1c of 37 mmol/mol, and then at the clinical cutoff of 48 mmol/mol, which is the diagnostic threshold for diabetes. Feature selection was performed with Boruta, an algorithm built around random forest significance testing that iteratively confirms which features carry real information beyond chance. Model evaluation used repeated nested cross-validation, a scheme in which feature selection happens strictly inside the inner folds so that no information from the held-out data leaks into training, a critical safeguard against the optimistic performance estimates that have plagued many biomarker studies. Three classifiers were compared: logistic regression, random forest, and XGBoost, a gradient-boosted tree method popular in tabular data competitions.

The results were telling in two ways. At the population median threshold, no proteins in the oral rinse could reliably discriminate between the lower and upper halves of the HbA1c distribution, a negative finding the authors report transparently. But at the clinically meaningful cutoff of 48 mmol/mol, the picture changed. Logistic regression, the simplest of the three models, achieved the best performance, with a receiver operating characteristic area under the curve of 0.78, plus or minus 0.09, and a 95 percent confidence interval of 0.74 to 0.82. That level of discrimination, while not yet sufficient for stand-alone diagnosis, is respectable for a screening marker derived from a non-invasive sample, and notably it came from the interpretable linear model rather than the more flexible tree-based ensembles, a common pattern in small biomedical datasets where overfitting threatens complex learners.

Within that model, one protein dominated: REG1A, also known as Pancreatic Stone Protein. To interpret which features drove the predictions and in which direction, the team used SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each feature a contribution value for every individual prediction. The SHAP analysis confirmed that REG1A was the sole oral biomarker distinguishing elevated HbA1c, with higher oral levels associated with higher blood sugar status. The biological plausibility is striking: REG1A is a protein linked to the regeneration of pancreatic beta cells, the insulin-producing cells that fail progressively in type 2 diabetes. Pancreatic Stone Protein has previously been studied mainly in the context of acute inflammation and sepsis, but its connection to pancreatic physiology makes it a compelling candidate to reflect metabolic stress in a fluid sampled far from the pancreas itself.

The authors did not stop at the classifier output. In post hoc multivariable analyses, they tested whether the association between oral REG1A and elevated HbA1c could be explained by confounders, most importantly periodontitis, the chronic inflammatory gum disease that is both common in diabetes and known to alter oral protein composition. After adjusting for potential confounding factors, REG1A remained significantly associated with elevated HbA1c. Moreover, REG1A levels showed significant positive trends across the clinically relevant categories of glycemia, rising stepwise from normal through prediabetes to diabetes. That graded relationship matters, because a useful screening marker should ideally track disease severity rather than merely separate two arbitrary groups, and it hints that oral rinse proteomics could in principle support staging as well as detection.

The implications reach beyond the laboratory. A validated oral rinse test for elevated HbA1c could be deployed at pharmacies, dental clinics, community health fairs or at home, feeding into point-of-care devices and removing one of the practical frictions that keeps many people from being screened for a disease that often goes undiagnosed for years. It could also be a gift to large epidemiological studies, which could collect oral samples at scale without the cost and logistics of blood draws. The researchers themselves are careful about the limits: the study involved 176 participants in a case-control design, the performance estimate carries a wide confidence interval, and the finding requires validation in larger and independent cohorts before any clinical use. The two-threshold result also implies the marker works best at distinguishing clinically elevated HbA1c rather than fine gradations within the normal range.

Still, the study is a notable proof of concept at the intersection of three trends: the maturation of multiplex proteomic assays sensitive enough for oral fluids, the adoption of rigorously cross-validated machine learning in biomarker discovery, and the growing recognition of the mouth as a diagnostic window onto systemic disease. The work was funded through the ORANGEForce project within the ORANGEHealth consortium under Health~Holland, with additional support from Dutch regional and institutional funders, and the underlying cohort study was approved by the Medical Ethical Committee of Maastricht University Medical Centre+ with written informed consent from all participants. If REG1A holds up in larger validation studies, the humble mouth rinse, paired with a modest logistic regression model, could become one of the simplest tools yet for catching elevated blood sugar before it does its silent damage.

Subject of Research: Identification of the protein REG1A in oral rinse samples as a machine learning-based non-invasive biomarker of elevated blood HbA1c levels

Article Title: REG1A as an oral biomarker for elevated blood levels of HbA1c: a machine learning approach in The Maastricht Study in a case control design

Article References: Stamatelou, E., Kosho, M. X. F., Gallucci, A., van der Kallen, C. J. H., van Greevenbroek, M. M. J., In ‘t Veld, S. G. J. G., Schut, M. C., Brouwers, M. C. G. J., Laine, M. L., & Loos, B. G. (2026). REG1A as an oral biomarker for elevated blood levels of HbA1c: a machine learning approach in The Maastricht Study in a case control design. Clinical Proteomics. https://doi.org/10.1186/s12014-026-09637-w

Image Credits: AI Generated

DOI: 10.1186/s12014-026-09637-w

Keywords: REG1A, HbA1c, type 2 diabetes, oral rinse, biomarker, machine learning, proteomics, Olink, SHAP, The Maastricht Study, pancreatic stone protein, nested cross-validation