Abstract
Aim:
To develop machine learning (ML) models integrating clinical and microbial predictors for early pancreatic cancer (PC) detection.
Background:
Pancreatic cancer is a leading cause of cancer-related mortality, with a 5-year survival rate of ~12%. Limited biomarkers and non-specific risk factors hinder early diagnosis. Emerging evidence links oral and gut microbiota, such as Fusobacterium nucleatum and Roseburia species, to PC risk, offering potential for non-invasive biomarkers.
Methods:
We analyzed a retrospective cohort of 40 participants (20 PC cases, 20 controls). Clinical (e.g., age, WBC) and microbial (e.g., Fusobacterium nucleatum, Roseburia-to-Fusobacterium ratio [RI/FN]) predictors were evaluated using five ML classifiers (logistic regression, SVM, random forest, naïve Bayes, neural network) under Leave-Group-Out Cross-Validation (LGOCV; 80/20 split, 200 repetitions). Elastic-net regularization and stability selection identified key predictors. Performance metrics included AUC, sensitivity, specificity, PPV, NPV, and accuracy. Nomograms were developed for clinical utility.
Results:
Age (AUC 97.4%) and microbial markers (e.g., RI/FN ratio, AUC 100%) showed excellent discrimination. Multivariable models using age and RI/FN achieved excellent performance (AUC 98–100%). Nomograms provided interpretable risk estimates.
Conclusions:
Integrating clinical and microbial predictors with ML offers a promising approach for non-invasive PC detection. The RI/FN ratio and age are robust biomarkers that warrant further validation in larger cohorts. However, the small sample size limits generalizability and warrants validation in larger cohorts.
Key Words: Pancreatic cancer, Machine learning, Microbiome, Biomarkers, Early detection
Introduction
Pancreatic cancer (PC) remains a formidable global health challenge, characterized by its insidious onset and poor prognosis (1, 2). In 2022, PC accounted for approximately 510,992 new cases and 467,409 deaths worldwide, reflecting its high lethality. Despite advances in therapeutic strategies and tumor biology, the 5-year survival rate in the United States hovers at approximately 12%, with PC ranking as the third leading cause of cancer-related mortality and projected to become the second by 2030 (1, 3). This persistent burden underscores the urgent need for improved early detection methods.
PC is characterized by a complex and multifactorial etiology, with well-established risk factors encompassing advanced age, chronic pancreatitis, long-standing diabetes, and lifestyle elements such as smoking and obesity (4). The non-specific nature of these risks, as they are prevalent in numerous other conditions and in the general population, renders them impractical as criteria for effective large-scale screening initiatives. Consequently, diagnosis is typically initiated only upon the manifestation of symptoms, which often signify advanced, unresectable disease (5). Current diagnostic paradigms rely heavily on imaging techniques. Yet, this such as contrast-enhanced computed tomography (CT), yet this modality is often inadequate for detecting small or early-stage tumors (6). Furthermore, the sole FDA-approved biomarker, carbohydrate antigen 19-9 (CA19-9), is limited by suboptimal sensitivity and specificity, confining its primary use to disease monitoring rather than initial detection (7).
Emerging research has identified a compelling link between microbial dysbiosis and PC risk, particularly involving oral and gut microbiota (8). Bacteria, such as Fusobacterium nucleatum and Porphyromonas gingivalis, have been associated with elevated PC risk (8, 9). Animal models have demonstrated that F. nucleatum accelerates pancreatic tumorigenesis, supporting its role in disease progression (10). Similarly, gut microbiota, including Roseburia intestinalis and Bifidobacterium species, are increasingly recognized for their diagnostic and prognostic potential, with elevated levels observed in PC patients and correlated with tumor characteristics (9).
Since the importance of microbial biomarkers is now widely recognized, using computational approaches can help further improve diagnostic accuracy. Artificial intelligence (AI), particularly machine learning (ML) and transfer learning (TL), is revolutionizing precision medicine by uncovering complex patterns in large-scale biological datasets (11, 12). ML algorithms excel at integrating diverse data types, such as clinical parameters, laboratory values, and microbiome profiles, to identify novel biomarkers with enhanced sensitivity and specificity (12). To date, few studies have integrated both clinical parameters and microbiome-derived features using machine learning for early PC detection. In this study, we leverage ML to analyze clinical and microbial predictors, focusing on oral and fecal microbiota, to develop predictive models for PC. By combining these data, we aim to create a robust, non-invasive framework for early detection and risk stratification, offering new insights into disease mechanisms and potential therapeutic targets for precision oncology. Given the rapid advances in artificial intelligence and machine learning, along with the high accuracy and significant impact of these approaches, this topic is expected to attract greater research attention from investigators in the coming years. Moreover, the lack of interdisciplinary and integrative studies in this area underscores the need for greater focus on this research.
Methods
Study design and data
Our study utilized a previously published retrospective dataset of 40 participants, with 20 diagnosed with PC and 20 serving as controls. The dataset analyzed in this study was originally generated and published in our prior work and is to develop predictive models for early PC detection (9). The primary outcome was a binary variable, "Group," with the Control group designated as the reference level.
Statistical analysis
Pre-processing
All predictors except sex were treated as continuous, while sex was handled as a binary factor (0/1). Missing values in constant variables were imputed using random forest multiple imputation. Predictors with fewer than two distinct values were discarded, and near-zero variance predictors were removed. Continuous variables were normalized by centering and scaling, while sex remained unchanged. For ROC-based modeling, the outcome was relabeled so that PC was explicitly defined as the positive class.
Model development and internal validation
Cross-validation for model tuning
Given the small sample size, we used Leave-Group-Out Cross-Validation (LGOCV) with 80/20 splits repeated 200 times. Each repetition produced out-of-fold (OOF) predictions for performance aggregation and plotting. Within caret, the summary function was used with the area under the ROC curve (AUC) as the primary metric.
Final performance estimation
To reduce optimism in the final performance of the logistic model, we applied the .632 bootstrap with 1000 resamples. Results from the bootstrap (AUC, sensitivity, specificity, positive predictive value, negative predictive value, and accuracy) are reported as the primary internal validation estimates.
Candidate models
We trained five classifiers on the selected predictor sets, including logistic regression, support vector machine with a radial kernel, random forest, naïve Bayes, and a neural network with a single hidden layer. All models were tuned using the LGOCV scheme, and out-of-fold predictions were retained to compute performance metrics and generate ROC and Precision–Recall curves.
Discrimination thresholds and performance metrics
For each resample, we computed a Youden-optimal threshold on the ROC curve to derive confusion-matrix metrics. The Youden index is defined as:
Youden index = sensitivity + specificity – 1.
At the threshold that maximized this index, we calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy. For each metric across resamples, we reported the mean and a 95% confidence interval using a t-based interval on the resample distribution. AUC was also computed per resample. Out-of-fold predictions were used to produce combined ROC curves and combined Precision–Recall curves.
Univariate analysis
To quantify under each variable's standalone predictive ability, using the LGOCV 80/20 × 200 scheme. For each variable, we reported AUC, sensitivity, specificity, PPV, NPV, and accuracy (mean and 95% CI). Continuous predictors were centered and scaled; sex was modeled as a factor.
Elastic-net modeling and stability selection
We used elastic-net logistic regression for feature screening, tuning hyperparameters over a grid of α values from 0.0 to 1.0 (in steps of 0.1) and λ on a logarithmic scale under the LGOCV scheme. Variables with non-zero coefficients at the best-tuned combination of α and λ were retained. To identify stable predictors, we performed bootstrap stability selection with 1000 resamples. In each resample, an elastic-net was fit with α fixed at the value selected during tuning, the model was extracted at the λ corresponding to the one-standard-error rule, and predictors with non-zero coefficients were recorded.
We computed each variable’s selection frequency (the proportion of bootstraps with a non-zero coefficient). Variables with frequency ≥ 0.60 were retained, subject to the events-per-variable (EPV) rule: with 20 PC events, we capped the model at 2 predictors (20/10 = 2). If more variables exceeded the threshold, we kept the highest-frequency ones; if none exceeded it, we kept the top-ranked variables up to the EPV cap. This stability-selected set was used to train the multivariable model ensemble (GLM, SVM, RF, NB, NNET) and to build the primary nomogram.
“Key markers” subset
In addition to the stability-selected model, we constructed a compact clinically oriented subset (“key markers”) by ranking predictors with a composite score that averaged their ranks across three criteria: stability selection frequency, univariate AUC, and the absolute value of elastic-net coefficients at the best λ. The top three markers (modifiable to two, depending on the events-per-variable constraint) were selected, and the model ensemble was refit on this subset. Performance was again summarized under the LGOCV scheme, and a second nomogram was developed using these key markers.
Nomograms
Logistic regression models were fitted for both the stability-selected predictors and the key-marker subset. Nomograms were constructed with probability scales spanning a broad range, with ticks from 0.05 to 0.95 for the key-marker model. Internal validation was performed using bootstrap resampling with 1000 repetitions to obtain optimism-corrected indices and calibration curves.
Visualization
ROC and Precision–Recall curves were produced from the out-of-fold predictions.
Outputs and reproducibility
The analysis was implemented in R (version 4.5.0; seed fixed at 20250917), and the full R code is provided in the Supplementary Materials.
Results
Univariate analysis
Table 1 summarizes the discriminative performance of individual clinical and microbial predictors for distinguishing pancreatic cancer (PC) from controls. Among clinical variables, Age achieved the highest accuracy with an AUC of 97.41% (95% CI: 96.72–98.10), sensitivity of 95.75%, and specificity of 98.62%. Several hematological markers, including hemoglobin, hematocrit, and RBC count, also showed moderate discriminatory ability (AUCs 78–83%). In contrast, variables such as WBC, MCV, and MCHC demonstrated limited predictive value (AUCs ≤ 55%).
Table 1.
Univariate classification performance of individual clinical and microbial predictors for discriminating pancreatic cancer (PC) from controls.
| Variable | AUC (95% CI) | Sens (95% CI) | Spec (95% CI) | PPV (95% CI) | NPV (95% CI) | ACC (95% CI) |
|---|---|---|---|---|---|---|
| Age | 97.41 (96.72, 98.10) | 95.75 (94.44, 97.06) | 98.62 (97.83, 99.42) | 98.90 (98.26, 99.54) | 96.60 (95.55, 97.65) | 97.19 (96.46, 97.92) |
| WBC (count/L) | 52.66 (50.03, 55.28) | 72.62 (69.12, 76.13) | 59.88 (55.51, 64.24) | 71.33 (68.72, 73.94) | 74.05 (70.73, 77.37) | 66.25 (64.70, 67.80) |
| RBC (count/L) | 78.81 (76.53, 81.10) | 82.62 (80.08, 85.17) | 87.25 (84.75, 89.75) | 89.79 (87.96, 91.61) | 86.37 (84.54, 88.20) | 84.94 (83.54, 86.34) |
| Hemoglobin (g/dl) | 81.03 (79.12, 82.95) | 89.88 (87.75, 92.00) | 75.62 (72.80, 78.45) | 81.70 (79.79, 83.60) | 91.45 (89.72, 93.18) | 82.75 (81.42, 84.08) |
| Hematocrit (%) | 82.92 (80.88, 84.97) | 91.38 (89.54, 93.21) | 78.25 (75.22, 81.28) | 83.90 (81.88, 85.93) | 92.37 (90.75, 93.98) | 84.81 (83.31, 86.31) |
| MCV (fL) | 36.31 (34.25, 38.37) | 73.38 (70.08, 76.67) | 41.75 (37.70, 45.80) | 59.87 (57.69, 62.05) | 63.05 (58.83, 67.26) | 57.56 (56.06, 59.06) |
| MCH (pg) | 41.20 (38.47, 43.93) | 82.62 (79.64, 85.61) | 39.38 (35.77, 42.98) | 60.44 (58.48, 62.40) | 73.10 (68.66, 77.54) | 61.00 (59.28, 62.72) |
| MCHC (g/dl) | 40.53 (37.92, 43.14) | 62.50 (58.89, 66.11) | 58.88 (54.03, 63.72) | 68.65 (65.73, 71.57) | 61.43 (57.49, 65.38) | 60.69 (59.02, 62.36) |
| RDW | 66.97 (64.12, 69.81) | 69.75 (66.99, 72.51) | 89.12 (86.29, 91.96) | 90.61 (88.44, 92.78) | 76.12 (73.83, 78.41) | 79.44 (77.74, 81.14) |
| Platelet count | 57.80 (54.98, 60.62) | 71.50 (68.11, 74.89) | 68.25 (63.90, 72.60) | 76.44 (73.75, 79.13) | 74.58 (71.35, 77.80) | 69.88 (68.11, 71.64) |
| AST (IU/L) | 81.28 (79.20, 83.36) | 84.75 (82.24, 87.26) | 81.12 (78.15, 84.10) | 85.50 (83.43, 87.56) | 87.84 (85.96, 89.73) | 82.94 (81.47, 84.40) |
| ALT (IU/L) | 66.91 (64.23, 69.58) | 78.38 (74.96, 81.79) | 72.25 (68.24, 76.26) | 80.29 (77.56, 83.02) | 84.21 (81.82, 86.60) | 75.31 (73.78, 76.84) |
| AST/ALT Ratio (De Ritis ratio) | 58.81 (56.23, 61.39) | 77.38 (73.98, 80.77) | 62.88 (59.26, 66.49) | 72.32 (70.11, 74.54) | 79.81 (77.03, 82.59) | 70.12 (68.58, 71.67) |
| AlkP (IU/L) | 74.75 (72.37, 77.13) | 77.38 (74.59, 80.16) | 83.75 (80.65, 86.85) | 87.25 (85.04, 89.46) | 82.25 (80.24, 84.26) | 80.56 (78.96, 82.16) |
| Total Bilirubin (mg/dl) | 86.25 (84.31, 88.19) | 85.25 (82.80, 87.70) | 92.00 (89.74, 94.26) | 94.04 (92.46, 95.62) | 88.80 (87.04, 90.56) | 88.62 (87.26, 89.99) |
| Direct Bilirubin (mg/dL) | 86.48 (84.73, 88.24) | 78.38 (75.70, 81.05) | 96.12 (94.15, 98.10) | 97.48 (96.23, 98.72) | 84.21 (82.42, 86.00) | 87.25 (85.85, 88.65) |
| Amylase (U/L) | 47.48 (44.52, 50.45) | 54.87 (51.46, 58.29) | 76.38 (72.16, 80.59) | 78.43 (75.20, 81.66) | 62.28 (59.34, 65.22) | 65.62 (63.68, 67.57) |
| Lipase (U/L) | 71.72 (69.04, 74.40) | 71.12 (68.24, 74.01) | 92.00 (89.85, 94.15) | 92.79 (90.98, 94.60) | 78.49 (76.57, 80.41) | 81.56 (79.99, 83.13) |
| PT (second) | 73.08 (70.31, 75.85) | 70.00 (67.18, 72.82) | 99.00 (97.75, 100.00) | 99.38 (98.63, 100.00) | 78.40 (76.49, 80.31) | 84.50 (83.00, 86.00) |
| PTT (second) | 89.89 (88.04, 91.74) | 87.12 (84.92, 89.33) | 99.12 (98.48, 99.77) | 99.25 (98.70, 99.80) | 90.15 (88.53, 91.77) | 93.12 (92.00, 94.25) |
| Neisseria elongata | 96.77 (95.81, 97.73) | 95.75 (94.44, 97.06) | 98.62 (97.83, 99.42) | 98.80 (98.10, 99.50) | 96.50 (95.41, 97.59) | 97.19 (96.38, 98.00) |
| Granulicatella adiacens | 98.75 (98.24, 99.26) | 98.50 (97.67, 99.33) | 98.38 (97.51, 99.24) | 98.70 (98.01, 99.39) | 98.80 (98.14, 99.46) | 98.44 (97.86, 99.02) |
| Fusobacterium nucleatum | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) |
| Roseburia intestinalis | 98.50 (97.86, 99.14) | 98.25 (97.12, 99.38) | 97.75 (96.75, 98.75) | 98.20 (97.40, 99.00) | 98.73 (97.93, 99.53) | 98.00 (97.27, 98.73) |
| Bifidobacterium bifidum | 81.41 (79.52, 83.29) | 87.88 (85.67, 90.08) | 81.38 (78.58, 84.17) | 85.92 (83.98, 87.86) | 90.16 (88.44, 91.88) | 84.62 (83.31, 85.94) |
| Fusobacterium-to-Bifidobacterium Ratio (FN/BF) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) |
| Roseburia-to-Fusobacterium Ratio (RI/FN) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) |
| Gender | 42.81 (40.90, 44.72) | 37.12 (33.92, 40.33) | 48.50 (44.58, 52.42) | 41.52 (38.75, 44.28) | 40.35 (37.85, 42.84) | 42.81 (40.90, 44.72) |
Note: Metrics include the area under the receiver operating characteristic curve (AUC), sensitivity (Sens), specificity (Spec), positive predictive value (PPV), negative predictive value (NPV), and overall accuracy (ACC), each with 95% confidence intervals. Results were obtained from univariate logistic regression models fit separately for each predictor. Performance estimates were derived using repeated 80/20 leave-group-out cross-validation (LGOCV) with 200 repetitions. Out-of-fold predictions from each repetition were aggregated, and optimal decision thresholds were chosen by the Youden index to compute sensitivity and specificity.
Among microbial markers, Fusobacterium nucleatum, the Fusobacterium-to-Bifidobacterium ratio (FN/BF), and the Roseburia-to-Fusobacterium ratio (RI/FN) achieved perfect discrimination (AUC = 100.0%, 95% CI: 100.0–100.0) with flawless sensitivity, specificity, PPV, NPV, and accuracy. Other taxa, including Granulicatella adiacens (AUC 98.75%) and Roseburia intestinalis (AUC 98.50%), also exhibited excellent individual performance.
Multivariable models: elastic-net selected predictors
Predictors selected through elastic-net regularization included Age and WBC. Multivariable classifiers trained on these variables demonstrated strong performance across all methods (Table 2). Logistic regression (GLM) achieved an AUC of 99.03% (95% CI: 98.35–99.71), with 100% sensitivity and 98% specificity. Other models, including SVM, RF, NB, and NNET, performed comparably, with AUCs ranging from 97.14% to 99.75%.
Table 2.
Multivariable classification performance based on predictors selected through elastic-net regularization (Age and WBC).
| Model | AUC (95% CI) | Sens (95% CI) | Spec (95% CI) | PPV (95% CI) | NPV (95% CI) | ACC (95% CI) |
|---|---|---|---|---|---|---|
| GLM | 99.03 (98.35, 99.71) | 100.00 (100.00, 100.00) | 98.00 (96.63, 99.37) | 98.65 (97.76, 99.55) | 100.00 (100.00, 100.00) | 99.00 (98.32, 99.68) |
| SVM | 98.91 (98.46, 99.36) | 99.00 (98.32, 99.68) | 97.88 (96.90, 98.85) | 98.30 (97.52, 99.08) | 99.20 (98.65, 99.75) | 98.44 (97.86, 99.02) |
| RF | 97.14 (96.37, 97.91) | 94.62 (93.19, 96.06) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 95.70 (94.55, 96.85) | 97.31 (96.59, 98.03) |
| NB | 99.59 (99.36, 99.83) | 94.62 (93.19, 96.06) | 98.62 (97.83, 99.42) | 98.90 (98.26, 99.54) | 99.90 (99.70, 100.00) | 99.25 (98.84, 99.66) |
| NNET | 99.75 (99.58, 99.92) | 94.62 (93.19, 96.06) | 99.00 (98.32, 99.68) | 99.20 (98.65, 99.75) | 100.00 (100.00, 100.00) | 99.50 (99.16, 99.84) |
Note: Performance metrics include the area under the receiver operating characteristic curve (AUC), sensitivity (Sens), specificity (Spec), positive predictive value (PPV), negative predictive value (NPV), and overall accuracy (ACC), each reported with 95% confidence intervals. Models were trained using Age and white blood cell count (WBC, per liter) as predictors. Estimates were obtained with repeated 80/20 leave-group-out cross-validation (200 repetitions). Compared classifiers include logistic regression (GLM), support vector machine with radial kernel (SVM), random forest (RF), naïve Bayes (NB), and a single-hidden-layer neural network (NNET).
Multivariable models: clinically prioritized key markers
Using clinically prioritized key markers, Age and the RI/FN ratio, all models achieved outstanding discriminatory accuracy (Table 3). The RF and NNET classifiers reached perfect prediction (AUC, sensitivity, specificity, PPV, NPV, and accuracy all 100%). GLM and SVM also yielded excellent performance, with AUCs exceeding 98%.
Table 3.
Multivariable classification performance based on clinically prioritized key markers (RI/FN ratio and Age).
| Model | AUC (95% CI) | Sens (95% CI) | Spec (95% CI) | PPV (95% CI) | NPV (95% CI) | ACC (95% CI) |
|---|---|---|---|---|---|---|
| GLM | 99.59 (99.29, 99.90) | 99.50 (99.01, 99.99) | 99.50 (99.01, 99.99) | 99.60 (99.21, 99.99) | 99.60 (99.21, 99.99) | 99.50 (99.16, 99.84) |
| SVM | 98.53 (97.99, 99.08) | 98.25 (97.36, 99.14) | 98.00 (97.05, 98.95) | 98.40 (97.64, 99.16) | 98.60 (97.89, 99.31) | 98.12 (97.50, 98.75) |
| RF | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) |
| NB | 97.81 (96.88, 98.74) | 97.75 (96.69, 98.81) | 98.62 (97.83, 99.42) | 98.90 (98.26, 99.54) | 98.23 (97.41, 99.05) | 98.19 (97.55, 98.83) |
| NNET | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) | 100.00 (100.00, 100.00) |
Note: Performance metrics include the area under the receiver operating characteristic curve (AUC), sensitivity (Sens), specificity (Spec), positive predictive value (PPV), negative predictive value (NPV), and overall accuracy (ACC), each reported with 95% confidence intervals. Models were trained using Age and the Roseburia-to-Fusobacterium ratio (RI/FN) as predictors, selected for their clinical relevance. Estimates were obtained with repeated 80/20 leave-group-out cross-validation (200 repetitions). Compared classifiers include logistic regression (GLM), support vector machine with radial kernel (SVM), random forest (RF), naïve Bayes (NB), and a single-hidden-layer neural network (NNET).
Combined ROC and PR curves
Figure 1 presents aggregated ROC and PR curves for all multivariable classifiers under repeated 80/20 leave-group-out cross-validation (200 repetitions). All models exhibited near-perfect classification performance, with overlapping ROC and PR profiles reflecting their consistently high discriminatory ability.
Figure 1.
Combined receiver operating characteristic (ROC) and precision–recall (PR) curves for multivariable classifiers under repeated 80/20 leave-group-out cross-validation (200 repetitions). The models compared include logistic regression (GLM), support vector machine with a radial kernel (SVM), random forest (RF), naïve Bayes (NB), and a single-hidden-layer neural network (NNET). Out-of-fold predictions from each repetition were aggregated to generate the curves.
Nomogram analysis
To enhance clinical interpretability, nomograms were constructed for visualizing the probability of PC based on selected predictors. The first nomogram (Figure 2, upper panel) was derived from the stability-selected predictors identified through elastic-net regularization and bootstrap stability analysis. The second nomogram (Figure 2, lower panel) was constructed from the clinically prioritized key markers, Age and the Roseburia-to-Fusobacterium ratio (RI/FN). Both nomograms provide individualized risk estimates on a probability scale, with the key-marker model offering a more compact and clinically oriented representation.
Figure 2.
Nomograms for predicting pancreatic cancer risk. The upper panel shows the model based on stability-selected predictors, and the lower panel shows the model based on clinically prioritized key markers (Age and RI/FN ratio).
Discussion
PC remains a lethal malignancy, largely due to the absence of reliable tools for early detection. The only FDA-approved biomarker, CA19-9, suffers from insufficient sensitivity and specificity for screening, creating a critical diagnostic void (13, 14). This diagnostic gap underscores the urgent need for novel, non-invasive biomarkers to enable earlier detection, similar to advances seen in breast cancer management through early screening (15, 16). Our study addresses this gap by demonstrating the profound potential of oral and gut microbiota as non-invasive biomarkers for PC. We show that specific microbial taxa and their ratios, integrated with clinical data using ML, can achieve exceptional discriminatory accuracy, paving the way for novel early detection strategies.
A key finding was the superior performance of microbial markers, particularly F. nucleatum and the Fusobacterium-to-Bifidobacterium (FN/BF) and Roseburia-to-Fusobacterium (RI/FN) ratios, which achieved an AUC of 1.0 in this dataset. This aligns with growing evidence of microorganisms like F. nucleatum may promote tumorigenesis by modulating the tumor microenvironment and fueling local inflammation (9). The fact that microbial ratios outperformed individual taxa underscores the importance of ecological dynamics within the microbiome; it is the shifting balance between species, rather than the presence or absence of a single organism, that provides a more comprehensive and powerful biomarker signature for PC.
In recent years, ML has emerged as a powerful tool for early PC detection, capable of integrating high-dimensional clinical and paraclinical data to uncover complex patterns beyond the reach of traditional statistical methods (17). In our study, ML was instrumental in distilling this complex, high-dimensional data into robust predictive models. Among all clinical variables, Age was the strongest standalone predictor (AUC 97.4%), consistent with established epidemiological risk profiles. However, most routine hematological and biochemical parameters showed only modest discriminatory value on their own. The integration of clinical and microbial data significantly enhanced classification performance. The elastic-net regularization model, utilizing only Age and White Blood Cell count (WBC), demonstrated near-perfect accuracy across multiple ML algorithms. Furthermore, the clinically prioritized key-marker model, based solely on Age and the RI/FN ratio, achieved perfect classification (AUC = 1.0) with both Random Forest and Neural Network models. The development of nomograms from these predictor sets provides a practical, interpretable tool for individualized risk assessment, enhancing their potential for clinical translation.
It is worth noting that microbiome sampling (oral or fecal) is non-invasive and increasingly feasible within clinical diagnostic workflows. The clinical implications of these findings are substantial. A model requiring only two variables, a readily available clinical parameter (Age) and a microbial ratio (RI/FN), to achieve such high accuracy suggests a path toward cost-effective, non-invasive screening tools. Such models could complement existing modalities, such as imaging and CA19-9 monitoring, facilitating earlier detection and timely intervention in high-risk populations. Previous research has shown that ML–microbiome integration provides strong diagnostic capabilities across multiple cancers. For instance, fecal microbiome-based ML classifiers have distinguished colorectal cancer and precancerous lesions from healthy controls (18). Similarly, ML models used in fecal and oral microbiome signatures have demonstrated high discriminatory power for detecting gastric cancer (19). These findings support the application of microbiome-driven machine learning approaches in cancer diagnosis.
Despite these promising results, our study has several limitations. First, the relatively small sample size, despite rigorous internal validation, necessitates caution regarding generalizability. External validation in larger, multi-center, and ethnically diverse cohorts is essential to confirm the robustness of these biomarkers. Second, the cross-sectional nature of our study design precludes causal inference. Longitudinal studies are required to determine whether the observed microbial shifts are a cause or a consequence of pancreatic carcinogenesis. Finally, reliance on relative abundance data from microbiome sequencing can introduce compositional biases. Future work would benefit from absolute quantification methods and integration with multi-omics data (e.g., metatranscriptomics, metabolomics) to elucidate functional mechanisms.
Conclusion
In conclusion, our study establishes that integrating host clinical data with gut microbiome signatures can generate highly accurate predictive models for PC. The combination of Age and the RI/FN ratio emerges as a particularly promising, clinically actionable biomarker set for risk stratification. Future research should focus on validating these signatures in independent cohorts, unraveling the mechanistic links between microbial dysbiosis and PC development, and integrating these biomarkers into multi-omic platforms for precision diagnosis. Ultimately, the synergy between host and microbiome data holds significant promise for improving early detection, patient prognosis, and personalized therapeutic strategies in PC.
Ethical issues
Research Ethics Committees of Islamic Azad Tehran Medical Sciences University - Pharmacy and Pharmaceutical Branches Faculty (No. IR.IAU.PS.REC.1402.369).
Acknowledgements
The authors would like to express their sincere gratitude to the esteemed members of the Research Institute for Gastroenterology and Liver Diseases at Shahid Beheshti University of Medical Sciences in Tehran, Iran.
Conflict of interests
The authors declare no conflict of interest.
References
- 1.Hesami Z, Olfatifar M, Sadeghi A, Zali MR, Mohammadi‐Yeganeh S, Habibi MA, et al. Global trend in pancreatic cancer prevalence rates through 2040: an illness‐death modeling study. Cancer Med. 2024;13:e70318. doi: 10.1002/cam4.70318. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68:394–424. doi: 10.3322/caac.21492. [DOI] [PubMed] [Google Scholar]
- 3.Leiphrakpam PD, Chowdhury S, Zhang M, Bajaj V, Dhir M, Are C. Trends in the global incidence of pancreatic cancer and a brief review of its histologic and molecular subtypes. J Gastrointest Cancer. 2025;56:71. doi: 10.1007/s12029-025-01183-2. [DOI] [PubMed] [Google Scholar]
- 4.Peduzzi G, Archibugi L, Farinella R, de Leon Pisani RP, Vodickova L, Vodicka P, et al. The exposome and pancreatic cancer, lifestyle and environmental risk factors for PDAC. Semin Cancer Biol . 2025;113:100–129. doi: 10.1016/j.semcancer.2025.05.004. [DOI] [PubMed] [Google Scholar]
- 5.Klein AP. Pancreatic cancer epidemiology: understanding the role of lifestyle and inherited risk factors. Nat Rev Gastroenterol Hepatol. 2021;18:493–502. doi: 10.1038/s41575-021-00457-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Hanna-Sawires RG, Schiphuis JH, Wuhrer M, Vasen HF, van Leerdam ME, Bonsing BA, et al. Clinical perspective on proteomic and glycomic biomarkers for diagnosis, prognosis, and prediction of pancreatic cancer. Int J Mol Sci. 2021;22:2655. doi: 10.3390/ijms22052655. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Lee T, Teng TZJ, Shelat VG. Carbohydrate antigen 19-9—Tumor marker: Past, present, and future. World J Gastrointest Surg. 2020;12:468. doi: 10.4240/wjgs.v12.i12.468. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Hesami Z, Sabzehali F, Khorsand B, Alipour S, Sadeghi A, Asri N, et al. Microbiota as a state-of-the-art approach in precision medicine for pancreatic cancer management: a comprehensive systematic review. iScience. 2025;28:112314. doi: 10.1016/j.isci.2025.112314. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Tavanaeian S, Feizabadi MM, Falsafi S, Aghdaei HA, Houri H. Oral and fecal microbiome alterations in pancreatic cancer: insights into potential diagnostic biomarkers. BMC Microbiol. 2025;25:624. doi: 10.1186/s12866-025-04344-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Hayashi M, Ikenaga N, Nakata K, Luo H, Zhong P, Date S, et al. Intratumor Fusobacterium nucleatum promotes the progression of pancreatic cancer via the CXCL1‐CXCR2 axis. Cancer Sci. 2023;114:3666–78. doi: 10.1111/cas.15901. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Khorsand B, Rajabnia M, Jahanian A, Fathy M, Taghvaei S, Houri H. Enhancing the accuracy and effectiveness of diagnosis of spontaneous bacterial peritonitis in cirrhotic patients: a machine learning approach utilizing clinical and laboratory data. Adv Med Sci. 2025;70:1–7. doi: 10.1016/j.advms.2024.10.001. [DOI] [PubMed] [Google Scholar]
- 12.Schattenberg JM, Chalasani N, Alkhouri N. Artificial intelligence applications in hepatology. Clin Gastroenterol Hepatol. 2023;21:2015–25. doi: 10.1016/j.cgh.2023.04.007. [DOI] [PubMed] [Google Scholar]
- 13.Rezaei-Tavirani M, Zamanian-Azodi M, Rajabi S, Masoudi-Nejad A, Rostami-Nejad M, Rahmatirad S. Protein clustering and interactome analysis in parkinson and alzheimer's diseases. Arch Iran Med. 2016;19:101–9. [PubMed] [Google Scholar]
- 14.Zhang Y, Yang J, Li H, Wu Y, Zhang H, Chen W. Tumor markers CA19-9, CA242 and CEA in the diagnosis of pancreatic cancer: a meta-analysis. Int J Clin Exp Med. 2015;8:11683. [PMC free article] [PubMed] [Google Scholar]
- 15.Half E, Keren N, Reshef L, Dorfman T, Lachter I, Kluger Y, et al. Fecal microbiome signatures of pancreatic cancer patients. Sci Rep. 2019;9:16801. doi: 10.1038/s41598-019-53041-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Farrell JJ, Zhang L, Zhou H, Chia D, Elashoff D, Akin D, et al. Variations of oral microbiota are associated with pancreatic diseases including pancreatic cancer. Gut. 2012;61:582–8. doi: 10.1136/gutjnl-2011-300784. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Khorsand B, Hesami Z, Alipour S, Farmani M, Houri H. Harnessing artificial intelligence for detection of pancreatic cancer: a machine learning approach. Clin Exp Med. 2025;25:228. doi: 10.1007/s10238-025-01761-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zeller G, Tap J, Voigt AY, Sunagawa S, Kultima JR, Costea PI, et al. Potential of fecal microbiota for early‐stage detection of colorectal cancer. Mol Syst Biol. 2014;10:766. doi: 10.15252/msb.20145645. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Rosero YC, Alzate LA, Parra Medina R. Insights on prevention and gastric cancer detection: an integrative approach through risk factors, microbiome, molecular markers and machine learning. Oncologie. 2025;27:643–58. [Google Scholar]