Gut microbiome signatures discriminate deep vein thrombosis through machine learning and metabolic analysis Authors: Chiung-Ray Lu, Wen-Pin Hu, Chung-Ho Hsu, Chun-Cheng Wang, Shih-Sheng Chang, Yen-Nien Lin, Yu-Kai Lin, Wei-Hsin Chung, Kuan-Cheng Chang, Ju-Ming Wang, Der-Yang Cho, Yi-Yao Hsu, Hsin-Hwa Tsai, Zi-Lun Lai, Yu-Jen Wang, Po-Ren Hsueh, Hong-Yue Lai Published: 2026 ABSTRACT AbstractDeep vein thrombosis (DVT) remains difficult to distinguish because of its often silent presentation and the limited specificity of current diagnostic tools. We aimed to evaluate whether integrating gut microbiome profiles with routine clinical data could enhance the classification performance for identifying DVT in a case–control cohort. Stool samples were collected from individuals with DVT (n = 58), coronary artery disease (CAD, n = 56), and healthy controls (HC, n = 500). Full-length 16S rRNA gene sequencing was used to characterize the gut microbiota at species-level resolution. A random forest classifier was trained using a nested cross-validation framework, with permutation importance and SHAP (Shapley additive explanations) analyses applied to assess model interpretability. Decision curve analysis (DCA) was employed to evaluate the discriminative value of the models in an independent test set. Following linear discriminant analysis (LDA) effect size (LEfSe) screening, 95 candidate microbial features were entered into a random forest framework. Features were reduced using mutual information filtering and embedded selection to retain the final 10 for DVT vs. non-DVT classification. The integrated microbial-clinical model demonstrated substantially improved discrimination compared with the clinical-only model, achieving higher ROC-AUC [0.947 (95% CI 0.870–0.991) vs. 0.874 (95% CI 0.794–0.941)] and PR-AUC [0.793 (95% CI 0.602–0.931) vs. 0.497 (95% CI 0.274–0.724)]. Importantly, the microbiome-derived signals were robustly associated with DVT risk after adjustment for clinical covariates. Functional prediction analysis indicated enrichment of vitamin K2 and lipopolysaccharide (LPS) biosynthesis pathways in DVT, suggesting potential microbial links to coagulation and inflammation, whereas healthy controls were predominantly enriched in NAD and tetrahydrofolate (THF) biosynthesis pathways. Together, these results demonstrate that microbiome-based classification provides complementary biological insights that distinguish DVT cases from controls.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-55650-2.Keywords: Deep vein thrombosis, Gut microbiome, Random forest, SHAP, Decision curve analysisSubject terms: Biomarkers, Cardiology, Computational biology and bioinformatics, Diseases, Medical research, Microbiology Introduction Venous thromboembolism (VTE) is a condition characterized by the formation of blood clots within the venous system. It encompasses both deep vein thrombosis (DVT), which involves thrombus formation in the deep veins, most commonly in the lower extremities, and pulmonary embolism (PE), which occurs when a dislodged thrombus migrates to the pulmonary vasculature, potentially leading to life-threatening complications. VTE remains the third leading cause of vascular death, following myocardial infarction and stroke1. Notably, over the past two decades, while age-standardized mortality from myocardial infarction and stroke has declined in high-income countries, VTE mortality has remained unchanged2. VTE poses substantial clinical challenges, as its frequently asymptomatic or nonspecific presentation could delay diagnosis and management, rendering it the most common preventable cause of mortality in the hospital3. Additionally, individuals experiencing a first episode of unprovoked VTE face a stronger tendency to a recurrent episode4. About 20% of individuals die within one year of a VTE diagnosis, either due to VTE itself or, more commonly, from the underlying conditions that triggered the event5. Accordingly, identifying specific biomarkers associated with VTE status may contribute to a more comprehensive understanding of the disease’s biological landscape, potentially supporting clinical evaluation pathways. The pathophysiology of VTE is unique among vascular diseases, as VTE risk is not linked to well-established atherosclerotic cardiovascular risk factors, such as hyperlipidemia, hypertension, and diabetes mellitus6. The risk of VTE is multifactorial, including basal risk factors (older age, male sex, sedentariness, smoking, taller height, obesity, or genetics), triggers (surgery, hospitalization, pregnancy, or exogenous estrogen), and acquired risk factors (cancer or the presence of antiphospholipid antibodies)7. Patients may experience complications, such as post-thrombotic syndrome and pulmonary hypertension, after a VTE event, leading to a decline in quality of life. Consequently, the greatest opportunity to lessen the burden of VTE lies in primordial prevention (preventing the development of underlying risk factors), primary prevention (managing risk factors in high-risk individuals), and secondary prevention (reducing the risk of recurrence)7. Nevertheless, these identified risk factors could not fully explain the incidence of VTE possibly due to race/ethnicity differences, and the use of anticoagulation therapy for secondary prevention may also increase the risk of bleeding. Since VTE is a globally under-recognized vascular disease and its comprehensive surveillance system is still lacking, further investigation into complementary markers is warranted to expand the current clinical knowledge base regarding VTE-associated profiles. Approximately 100 trillion microorganisms, collectively known as the gut microbiota, have co-evolved in a symbiotic relationship with their human host within the gastrointestinal tract8. The majority of these microbial communities are bacteria, comprising over 1,000 species and totaling around 1014 organisms9. The gut microbiota also harbors over 150 times more genetic information than the human genome, collectively referred to as the gut microbiome8. The gut microbiota is essential for maintaining host health, including killing pathogens by releasing inhibitory metabolites, producing various antimicrobial substances, and strengthening immune barrier function10. Additionally, the gut microbiota also inhibits the growth and spread of pathogens by competing with them for both nutritional resources and physical space. Microbial dysbiosis is characterized by an imbalance in gut bacteria, involving changes in their composition, metabolic activity, or spatial distribution11. This imbalance has been associated with various diseases, including cardiovascular disease (CVD)12. Several factors can influence the composition and function of the gut microbiota. Among these factors, some are unmodifiable, comprising the individual genetic background and types of delivery; some are difficult to modify, including persistent viral infection and geographical and lifestyle factors; and others could potentially be modified, containing dietary choices and medication usage13. Collectively, distinct gut microbial signatures and their associated products may alert clinicians and provide additional biological perspectives associated with VTE. The 16S rRNA gene has long served as a cornerstone for sequence-based bacterial analysis. Utilizing 16S rRNA gene sequencing of fecal samples from Asian participants, a recent study found that, at the genus level, VTE patients (n = 8) exhibited an increased abundance of Blautia, Roseburia, Coprococcus, and Ruminococcus compared to healthy controls (n = 7)14. However, the potential causal relationship between the gut microbiota and VTE is still unclear. Mendelian randomization is a method that uses genetic variants, such as single-nucleotide polymorphisms (SNPs) identified through genome-wide association study (GWAS), to explore and uncover causal relationships between complex traits. Using Mendelian randomization analysis, Yang et al. found that the family Streptococcaceae had a positive genetic causal relationship with DVT15, while Xi et al. observed that the positive genetic causal effect of the genus Coprococcus1 on VTE16. Nonetheless, the genetic data from these two studies were largely obtained from European individuals. Together, previous research suggests that the gut microbiota may be associated with the development of VTE, but the findings have been highly variable, likely due to factors such as ethnic differences and limited phylogenetic resolution at the species level. Fortunately, with advances in long-read third-generation sequencing, analyzing full-length 16S rRNA gene variants can enhance the accuracy of bacterial identification at the species and strain levels17. Moreover, as our study focuses on the gut-vascular axis, our hypothesis suggests that thromboinflammatory processes may bridge gut dysbiosis with in situ thrombus formation (DVT) rather than embolic propagation (PE). Accordingly, this study aimed to identify distinct gut microbial signatures from a species-level resolution in our well-characterized DVT cohort from an Asian population. Patients and methods Patient recruitment We recruited a cohort of 500 healthy controls (HC), 56 patients diagnosed with coronary artery disease (CAD) without deep vein thrombosis (DVT), and 58 patients diagnosed with DVT from October 2023 to September 2024 at China Medical University Hospital, Taiwan (CMUH112-REC2-087). The sample size was based on the availability of patients with a diagnosis of DVT during the study period; no formal sample size calculation was undertaken. Fecal samples were collected during admission for gut microbiota analysis. No formal study protocol was prepared or registered. Healthy controls were recruited as community volunteers from the same geographic area as the patient cohorts to ensure consistency in environmental and dietary exposures. This strategy allowed for a robust comparison between the disease-associated dysbiosis and the representative biological baseline of the local community. Definition of CAD patients: Patients with a confirmed diagnosis of significant coronary artery obstruction, defined as stenosis greater than 50%, regardless of whether they have undergone catheter-based interventions or are being treated with medication alone. Definition of DVT patients: Patients with a confirmed diagnosis of peripheral venous thrombosis based on imaging studies, such as peripheral ultrasound or computed tomography. Cases were defined as onset within 30 days. Data on pre-admission antithrombotic therapy and laboratory parameters from the first blood draw in the emergency department were recorded for patients with CAD and DVT (Supplementary Tables 1 and 2). Inclusion criteria: Age and consent: Participants must be at least 20 years old, willing to participate, and capable of completing the study requirements. Written informed consent was obtained from the participants or their legal representatives. Age and consent: Participants must be at least 20 years old, willing to participate, and capable of completing the study requirements. Written informed consent was obtained from the participants or their legal representatives. Blood donation history: No history of blood donation exceeding the specified amounts within the past three months:Female: 250 mL.Male: 500 mL. Blood donation history: No history of blood donation exceeding the specified amounts within the past three months:Female: 250 mL.Male: 500 mL. Female: 250 mL. Female: 250 mL. Male: 500 mL. Male: 500 mL. Antibiotic use: No antibiotic use within two weeks prior to hospital admission. Antibiotic use: No antibiotic use within two weeks prior to hospital admission. Exclusion criteria: Vulnerable populations:Individuals under 20 years of age.Incarcerated individuals.Indigenous populations.Pregnant women.Individuals with psychiatric disorders. Vulnerable populations:Individuals under 20 years of age.Incarcerated individuals.Indigenous populations.Pregnant women.Individuals with psychiatric disorders. Individuals under 20 years of age. Individuals under 20 years of age. Incarcerated individuals. Incarcerated individuals. Indigenous populations. Indigenous populations. Pregnant women. Pregnant women. Individuals with psychiatric disorders. Individuals with psychiatric disorders. Procedure limitation: Patients who are unable to undergo catheter-based deep vein intervention procedures. Procedure limitation: Patients who are unable to undergo catheter-based deep vein intervention procedures. Missing data were handled by excluding incomplete cases. Missing data were handled by excluding incomplete cases. The workflow of PacBio full-length 16S rRNA gene sequencing and analysis Fecal specimens were naturally collected, and genomic DNA was extracted using the QIAcube HT (9,001,896, QIAGEN), an automated nucleic acid purification system known for its speed and reliability, in combination with the QIAamp 96 PowerFecal QIAcube HT Kit (51,531, QIAGEN). The full-length 16S rRNA gene (V1–V9 region, ~ 1.5 kb) was then amplified using universal primers 27F (5′-AGR GTT YGA TYM TGG CTC AG-3′) and 1492R (5′-RGY TAC CTT GTT ACG ACT T-3′), along with high-fidelity DNA polymerase from the KAPA HiFi HotStart ReadyMix PCR Kit. PCR products were verified by electrophoresis on a 0.8% agarose gel and purified using magnetic beads (AMPure PB, PacBio). The purified amplicons were then processed for end repair and adapter ligation to construct SMRTbell libraries using the SMRTbell® Prep Kit 3.0 (102-141-700, PacBio). After evaluating the integrity and size distribution of the SMRTbell libraries using the Agilent 4200 TapeStation system with the D5000 ScreenTape assay, single-molecule real-time (SMRT) sequencing was conducted on the PacBio Sequel IIe system using the SMRT Cell 8 M. Highly accurate circular consensus sequencing (CCS) reads were generated from raw PacBio sequencing data using SMRT Link software (v9.0). After demultiplexing with LIMA (v2.0), the CCS reads underwent quality control, denoising, amplicon sequence variant (ASV) inference, and chimera removal using the DADA2 R package (v4.2.0). Finally, the taxonomic classification of ASVs at the species level was conducted using the SILVA 16S rRNA reference database (v132). Alpha and beta diversity analysis QIIME2 (quantitative insights into microbial ecology) was used for analyzing microbiome alpha and beta diversity. Alpha diversity metrics were calculated to evaluate microbial diversity within samples, where richness represents the number of taxa present and evenness reflects the balance of their relative abundances. Three complementary indices were assessed: Pielou’s evenness (community evenness), observed features (ASV richness), and the Shannon diversity index, which incorporates both evenness and richness. Group-wise comparisons of alpha diversity were performed using the Kruskal–Wallis rank-sum test for overall group comparisons, followed by pairwise Wilcoxon rank-sum tests for post hoc analysis where appropriate. Beta diversity was evaluated using Bray–Curtis dissimilarity, which accounts for differences in microbial abundance between samples. Group differences in community composition were tested with pairwise PERMANOVA (permutational multivariate analysis of variance) using 999 permutations. Principal coordinates analysis (PCoA) based on Bray–Curtis distances was performed at the taxonomic level and visualized using the vegan package in R. Ellipses denote 95% confidence intervals, highlighting clustering patterns and the distribution of gut microbial communities across the three study groups. Microbial differential abundance analysis using LEfSe Linear discriminant analysis (LDA) effect size (LEfSe) is a biomarker discovery and explanation method specifically designed for high-dimensional data, such as microbiome profiles. LEfSe was used to identify candidate microbial taxa between groups by combining non-parametric statistical testing with LDA to estimate the effect size of each feature, prioritizing both statistical significance and biological relevance. Taxonomic features were first analyzed using the Kruskal–Wallis rank-sum test to detect overall group-wise differences, followed by pairwise Wilcoxon rank-sum tests to assess between-group significance. Features meeting the significance threshold (p < 0.05) were then ranked according to their LDA scores. Results were visualized using a taxonomic cladogram to illustrate phylogenetic relationships and highlight enriched taxa, alongside bar plots of log10-transformed LDA scores (with LDA score ≥ 2.0) to reflect the magnitude of differential abundance. All analyses were conducted using the Galaxy platform (Huttenhower Lab, Harvard) with default parameters unless otherwise specified. Random forest modeling and feature interpretation Data splitting and preprocessing: The study cohort was partitioned into a training set (70%) and an independent test set (30%) using an ordered within-group split to preserve temporal integrity. Features included clinical variables and microbial taxa. Categorical variables were encoded using one-hot encoding. Model pipeline and nested cross-validation: LEfSe-based microbial screening was performed using only the training set before nested cross-validation as an initial biologically informed prescreening step to reduce the feature space. The retained microbial features were then entered into a nested cross-validation framework consisting of a fivefold outer loop repeated three times and a threefold inner loop. Within each inner-loop training fold, mutual information filtering was applied to retain the top 30 features, followed by embedded feature selection via SelectFromModel (top 10 features), and then hyperparameter tuning of the random forest classifier. After the optimal settings were identified in the inner loop, the model was retrained on the full outer-loop training set and evaluated on the corresponding unseen outer-loop test fold. To address class imbalance, random forest weights were set to balanced. Performance metrics and probability calibration: To evaluate model performance, we utilized the area under the receiver operating characteristic (ROC-AUC) and precision–recall (PR-AUC) curves as primary metrics, the latter being particularly robust for class imbalance. Furthermore, threshold-dependent metrics, including sensitivity, specificity, accuracy, balanced accuracy, precision, and F1-score, were calculated based on the optimal Youden’s index threshold. Predicted probabilities were calibrated using Platt scaling (sigmoid method). Calibration quality was assessed using the expected calibration error (ECE) and Brier score. Calibration was assessed through quantile binning (4 bins), ensuring that all performance assessments accounted for potential sampling variability. To ensure the reliability of these point estimates, 95% confidence intervals (CIs) were derived for each metric using 2000 bootstrap iterations, providing a rigorous assessment of the model’s discriminative stability. Feature importance and interpretability: Permutation importance was used to quantify the average decrease in AUC across outer validation folds when each feature was randomly permuted. SHAP (Shapley additive explanations) analysis was applied to estimate global feature importance by averaging absolute SHAP values across outer folds, and to visualize feature effects and interactions in the final refitted model18. Additionally, feature selection frequencies across outer folds were recorded to assess the stability and robustness of the identified biological markers. Clinical utility and inference modeling We utilized decision curve analysis (DCA) to examine the model’s stability and evaluate if the integration of microbial data offers incremental net benefit (NB). A paired bootstrap approach (n = 2000 iterations) was employed to compare the NB of the integrated (microbial + clinical) model against a clinical-only baseline. Analysis was conducted across an expanded grid of decision thresholds (Pt = 0–0.3) and simulated disease prevalence (π = 0–0.4) using synthetic prior shift analysis via odds-scaling to ensure model robustness across diverse clinical settings. The optimal clinical decision zone was identified using out-of-fold (OOF) data by selecting cells where the lower bound of the 95% confidence interval (CILo) for the change in net benefit (ΔNB) was greater than zero. To ensure decision stability, we identified a “heat-zone” range around the best-performing cell, defined as cells maintaining ≥ 85% of the maximum ΔNB. This OOF-defined decision strategy was independently validated on the test set to report generalizable metrics, including the median true positive gain and net reduction in false-positive classifications per 100 patients. Furthermore, specific decision thresholds were evaluated via Youden’s index and a fixed specificity target (≥ 90%). To quantify the contribution of the microbiome score while adjusting for traditional risk factors, we constructed a generalized linear model using the test set. The primary predictor was Δlogit, defined as the difference in the log-odds of DVT predicted by the integrated model versus the clinical-only model. The model was adjusted for age, gender, BMI, hypertension, and the use of DOACs or antiplatelet agents. Robustness was confirmed through a parallel ridge logistic regression with bootstrap-derived 95% CIs (n = 2000 iterations) to account for potential overfitting in the test set. Results Clinical characteristics of the DVT, CAD, and healthy control cohort A detailed flowchart outlining the overall study design and analysis is illustrated in Fig. 1. Because DVT and CAD share overlapping risk factors and may present subclinically, differentiating DVT from clinically relevant non-DVT cardiovascular conditions remains challenging. CAD patients were therefore included as disease controls and combined with healthy individuals as the non-DVT class to reflect real-world diagnostic complexity. To address this issue, we recruited 58 patients with DVT, 56 patients with CAD, and 500 healthy controls (HC). The DVT cohort was defined as onset within 30 days. Within this group, 34.5% were classified as provoked DVT associated with recent surgery or active malignancy, while 65.5% were unprovoked. Stool samples were collected at the time of admission and before any catheter-based thrombolysis or thrombectomy, minimizing the potential impact of acute hospital-based interventions on the microbiome profiles. The cohort was split according to the order of enrollment. Within each group, the earlier 70% of cases were assigned to the training set (350 HC, 39 CAD, and 40 DVT) for feature selection and model development, and the later 30% were reserved as an independent test set (150 HC, 17 CAD, and 18 DVT) for validation. Clinical characteristics of the total cohort, training, and test groups are summarized in Table 1 and Supplementary Tables 3–4. These characteristics were categorized into demographics and lifestyle, genetics, medical history and comorbidities, and recent medical interventions. According to 2023 demographic data from Taiwan (https://pop-proj.ndc.gov.tw/main_en/), the median age of the population was 44.4 years, with males accounting for 49.3%. In the total cohort (Table 1), the median age of the HC was 42 years (interquartile range, IQR: 34–51), and 45.2% were male, which is comparable to the population-level demographic data. The median age of patients with CAD was 65.5 years (IQR: 60–72), with 82.1% being male. In contrast, the median age of patients with DVT was 66.5 years (IQR: 53–72.75), with 55.2% being male. An overview of the study design and analytical workflow. Table 1Summary of clinical characteristics among the total cohort.FeaturesHC (n = 500)CAD (n = 56)p-value(HC vs CAD)DVT (n = 58)p-value(HC vs VTE)p-value(CAD vs VTE)Demographics and lifestyle Age (years)42 (IQR: 34–51)65.5 (IQR: 60–72)< 0.00166.5 (IQR: 53–72.75)< 0.0010.55 Gender (male)226 (45.2%)46 (82.1%)< 0.00132 (55.2%)0.1490.002 Height (cm)171.97 ± 5.96 (male)166.17 ± 6.88< 0.001169.78 ± 7.350.1170.029158.98 ± 5.11 (female)154.19 ± 6.110.036156.75 ± 5.440.0540.23 BMI (kg/m2)23.7 ± 3.5625.02 ± 3.650.01224.7 ± 4.110.0770.668 Alcohol149 (29.8%)2 (3.6%)< 0.0017 (12.1%)0.0040.093 Smoking No456 (91.2%)26 (46.4%)< 0.00140 (69%)< 0.0010.036 Past24 (4.8%)22 (39.3%)11 (19%) Current20 (4%)8 (14.3%)7 (12.1%)Genetics Blood type O223 (44.6%)25 (44.6%)0.96814 (24.1%)0.0110.072 A124 (24.8%)15 (26.8%)15 (25.9%) B129 (25.8%)14 (25%)24 (41.4%) AB24 (4.8%)2 (3.6%)5 (8.6%)Medical history and comorbidities History of CAD14 (2.8%)34 (60.7%)< 0.0011 (1.7%)0.632< 0.001 History of VTE(> 30 days)0 (0%)0 (0%)–10 (17.2%)< 0.0010.001 Hypertension47 (9.4%)41 (73.2)< 0.00133 (56.9%)< 0.0010.068 Diabetes16 (3.2%)22 (39.3%)< 0.00110 (17.2%)< 0.0010.009 Liver disease21 (4.2%)1 (1.8%)0.3795 (8.6%)0.1310.102 Renal disease5 (1%)18 (32.1%)< 0.00112 (20.7%)< 0.0010.165 CVA3 (0.6%)4 (7.1%)< 0.0014 (6.9%)< 0.0010.959 Coagulopathy0 (0%)0 (0%)–17 (29.3%)< 0.001< 0.001 Hyperhomocysteinemia0 (0%)0 (0%)–5 (8.6%)< 0.0010.025 Chronic anemia(Hb < 10 g/dL)25 (5%)5 (8.9%)0.2178 (13.8%)0.0070.414 APS0 (0%)0 (0%)–5 (8.6%)< 0.0010.025 Malignancy11 (2.2%)7 (12.5%)< 0.00118 (31%)< 0.0010.017Medical interventions Operation(within 30 days)0 (0%)0 (0%)–8 (13.8%) < 0.0010.004 Aspirin and/or P2Y12i0 (0%)49 (87.5%)< 0.0015 (8.6%) < 0.001 < 0.001 DOAC0 (0%)2 (3.6%)< 0.00116 (27.6%) < 0.001 < 0.001HC: healthy control; CAD: coronary artery disease; DVT: deep vein thrombosis; BMI: body mass index; CVA: cerebrovascular accident; Hb: hemoglobin; APS: antiphospholipid syndrome; P2Y12i: P2Y12 inhibitors; DOAC: direct oral anticoagulant. Discrete variables were reported as median (interquartile range, IQR) and compared using the Mann–Whitney U test. Continuous variables were presented as mean ± standard deviation and analyzed using a Student’s t-test or Welch’s t-test (when equal variance was not assumed). Categorical variables were expressed as numbers (percentages) and analyzed using the Chi-square test.Significant values are in bold. Summary of clinical characteristics among the total cohort. History of VTE (> 30 days) Chronic anemia (Hb < 10 g/dL) Operation (within 30 days) HC: healthy control; CAD: coronary artery disease; DVT: deep vein thrombosis; BMI: body mass index; CVA: cerebrovascular accident; Hb: hemoglobin; APS: antiphospholipid syndrome; P2Y12i: P2Y12 inhibitors; DOAC: direct oral anticoagulant. Discrete variables were reported as median (interquartile range, IQR) and compared using the Mann–Whitney U test. Continuous variables were presented as mean ± standard deviation and analyzed using a Student’s t-test or Welch’s t-test (when equal variance was not assumed). Categorical variables were expressed as numbers (percentages) and analyzed using the Chi-square test. Significant values are in bold. Distinct microbial diversity and composition in DVT patients suggest discriminative features for disease classification Traditional DVT risk scores rely on clinical features that may be insufficient in atypical or subclinical cases. By contrast, the gut microbiome offers a biologically rich biomarker source that enhances conventional assessment by capturing risk dimensions not reflected in clinical parameters. To minimize bias, fecal samples were collected naturally from participants who had not used antibiotics within 14 days prior to admission. These samples underwent PacBio full-length 16S rRNA gene sequencing and analysis. Significant differences in gut microbial communities at the phylum level were observed among control, DVT, and CAD groups, with Proteobacteria enriched in both DVT and CAD patients. Notably, DVT patients exhibited lower levels of Firmicutes compared to those with CAD (Fig. 2A). In terms of α diversity, DVT patients showed significantly lower gut microbial richness (Shannon index: p = 6.35 × 10−4 and 7.45 × 10−4; observed features: p = 1.60 × 10−7 and 1.80 × 10−5) compared to healthy controls and CAD patients, respectively (Table 2; Fig. 2B). For β diversity, based on Bray–Curtis dissimilarity, PERMANOVA revealed that the microbial community composition of DVT patients was significantly different from those of healthy controls and CAD patients (p = 0.001 and 0.003, respectively; Table 2). To visualize these differences, principal coordinate analysis (PCoA) was performed, projecting high-dimensional taxonomic data into a lower-dimensional space while preserving variance and structural integrity. The first two principal coordinates (PCoA1 and PCoA2) explained 10.15% and 6.26% of the total variation, respectively (Fig. 2C). Although partial overlap was observed among the three groups, the 95% confidence ellipses highlight significant shifts in community centroids. Collectively, these findings suggest that DVT may be identifiable based on low-diversity dysbiosis and distinct microbial community profiles, supporting the potential of microbiota as discriminative markers for disease classification. Comparisons of fecal microbiome among control, DVT, and CAD groups. (A) Stacked bar chart illustrating the relative abundance of the six dominant phyla of fecal microbiome among healthy controls (n = 500), DVT patients (n = 58), and CAD patients (n = 56). Phyla not displayed individually are grouped under “Others”. (B) Box plots depicting the Shannon diversity index and observed species richness of the fecal microbiome. Each plot displays the median (center line), interquartile range (25th to 75th percentiles; box), and the 5th to 95th percentile range (whiskers). P values were determined using two-sided Wilcoxon rank-sum tests. (C) PCoA based on Bray–Curtis distances at the taxonomic level revealed distinct microbial compositions among groups, visualized using the vegan package in R. Ellipses indicate 95% confidence intervals, illustrating the clustering and distribution of gut microbial communities across the three study groups. Table 2Alpha and beta diversity analysis.Groupsα diversityβ diversity(Bray–Curtis dissimilarity)Evenness(Pielou’s evenness)Richness(Observed features)Evenness + Richness(Shannon index)DVT vs HCp = 0.947745p = 1.60 × 10−7p = 6.35 × 10−4p = 0.001CAD vs HCp = 0.398164p = 1.70 × 10−1p = 1.85 × 10−1p = 0.001DVT vs CADp = 0.574704p = 1.80 × 10−5p = 7.45 × 10−4p = 0.003DVT deep vein thrombosis, HC healthy control, CAD coronary artery disease.Significant values are in bold. Alpha and beta diversity analysis. DVT deep vein thrombosis, HC healthy control, CAD coronary artery disease. Significant values are in bold. LEfSe-based identification of candidate microbial features for DVT classification To reduce the high dimensionality of the microbiome data, we employed linear discriminant analysis (LDA) effect size (LEfSe) as an exploratory, biologically informed prescreening step. A key advantage of LEfSe lies in its interpretability, as it ranks features by effect size, enabling clear comparisons of microbial patterns across clinical groups. LEfSe was applied to the training set to identify candidate microbial taxa. The cladogram illustrates the hierarchical relationships among differentially abundant taxa, with node colors indicating the group in which each taxon is most enriched. A circular cladogram revealed that the phyla Synergistota, Proteobacteria, and Fusobacteria were enriched in DVT patients (green), while Firmicutes and Actinobacteria predominated in healthy controls (red) (Fig. 3A). Additionally, Actinobacteria was the most abundant phylum in CAD patients (blue) compared to DVT patients (green) (Fig. 3B). Histograms ranked by log10-transformed LDA scores (LDA scores ≥ 2.0) highlighted discriminative features differentiating DVT patients from healthy controls (Supplementary Fig. 1A) and from CAD patients (Supplementary Fig. 1B). At the species level, these differences were illustrated in Fig. 3C (DVT vs HC) and Fig. 3D (DVT vs CAD). We found that the top 3 species enriched in DVT patients were Escherichia coli, Klebsiella pneumoniae, and Fusobacterium varium, while Bacteroides coprocola, Bifidobacterium pseudocatenulatum, and Collinsella aerofaciens were most abundant in healthy controls. In comparison to CAD patients, the top 3 species enriched in DVT patients were Klebsiella pneumoniae, Enterococcus faecium, and Clostridium bolteae. Taxonomic cladogram and species-level visualization of candidate microbiota identified by LEfSe. (A–B) The cladogram illustrates the hierarchical taxonomic structure of microbial communities. Nodes closer to the center represent higher taxonomic levels, starting from phylum (p), followed by class (c), order (o), and family (f) as the branches extend outward. Somewhat counterintuitively, the outermost ring displays the full names of phyla, but only for those showing statistically significant differences between groups. Colored nodes and sectors indicate the group in which each taxon is most abundant: red for healthy controls, green for DVT patients, blue for CAD patients, and yellow for taxa without significant differences. Taxonomic identifiers (letters adjacent to the nodes) correspond to class, order, and family levels and are provided in the accompanying legend to aid interpretation. (C–D) Species (s)-level taxa were extracted from LEfSe histograms based on discriminative features (log10-transformed LDA scores ≥ 2.0) and visualized to highlight key microbial signatures differentiating the clinical groups. Model discrimination, calibration, and feature robustness To classify DVT versus non-DVT cases (HC and CAD), we first constructed a candidate feature set by taking the union of biologically relevant microbial species identified in Fig. 3C and D, yielding 95 microbial features for downstream modeling. Given the high feature-to-sample ratio of the microbiome data, we applied a random forest-based discriminative framework. A two-step feature selection strategy was then implemented, consisting of initial filtering by mutual information (top 30 features), followed by embedded feature selection within the model to retain the top 10 features. Model tuning and performance evaluation were conducted using nested cross-validation (fivefold outer loop repeated three times, with a threefold inner loop). Class imbalance was addressed by applying balanced class weights. Model performance was primarily evaluated using ROC-AUC and precision–recall AUC (PR-AUC). Additionally, threshold-dependent metrics, including sensitivity, specificity, accuracy, balanced accuracy, precision, and F1-score, were determined based on the optimal Youden’s index. Predicted probabilities were calibrated with Platt scaling, and calibration quality was assessed using the Brier score and expected calibration error (ECE). Because traditional DVT scores are symptom-driven and not designed for subclinical populations, we assessed whether integrating microbiome-derived features into a non-symptom-based clinical model (22 predictors; Table 1) provides incremental discriminative value in clinically ambiguous settings where existing tools are least informative. Discrimination analysis on the test set revealed a significant performance improvement for the integrated microbial-clinical model, which outperformed the clinical-only model in ROC-AUC [0.947 (95% CI 0.870–0.991) vs. 0.874 (95% CI 0.794–0.941)], PR-AUC [0.793 (95% CI 0.602–0.931) vs. 0.497 (95% CI 0.274–0.724)], F1-score [0.627 (95% CI 0.462–0.774) vs. 0.399 (95% CI 0.254–0.538)], and balanced accuracy [91.6% (95% CI 84.7%–96.2%) vs. 79.3% (95% CI 69.2%–87.5%)] (Fig. 4A–B; Table 3). Calibration analysis on the test set demonstrated that the integrated model achieved a lower Brier score (0.045 vs. 0.069) and comparable ECE (0.029 vs. 0.022) (Fig. 4C and G, left), indicating overall better agreement between predicted and observed risks. Additionally, the distribution of predicted probabilities showed improved separation between DVT and non-DVT cases in the integrated model, with DVT cases being more frequently assigned higher predicted risks (Fig. 4C and G, right). To assess model robustness and interpretability, we evaluated feature contributions from complementary perspectives. Feature stability across outer validation folds was first examined to identify consistently selected predictors (Supplementary Fig. 2A–B). Global feature importance was then quantified using absolute SHAP values averaged across outer folds (Fig. 4D and H) and visualized in the final refitted model (Fig. 4E and I), highlighting the relative contribution of each feature to model predictions. Finally, permutation importance was applied as a model-agnostic check by measuring the mean decrease in AUC when individual features were randomly permuted (Fig. 4F and J). Together, these analyses demonstrated a highly consistent ranking of key features, with Fusicatenibacter saccharivorans, Clostridium bolteae, and age emerging as the most influential predictors in the integrated model, and age, hypertension, and pre-admission use of DOACs in the clinical-only model, supporting the stability and robustness of the respective models. Notably, pre-admission DOAC use in our cohort primarily reflects long-term management for conditions such as atrial fibrillation, previous VTE history, or apex thrombus, rather than acute treatment for the current episode. Additionally, SHAP interaction heatmaps revealed minimal cross-feature interactions, indicating that predictors primarily acted through additive effects, with limited synergistic influence (Supplementary Fig. 2C–D). Performance, calibration, and interpretability of the clinical-only and integrated model. (A–B) ROC and PR curves were used to compare the discriminative performance of the clinical-only model and the integrated microbial-clinical model on the test set. (C, G) Calibration curves (left) and distributions of predicted probabilities (right) on the test set for the clinical-only model and the integrated model. (D, H) Global feature importance estimated by mean absolute SHAP values across outer validation folds. (E, I) SHAP summary plot for the final refitted model, illustrating the direction and magnitude of individual feature effects on predicted DVT risk. Each dot represents the SHAP value for a specific feature in an individual sample; the overall distribution reflects the feature’s influence on the model’s prediction. Positive SHAP values indicate features that contribute to predicting DVT, while negative values push the prediction toward non-DVT. (F, J) Permutation importance was quantified as the mean decrease in ROC-AUC across outer validation folds when each feature was randomly permuted. Blood type was encoded as: 0 = O, 1 = A, 2 = B, 3 = AB; DOAC use: 0 = no use, 1 = use; alcohol consumption: 0 = non-drinker, 1 = drinker; gender: 0 = female, 1 = male; aspirin and/or P2Y12 inhibitor use: 0 = not taking, 1 = taking; smoking status: 0 = never smoker, 1 = past smoker, 2 = current smoker. Table 3Evaluation of classification performance on the test set.DVT vs non-DVT(HC + CAD)SensitivitySpecificityPrecisionF1-scoreAccuracyBalanced accuracyROC AUCPR AUCClinical-only83.6% (95% CI 64.7–100.0%)74.9% (95% CI 68.3–81.1%)26.5% (95% CI 15.6–38.5%)0.399 (95% CI 0.254–0.538)75.8% (95% CI 69.2–81.6%)79.3% (95% CI 69.2–87.5%)0.874 (95% CI 0.794–0.941)0.497 (95% CI 0.274–0.724)Microbial + Clinical94.5% (95% CI 81.8–100.0%)88.7% (95% CI 83.5–93.4%)47.4% (95% CI 31.0–64.7%)0.627 (95% CI 0.462–0.774)89.3% (95% CI 84.3–93.5%)91.6% (95% CI 84.7–96.2%)0.947 (95% CI 0.870–0.991)0.793 (95% CI 0.602–0.931)DVT: deep vein thrombosis; HC: healthy control; CAD: coronary artery disease; ROC AUC: receiver operating characteristic area under the curve; PR AUC: precision–recall AUC; CI: confidence interval. Evaluation of classification performance on the test set. DVT: deep vein thrombosis; HC: healthy control; CAD: coronary artery disease; ROC AUC: receiver operating characteristic area under the curve; PR AUC: precision–recall AUC; CI: confidence interval. The potential of microbiome-derived signals to provide complementary biological information In the standard diagnostic pathway for suspected DVT, patients at high clinical risk are directly referred for imaging, whereas those with low clinical risk and a negative D-dimer test can be safely ruled out. The main diagnostic uncertainty lies in patients with low-to-intermediate clinical risk but a positive D-dimer result. Although D-dimer is highly sensitive, its specificity is limited, resulting in a low positive predictive value (PPV, less than 20%)19,20. Consequently, decisions to proceed with imaging in this “gray zone” depend largely on clinical judgment. This suggests that microbiome-derived signatures may serve as complementary biological markers that reflect the distinct profile of DVT patients beyond traditional clinical risk factors. To evaluate the model’s behavioral stability, we employed decision curve analysis (DCA) as an exploratory framework. This analysis assesses whether the integrated microbial-clinical signatures provide incremental net benefit (NB) across a spectrum of threshold probabilities, thereby refining the biological characterization of patients within clinically ambiguous decision ranges. On the test set, DCA demonstrated that the integrated model provided a positive net benefit (ΔNB > 0) over the clinical-only model across a broad range of threshold probabilities (Pt = 0–0.3) and plausible disease prevalences (π = 0–0.4). Regions where the lower bound of the 95% confidence interval (CILo) for ΔNB exceeded zero were contoured (Fig. 5A), indicating exploratory discriminative potential. To reflect the low PPV of D-dimer in patients with low-to-moderate clinical suspicion, post-D-dimer risk was approximated using clinically plausible values (π = 0.1, 0.15, and 0.2). When π was fixed, the benefits were greatest at lower π values (Fig. 5B–D). At π = 0.10, the model achieved a median true positive gain of 3.78 per 100 patients (IQR, 3.65–3.78) and a median net reduction of 22.17 false-positive classifications per 100 patients (IQR, 19.38–24.05), indicating enhanced clinical utility in diagnostically ambiguous settings. These benefits remained present but were attenuated at higher π values. We further quantified the contribution of the microbiome-derived signal using a generalized linear model (GLM). The Δlogit (microbiome score), defined as the difference in predicted log-odds between the integrated and clinical-only models for each individual, showed a strong positive association with DVT risk, with higher Δlogit values corresponding to a monotonic increase in predicted DVT probability following adjusting for clinical priority (age, gender, BMI, and pre-admission antithrombotic therapy) and hypertension (among the top-ranked features in our integrated model) (Fig. 5E). This association was further validated using ridge logistic regression, which yielded a significant odds ratio for Δlogit (Supplementary Fig. 3), confirming the stability of the microbiome-derived effect. Incremental discriminative value of microbiome-derived signals. (A) Heatmap of the change in net benefit (ΔNB, clinical + microbial − clinical-only) comparing the integrated model with the clinical-only model on the test set across threshold probabilities (Pt = 0–0.3) and assumed disease prevalences (π = 0–0.4). Black contours indicate regions where the lower bound of the 95% confidence interval (CILo) for ΔNB exceeded zero, denoting exploratory evidence of discriminative gain. Positive ΔNB corresponds to a higher number of true positives detected and fewer false-positive classifications per 100 patients. (B–D) Decision curve analysis at fixed disease prevalences (π = 0.1, 0.15, and 0.2). (E) Partial effect plot from a generalized linear model illustrating the relationship between Δlogit [microbiome score, logit(Pclin + micro​) − logit(Pclin​)] and predicted DVT probability. Functional profiling of microbial signatures reveals distinct metabolic features associated with DVT To gain biological insight into the microbial markers identified by the classification model, we applied PICRUSt2 (phylogenetic investigation of communities by reconstruction of unobserved states) to predict functional metabolic capacities from 16S rRNA gene profiles. The inferred metabolic pathways enriched in DVT patients were primarily associated with vitamin K2 (menaquinone) biosynthesis, lipopolysaccharide (LPS) biosynthesis, and energy metabolism (TCA cycle variants and glyoxylate bypass) (Supplementary Fig. 4). Energy metabolism provides key intermediates (e.g., acetyl-CoA, NADH) that fuel both menaquinone and LPS biosynthesis. Menaquinone, in turn, functions as an electron carrier in the bacterial electron transport chain, supporting energy production. LPS biosynthesis depends on acetyl-CoA-derived lipids and sugar nucleotides derived from central carbon metabolism. Together, these pathways form a tightly linked functional module that may support bacterial persistence and prothrombotic activity in the context of DVT. In contrast, healthy controls were mainly enriched in NAD biosynthesis, tetrahydrofolate (THF) biosynthesis, and polyamine biosynthesis (Supplementary Fig. 4). THF acts as a carrier of one-carbon units essential for nucleotide and methionine synthesis, as well as NAD biosynthesis. NAD serves as a key redox cofactor, and its synthesis relies on THF-derived intermediates. Polyamine biosynthesis depends on S-adenosylmethionine (SAM), which is regenerated via THF-dependent methylation. These pathways collectively support biosynthetic demands and redox balance, forming a coordinated metabolic network that promotes vascular and immune homeostasis. Compared to DVT, CAD patients exhibited enrichment in polyamine biosynthesis only (Supplementary Fig. 5), suggesting a more comparable metabolic profile between DVT and CAD groups. To further explore the functional relevance of discriminative taxa in DVT, we conducted Spearman correlation analysis between microbial species and predicted metabolic pathways in the DVT versus healthy control comparison. Most taxa identified by permutation importance and SHAP analysis were found to be functionally linked to specific metabolic capacities (Supplementary Fig. 6), providing a hypothesis-generating context for the microbial features. Together, these pathway predictions and correlation patterns offer supportive evidence that strengthens the biological relevance and interpretability of key microbial contributors to DVT. Discussion The ABO blood types are a well-established and independent genetic risk factor for VTE, accounting for approximately 30% of cases21. The ABO locus encodes glycosyltransferases A and B, which mediate saccharide transfer to the surfaces of red blood cells and also influence von Willebrand factor (vWF) glycosylation, thereby affecting its clearance and cleavage by ADAMTS1322. Individuals with non-O blood types (A, B, and AB) have a higher risk of VTE, likely due to elevated plasma levels of vWF and factor VIII (FVIII)21. However, the influence of the ABO locus on VTE risk cannot be fully explained by vWF-dependent mechanisms, as the association persists even after adjusting for vWF and FVIII levels22. According to data published by the Taiwan Blood Services Foundation in April 2025 (https://www.blood.org.tw/xmdoc/cont?xsmsid=0P063554032741653967&sid=0P065535994814600583), the distribution of ABO blood types in Taiwan is as follows: type O (43.2%), type A (26.6%), type B (24%), and type AB (6.1%). The ABO blood type distribution among our healthy controls was O (44.6%), A (24.8%), B (25.8%), and AB (4.8%) (Table 1), which closely resembles the population-level data in Taiwan. Interestingly, we observed that the proportions of blood types O and B among our DVT patients were 24.1% and 41.4%, respectively (Table 1). Moreover, the SHAP analysis of the random forest model identified blood type B as one of the most predominant features associated with DVT (Fig. 4D–F). Although the strong association between ABO blood types, particularly blood type B, as observed in this study, and the risk of DVT is well established, the underlying biological mechanisms remain unclear and warrant further investigation. Arterial and venous thromboses differ in their pathophysiology: arterial thrombosis typically occurs at sites of plaque rupture under high shear stress and is platelet-rich, whereas venous thrombosis develops in low-shear environments and is rich in red blood cells and fibrin, primarily driven by coagulation factors23. Despite these differences, inflammation and endothelial dysfunction are common contributors to both. Symptomatic atherosclerotic events increase the risk of VTE, and patients with unprovoked VTE also face an elevated risk of arterial thrombosis24. In our study, several gut microbial taxa were shared between DVT and CAD, indicating overlapping dysbiotic signatures across venous and arterial thrombosis. These shared taxa aligned with similar clinical-microbial correlation patterns, suggesting that common host factors may shape the microbiome, while microbial alterations may in turn influence these same factors. Accordingly, this overlap supports the concept that DVT and CAD may arise from partly shared, microbiome-linked pathophysiological pathways. Most taxa identified by SHAP analysis were functionally linked to distinct metabolic capacities. In DVT patients, enriched pathways primarily involved vitamin K₂ biosynthesis, LPS biosynthesis, and energy metabolism, whereas healthy controls showed enrichment in NAD, tetrahydrofolate, and polyamine biosynthesis, providing a hypothesis-generating context to the microbial signatures (Supplementary Figs. 4 and 6). The overrepresentation of vitamin K₂ biosynthesis pathways (e.g., menaquinol-6 to -13) suggests a potential microbial contribution to coagulation, given vitamin K’s critical role in activating clotting factors25. Enrichment of LPS biosynthesis pathways, such as Kdo2-lipid A and UDP-glucose-derived O-antigen, points to increased Gram-negative bacterial components, which may drive systemic inflammation implicated in both venous and arterial thrombosis26. The presence of energy metabolism pathways, including multiple TCA cycle variants and glyoxylate bypass routes, may reflect microbial adaptation to hypoxic or inflammatory environments typical in thrombosis23,27. Conversely, enrichment of NAD biosynthesis pathways (e.g., NAD salvage pathway I and de novo synthesis from aspartate) in healthy controls may indicate reduced inflammation and oxidative stress, contributing to vascular stability28,29. Tetrahydrofolate (THF) biosynthesis pathways, including both de novo synthesis and salvage routes, support one-carbon metabolism and the remethylation of homocysteine to methionine30. Their upregulation may reduce circulating homocysteine, a known risk factor for both arterial and venous thrombosis31. Additionally, enrichment of polyamine biosynthesis may suggest an anti-inflammatory role with potential cardiovascular protection32. On the other hand, predicted microbial functional pathways showed only limited differences between DVT and CAD patients (Supplementary Fig. 5), suggesting that the two groups share broadly similar gut microbial functional profiles. To assess the presence of gut-derived bacterial LPS, we conducted liquid chromatography–tandem mass spectrometry (LC–MS/MS) analysis on fecal samples and observed a consistent upward shift of two lipid A metabolites, 3-hydroxydodecanoic acid and 3-hydroxytetradecanoic acid, in DVT patients compared to healthy controls (Supplementary Fig. 7A). Gut microbial dysbiosis, in combination with inflammation, can increase intestinal permeability, promoting the translocation of bacterial products such as LPS into the systemic circulation, thereby contributing to atherothrombotic processes26. In response to LPS, LPS-binding protein (LBP), primarily produced by the liver, serves as a biomarker of systemic inflammation. To assess this, plasma levels of LBP were measured using enzyme-linked immunosorbent assay (ELISA). ELISA results revealed that plasma levels of LBP were significantly higher in DVT patients (16.84 ± 0.32 μg/mL, n = 58) compared to those with CAD (10.77 ± 1.26 μg/mL, n = 46), although levels in both groups exceeded the normal range (LBP: 5–10 μg/mL) (Supplementary Fig. 7B). While these pilot validation steps are preliminary, the concordance observed between inferred pathways and measured biomarkers suggests that these microbial signatures likely reflect underlying metabolic alterations. Interestingly, a recent study suggests that elevated plasma LBP levels can predict the risk of venous thrombosis in women33. However, in this study, we found no significant difference in plasma LBP levels between male and female participants. LBP is not disease-specific and is known to increase across diverse inflammatory, infectious, and metabolic conditions34; however, within the cardiovascular disease spectrum, elevated LBP appears to mark a distinct phenotype that is more prominent in DVT than CAD. Although LBP could theoretically enhance discriminative performance as a biomarker, it was not incorporated into the random forest model for DVT versus non-DVT (HC + CAD) because LBP measurements were unavailable for healthy controls. This hypothesis highlights an important avenue for future investigation. A major limitation of this study is the potential for residual confounding due to the composition of the control group. The non-DVT group primarily consists of healthy individuals who differ substantially from DVT patients in terms of age, comorbidities, and overall clinical context. Consequently, the identified microbial signatures may partly reflect generalized dysbiosis associated with chronic illness or aging, rather than being uniquely specific to the pathophysiology of DVT. Although we minimized hospital-related confounding by collecting samples during the early stage of admission, we recognize that unmeasured confounders remain a potential source of variation. Moreover, while we controlled for several key clinical variables, certain pre-analytical factors, such as specific dietary patterns, bowel habits, inflammatory state, and the use of medications like proton pump inhibitors, metformin, or statins, were not systematically documented for all participants. Furthermore, this study is limited by the lack of external validation in larger, multi-center, and multi-ethnic cohorts. Future research employing a prospective design with comprehensive clinical phenotyping is essential to validate these findings and ensure their robustness and reproducibility across diverse populations. Nevertheless, we employed internal validation using nested cross-validation to minimize optimistic bias, and applied permutation importance to assess feature robustness and model interpretability. Notably, the scarcity of publicly available gut microbiome datasets for DVT highlights an unmet need for data sharing and collaborative efforts in this area. Conclusion Our integrated microbial-clinical model demonstrated incremental discriminative value over a clinical-only model under a nested cross-validation framework designed to reduce overfitting and limit information leakage. Notably, permutation importance analysis identified microbial features as eight of the top ten signatures, outperforming established clinical variables. This suggests that the gut microbiome serves as a useful biological marker rather than a mere proxy for clinical history. Furthermore, the improved net benefit in DCA suggests that microbiome-derived signals provide complementary biological information, which enhances the discriminative characterization of DVT cases within this study population. Supplementary Information Supplementary Information 1. Supplementary Information 2. Supplementary Information 3. Supplementary Information 4. Acknowledgements The authors thank Dr. Wei-Chun Wang from the Artificial Intelligence Center, China Medical University Hospital for his kind consultation. Author contributions Conceptualization: H.-Y. Lai and C.-R. Lu; methodology: C.-R. Lu, W.-P. Hu, C.-H. Hsu, C.-C. Wang, S.-S. Chang, Y.-N. Lin, Y.-K. Lin, W.-H. Chung, K.-C. Chang, J.-M. Wang, D.-Y. Cho, Y.-Y. Hsu, H.-H. Tsai, Z.-L. Lai, Y.-J. Wang, and P.-R. Hsueh; investigation: C.-R. Lu, W.-P. Hu, C.-H. Hsu, C.-C. Wang, S.-S. Chang, Y.-N. Lin, Y.-K. Lin, W.-H. Chung, K.-C. Chang, J.-M. Wang, D.-Y. Cho, Y.-Y. Hsu, H.-H. Tsai, Z.-L. Lai, Y.-J. Wang, and P.-R. Hsueh; formal analysis: C.-R. Lu, W.-P. Hu, C.-H. Hsu, C.-C. Wang, S.-S. Chang, Y.-N. Lin, Y.-K. Lin, W.-H. Chung, K.-C. Chang, J.-M. Wang, D.-Y. Cho, Y.-Y. Hsu, H.-H. Tsai, Z.-L. Lai, Y.-J. Wang, and P.-R. Hsueh; resources: J.-M. Wang, D.-Y. Cho, Y.-Y. Hsu, H.-H. Tsai, Z.-L. Lai, Y.-J. Wang, and P.-R. Hsueh; validation: C.-R. Lu, W.-P. Hu, C.-H. Hsu, C.-C. Wang, S.-S. Chang, Y.-N. Lin, Y.-K. Lin, W.-H. Chung, and K.-C. Chang; visualization: C.-R. Lu, W.-P. Hu, C.-H. Hsu, C.-C. Wang, S.-S. Chang, Y.-N. Lin, Y.-K. Lin, W.-H. Chung, and K.-C. Chang; writing—original draft: H.-Y. Lai and C.-R. Lu; writing—review & editing: H.-Y. Lai; funding acquisition: H.-Y. Lai and C.-R. Lu; supervision: H.-Y. Lai. All the authors contributed to the article and approved the submitted version. Funding This work was financially supported by the China Medical University grants CMU114-MF-94 and CMU113-N-11, the China Medical University Hospital grant DMR-113-007, and the National Science and Technology Council grant NSTC114-2314-B-039-053-MY2. Data availability The authors confirm that the data supporting the findings of this study are available within the article and its supplementary materials. All sequencing data used in this manuscript have been deposited in the National Center for Biotechnology Information (NCBI) Sequence Read Archive under project number PRJNA1290843 (https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1290843). Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate This study was approved by the Institutional Review Board of China Medical University Hospital (IRB No. CMUH112-REC2-087). All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and national research committees and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. Written informed consent was obtained from all participants prior to sample collection. Patients or the public were not involved in the design, conduct, reporting, interpretation, or dissemination of this research. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Information 1. Supplementary Information 2. Supplementary Information 3. Supplementary Information 4. Data Availability Statement The authors confirm that the data supporting the findings of this study are available within the article and its supplementary materials. All sequencing data used in this manuscript have been deposited in the National Center for Biotechnology Information (NCBI) Sequence Read Archive under project number PRJNA1290843 (https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1290843).