The study encompassed individuals undergoing physical examinations at the Shanghai Gaohang Community Service Center between 2022 and 2023. Following the application of inclusion and exclusion criteria, 28 healthy controls, 30 individuals with Pre-DM, and 37 T2DM patients were selected. Informed consent, approved by the ethics committee of Shuguang Hospital, affiliated with Shanghai University of TCM, was obtained from all participants.

Participants were considered eligible for the Pre-DM group based on one or more of the following criteria: (1) FBG ranging from 5.6 to 6.9 mmol/L; (2) 2-hour postprandial blood glucose (2hPG) between 7.8 and 11.0 mmol/L; (3) Glycosylated hemoglobin (HbA1C) levels between 5.7% and 6.4%. For T2DM, inclusion required meeting one or more of the following: (1) random blood glucose ≥11.1 mmol/L; (2) FBG ≥7.0 mmol/L; (3) 2hPG ≥11.1 mmol/L; (4) HbA1C ≥6.5%.

The exclusion criteria included: (1) patients diagnosed with type 1 or specific types of DM, or those experiencing acute complications such as ketoacidosis; (2) recent use of antibiotics, probiotics, traditional Chinese medicines, or immunosuppressive drugs within the past two weeks; (3) coexisting oral conditions or complications, including periodontitis, pulpitis, or oral cancer; (4) individuals with severe systemic diseases, such as malignant tumors, immune disorders, or hematological diseases; (5) pregnant or breastfeeding women; (6) individuals unable to provide fully informed consent due to mental health symptoms, behavioral disorders, or cognitive impairments; (7) participants with distinct dietary habits; (8) participants with irregular bowel movements; (9) smokers.

The clinical characteristics of the enrolled patients were examined in this cross-sectional study, with data collected on demographic factors such as age, gender, body mass index (BMI), and waist-to-hip ratio (WHR). Personal lifestyle factors included smoking and drinking history, along with dietary preferences categorized as spicy, sweet, or no specific preference. The laboratory-developed TCM Clinical Diagnosis Record Form documented key symptoms, including dry mouth, bitter taste, constipation, and diarrhea. Blood pressure was assessed according to World Health Organization (WHO) criteria, classifying hypertension into grade 1 (140–159 mmHg systolic/90–99 mmHg diastolic), grade 2 (160–179 mmHg systolic/100–109 mmHg diastolic), and grade 3 (systolic ≥ 180 mmHg/diastolic ≥ 110 mmHg).

The tongue image acquisition process utilized the Tongue Diagnostic Instrument (TFDA-1), a device engineered by the Intelligent Diagnostic Laboratory at Shanghai University of Traditional Chinese Medicine, as depicted in Figure 1 . Detailed data collection procedures can be found in our earlier publication (Soper, 2021). Key technical specifications of the device included a manual operating mode, 1/125 shutter speed, F6.3 aperture setting, ISO sensitivity of 200, correlated color temperature ranging from 4500K to 7000K, and illumination at 4800 ± 10% (unit: lx).

Oral microbiota samples were obtained from the central region of the tongue dorsum using aseptic pharyngeal swabs, with at least ten rotational movements. Stool samples were self-collected by participants in the morning using a sterile fecal sampler, targeting the central portion of the stool. Following cryopreservation, the samples were immediately dispatched to researchers on the same day. Each specimen was sealed in sterile, enzyme-free Eppendorf tubes, kept on ice, and transferred to a −80°C freezer within 30 minutes for preservation prior to sequencing. Participants were instructed to fast before sample collection.

Tongue image features were extracted using the established methods previously developed by our research group. An intelligent quality assessment model was employed to screen all collected tongue images, ensuring they met the required quality standards (Jiang et al., 2021). Feature extraction was performed using the adversarial generation network, Tongue-GAN, as described in our earlier publications (Li et al., 2021b; Jiang et al., 2022).

Bacterial DNA from the tongue dorsum and fecal samples was extracted using a swab genomic DNA extraction kit (CW2654, CwBiotech, Beijing, China) and an intestinal DNA extraction kit (TIANamp Stool DNA Kit, DP328, Tiangen Biotech, Beijing, China), respectively. The 16S rDNA full-length assembly sequencing technology (16S-FAST) enabled species-level classification through analysis of bacterial ribosomal 16S RNA sequences, encompassing nine variable and ten conserved regions (Karst et al., 2018). For both qualitative and quantitative assessment, as well as quality control, 10 ng of DNA was utilized. Splice and link libraries were constructed, followed by data assembly from electrophoresis and Qubit concentration measurements to ensure quality before proceeding with sequencing on the Illumina NovaSeq 6000 platform (Illumina, USA). The methodology has been thoroughly detailed in a previous study by members of the research team (Guo et al., 2023).

A cloud platform (https://www.genescloud.cn/home) was employed for sequence analyses, utilizing QIIME2 (2019.4), R language (v3.2.0), the ggplot2 package, and Python. ASV-level alpha diversity, including Shannon diversity indices, was calculated via the ASV table in QIIME2 and visualized through box plots. To assess the significance of differences, the Kruskal-Wallis rank sum test followed by Dunn’s post-hoc test was applied. Ranked abundance curves at the ASV level were generated to evaluate richness and evenness across samples. Beta diversity analysis, leveraging UniFrac distance metrics, explored microbial community structural variation, visualized through principal coordinate analysis (PCoA) and hierarchical clustering. Microbial structure differentiation among groups was quantified using permutational multivariate analysis of variance (PERMANOVA) in QIIME2. Linear discriminant analysis effect size (LEfSe) analysis identified differentially abundant taxa among groups under default settings. Random forest analysis in QIIME2, with default parameters, was employed to classify samples from distinct groups, utilizing nested stratified k-fold cross-validation for automated hyperparameter optimization and sample prediction. Co-occurrence network analysis was conducted via SparCC, with a pseudo-count set to 106. Correlation coefficient cutoffs were established at 0.70 using random matrix theory-based methods in the R package RMThreshold, with Cytoscape (v3.9.0) constructing the network visualization. The R language facilitated analysis of the network’s topological structure, with key species identified through topological indices and visualized using the ZiPi plot. Phylogenetic Investigation of Communities by Reconstruction of Unobserved States (PICRUSt2) predicted microbial functions based on MetaCyc (https://metacyc.org/).

Logistic regression with backward selection was applied, incorporating L2 regularization, a tolerance of 1e-4, an inverse regularization strength (C) of 1.0, and the lbfgs solver, with a maximum of 100 iterations. Tongue image features, clinical indicators, and microbial data were screened for Pre-DM and T2DM classification, allowing the removal of insignificant variables while addressing multicollinearity (Chicco et al., 2021). Model fit was evaluated via maximum likelihood and the Hosmer-Lemeshow test. A combined tongue-microbiota classification model for Pre-DM and T2DM was subsequently developed and validated. To ensure robustness, 5-fold cross-validation was employed, partitioning data into five subsets and averaging performance metrics such as ROC AUC, accuracy, sensitivity, and specificity (Lin et al., 2020; Soper, 2021). Python 3.10.9 facilitated machine learning techniques aimed at capturing non-linear relationships, utilizing models like support vector machines (SVM), random forests (RF), gradient boosting, adaptive boosting (AdaBoost), and K-nearest neighbors (KNN). Classification results were calculated using the sklearn library (Version 1.3.1).

Data analysis utilized SPSS v. 25.0 (IBM Corp., Armonk, NY, USA). Variable distribution normality and variance homogeneity were assessed using the Shapiro–Wilk and Levene tests, respectively. For data meeting normal distribution and variance homogeneity criteria, a t-test was applied; otherwise, non-parametric methods were employed. Categorical variables were analyzed using Fisher’s exact test, while continuous variables were evaluated via the Wilcoxon rank-sum test. Relationships among independent variables were examined through Spearman’s rank correlation, with p-values adjusted for multiple comparisons using the Bonferroni correction.

Following the screening process, 28 healthy controls, 30 Pre-DM patients, and 37 T2DM patients were included in the study ( Figure 2 ), with baseline clinical characteristics detailed in Table 1 . The Pre-DM and T2DM groups exhibited older average ages compared to the control group, although no significant age difference was observed between the Pre-DM and T2DM groups. Hypertension was identified as a risk factor for diabetes, with a markedly higher prevalence of hypertension among T2DM patients. BMI and WHR measurements indicated greater obesity in both the Pre-DM and T2DM groups relative to the control group.

Following the extraction of tongue image features, three sets of crowd computer tongue image parameters were identified. Notably, a significant increase in preALL was observed in both the Pre-DM and T2DM groups, indicating a marked thickening and greasiness of the tongue coating. The Con* and MEAN* values for both the tongue body and coating in the T2DM group were substantially higher compared to the other groups. Furthermore, the ASM* and ENT* values for the tongue coating displayed significant variation when compared to the Control and Pre-DM groups. These results suggest a progressive transition in the tongue texture from smooth to rough, indicative of aging as diabetes progresses. Analyzing the color parameters of the tongue and coating, a gradual shift towards paler and whiter shades was noted, as shown in Table 2 and Figure 3 .

α-diversity analysis revealed a marked reduction in the richness and evenness of gut microbiota as diabetes progressed, contrasted by a significant rise in oral microbiota species abundance in both the pre-DM and T2DM groups ( Figures 4A, B ). This discrepancy may be attributed to the thickened tongue coating observed in these groups. PCoA analysis further demonstrated partial overlap between the pre-DM and control groups, while the T2DM group exhibited distinct shifts in community composition, highlighting notable alterations in microbial structure ( Figures 4C, D ).

Significant variations were observed in the oral and gut microbiota at both the phylum and genus levels ( Figure 5 ). In the oral microbiota, diabetes progression was associated with a marked increase in Firmicutes-C and a decrease in Fusobacteriota. Notably, the Pre-DM group exhibited a rise in Bacteroidota and a reduction in Actinobacteriota compared to the control and T2DM groups. In the gut microbiota, Bacteroidota levels were significantly elevated, while Firmicutes-A and Actinobacteriota showed substantial reductions in the T2DM group relative to other groups. The microbiota shifts in Pre-DM largely mirrored those in T2DM, though the increase in Firmicutes-C was most prominent in the Pre-DM group. A significant rise in Proteobacteria was observed exclusively in the T2DM group.

At the genus level, Pauljensenia and Veillonella-A exhibited significant enrichment, while Neisseria showed a slight reduction in the T2DM group, aligning with the trends observed in Pre-DM. Additionally, Haemophilus-D and Porphyromonas-A were specifically reduced in the T2DM group. In the gut microbiota, Bacteroides levels increased, whereas Faecalibacterium and Bifidobacterium declined as the disease progressed. Notably, Phocaeicola exhibited the most significant increase in the Pre-DM group. Escherichia was markedly enriched in the T2DM group, where it accounted for the highest relative abundance.

LEfSe analysis revealed significant distinctions in tongue coating and fecal microbiota between the Pre-DM and T2DM groups ( Figure 6 ). At the genus level, Porphyromonas, AlloPrevotella, and Staphylococcus within the oral microbiota, alongside Blautia, Lactiplantibacillus, and Romboutsia-B in the gut microbiota, emerged as potential microbial markers for Pre-DM. In contrast, Escherichia, Klebsiella, and AlloPrevotella in the gut microbiota were identified as potential indicators of T2DM.

Abnormal FBG patients exhibiting thin white fur (TW-fur) and white greasy fur (WG-fur) were selected to investigate the relationship between tongue characteristics and the oral-gut axis microbiota. At the genus level, WG-fur was associated with elevated levels of Veillonella-A and Streptococcus in the oral cavity, while increased Blautia and Prevotella were observed in the gut. The majority of the altered microbiota belonged to the phylum Firmicutes. Pearson correlation analysis revealed significant associations between oral-gut microbiota and tongue parameters. Specifically, perALL showed a positive correlation with Phocaeicola-A and Veillonella-A, while TB-a exhibited the strongest correlation with Blautia-A (correlation coefficient 0.28, P<0.01). Additionally, Escherichia was positively correlated with multiple tongue image parameters, including TB-L, TC-L, and TC-b ( Figure 7 ).

To investigate the mechanism underlying greasy fur formation, an abundance analysis of KEGG functional pathways was performed. The findings indicate that the differential alterations in tongue fur were predominantly associated with pathways related to Metabolism, with the highest abundance observed in metabolic cofactors and vitamins ( Figure 7 ). According to TCM, tongue coating results from “disharmony of the viscera and fumigation of the spleen and stomach.” Increased metabolic activity elevates blood flow to the tongue, leading to a thickened coating. The metabolism of cofactors, vitamins, and other nutrients indirectly influences the tongue coating by impacting nutritional status, metabolic function, and visceral health.

A diagnostic model was developed to assess the predictive value of tongue images and microbiota in identifying Pre-DM and T2DM. Variables included gender, age, BMI, WHR, tongue image parameters (perALL, TB-Con, TC-Con, TB-ASM, TB-ENT, TB-MEAN, TC-ASM, TC-ENT, TC-MEAN, TB-L, TB-a, TB-b, TC-L, TC-a, and TC-b), and key microorganisms from the oral and gut microbiome (Bacteroides-H, Phocaeicola-A, Escherichia, and Porphyromonas-A). Six modelling techniques were applied—Logistic Regression, SVM, Random Forest, Gradient Boosting, AdaBoost, and KNN—to differentiate Pre-DM, T2DM, and healthy controls. Among the three classification models, SVM achieved the highest accuracy, followed by KNN, while Gradient Boosting demonstrated the poorest performance for classifying Pre-DM and T2DM ( Figure 8 , Table 3 ). Feature Importance analysis from the Random Forest model highlighted TB-a and perALL from the tongue image parameters, along with Escherichia and Porphyromonas-A, as key contributors to the model’s predictive performance, serving as primary classification indicators for Pre-DM and T2DM diagnosis ( Figure 9 ).