Machine learning-based clinical prediction model and multi-omics integration for assessing pancreatic cancer risk in new-onset diabetes

Abstract

Background

Given that pancreatic cancer (PC) is typically diagnosed at an advanced stage but is often preceded by new-onset diabetes mellitus (NODM), providing a window for early detection, we sought to develop and validate an interpretable machine-learning model integrated with multi-omics profiling to identify early biomarkers of NODM-associated PC.

Methods

In a population-based cohort, individuals with NODM-associated PC and NODM without PC were identified and randomly divided (70:30) into training and validation sets after feature selection. Eight machine learning (ML) classifiers were compared using fivefold cross-validation, and model performance was evaluated in terms of discrimination, calibration, and decision curve–based clinical utility. We evaluated interpretability using the Shapley additive explanations (SHAP) analyses. Mechanistically, Olink proteomic profiling and metabolomics were analyzed through clinical classifications and model-defined risk strata.

Results

Categorical boosting achieved the best performance in the independent validation set (AUROC = 0.844). The NODM cohort was stratified into high- (n = 2,362) and low-risk (n = 5,030) groups, and internal validation together with SHAP analyses demonstrated consistent model performance and identified clinically interpretable predictors. Proteomic and metabolomic analyses under clinical and risk-based grouping identified 39 overlapping differentially expressed proteins and 145 overlapping metabolites with enriched across 11 shared KEGG pathways. Cross-platform validation highlighted PLTP, CRTAC1, and ITGAV as serum biomarkers with a strong potential for early NODM-PC detection.

Conclusions

We developed an interpretable ML framework centered on NODM enables practical risk stratification for early PC detection by multi-omics and provides a pathway of ML-based triage followed by biomarker confirmation for earlier detection and diagnosis.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-07767-1.

Introduction

Pancreatic cancer (PC) is insidious, rapidly progressive, and prone to early metastasis [1]. Over 80% of patients present with advanced-stage disease at diagnosis, contributing to a global 5-year overall survival rate of approximately 13% [2, 3]. This burden continues to rise and is projected to become the second leading cause of cancer-related mortality in several high-Human Development Index countries [4, 5]. Despite advances in surgery, with neoadjuvant or adjuvant therapy, systemic chemotherapy, and immuno- or targeted therapies, overall long-term survival has improved only marginally: merely 15–20% of patients qualify for potentially curative resection, and recurrence after complete resection remains frequent, indicating a therapeutic plateau [6–10]. These realities underscore the need for a “stage shift” through earlier detection and precision screening to improve outcomes, enable individualized therapy, and increase the likelihood of cure.

PC is closely linked to disordered glucose metabolism, and new-onset diabetes mellitus (NODM) is recognized as a key early clinical signal and practical screening filter for identifying at-risk, asymptomatic individuals [11–13]. In recent years, the risk stratification models for the NODM population are developing rapidly, dysglycemia typically develops within 3 years before PC diagnosis, and approximately half of patients with resectable or small-volume tumors exhibit impaired glucose tolerance or diabetes [14–16]. Among individuals with NODM, the 3-year cumulative incidence of PC after NODM was 0.51%, corresponding to a 4.6-fold higher risk based on the standardized incidence ratio [17, 18]. Although genetic and multi-omics studies have improved risk modeling, biologically interpretable, clinically validated biomarkers and actionable targets remain limited [19–22]. Thus, using NODM as a high-yield entry point to integrate interpretable machine learning (ML) with multi-omics discovery represents a practical pathway for developing clinically meaningful biomarkers and enabling earlier, individualized PC detection.

Herein, we integrated the clinical features of PC arising in the context of NODM and developed an interpretable ML model to identify early PC within NODM populations. We then coupled this model with integrative proteomic and metabolomic analyses to identify serum-detectable biomarkers and delineate a practical, biomarker-assisted diagnostic strategy for the early detection of PC associated with NODM (NODM-PC).

Methods

The overall workflow of this study is presented in Fig. 1.

Fig. 1.

Schematic diagram of study design and model-development workflow. The analysis utilized a UK Biobank cohort, followed by clinical data extraction and feature screening and selection. Candidate models were trained and validated to determine the classifier with the best performance for identifying NODM-PC. Finally, integrative multi-omics analyses and clinical assays were used to nominate and validate serum biomarkers for PC early detection

Study design and population

This study utilized data from the UK Biobank (UKB), a population-based prospective cohort comprising approximately 500,000 participants aged 37–73 years recruited between 2006 and 2010. At baseline, the participants completed demographic and lifestyle questionnaires, underwent anthropometric assessments, and provided biological samples. Outcomes were ascertained longitudinally via linkage to hospital episode statistics, national cancer registries, and death records. The UK NHS Research Ethics Committee approved the study (21/NW/0157), and all participants provided written informed consent. Follow-up was completed on May 8, 2024. PC cases were identified using International Classification of Diseases, 10th Revision (ICD-10)/ International Classification of Diseases, 9th Revision (ICD-9) codes from hospital admissions and national registries, supplemented with self-reported cancer illness codes. Diabetes was ascertained using ICD-10/ICD-9 hospital codes, self-reported non-cancer illness codes, and self-reported physician diagnoses (Table S1). Following previous work, unspecified diabetes with onset after 35 years of age was classified as type 2 diabetes mellitus (T2DM) [23]. NODM was defined as the first T2DM diagnosis occurring within 3 years before PC diagnosis or cohort baseline. To ensure temporal precedence and mitigate reverse causation arising from simultaneous detection, patients with synchronous diagnoses, defined as a diagnosis of T2DM and PC on the same date, were excluded from the sensitivity analysis. Individuals with missing diagnostic dates for PC or diabetes were excluded.

Variables and candidate predictors

From the UKB, 229 variables encompassing clinical history, biochemistry, anthropometry, and lifestyle factors were extracted (Tables S2 and S3). We compared individuals with PC in the NODM-PC setting with three reference groups: PC with long-term diabetes mellitus (LTDM-PC), PC without diabetes mellitus (nonDM-PC), and T2DM without cancer. Seventeen variables showed consistent directions of effects across comparisons. Combined with five core covariates (age, sex, ethnicity, smoking, and alcohol use), these variables yielded 22 candidate predictors (Table S4). Continuous variables were standardized, categorical variables were one-hot encoded, and missing data were addressed using multiple imputations.

Feature selection, model development, and validation

The primary objective was to distinguish NODM-PC from NODM without PC (NODM-nonPC). Participant data were randomly divided into training and validation sets in a 7:3 ratio. Continuous features were standardized using z-scores, and categorical variables were one-hot encoded. Seven ML algorithms were used to build models, including categorical boosting (CatBoost), Gaussian naive Bayes (GaussianNB), light gradient-boosting machine (LightGBM), linear support vector machine (Linear SVM), logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost). We applied class weighting during model training to avoid potential bias from synthetic samples. For model development, we employed a grid search strategy combined with stratified five-fold cross-validation for hyperparameter tuning. The top three feature combinations and corresponding hyperparameters from the top-performing models (LR, Linear SVM, XGBoost, LightGBM, and CatBoost) were selected as the base models for stacked ensemble (Table S5).

RF was used as the meta-classifier, which further processed the aggregated results of the base models to construct the stacked ensemble algorithm model. Within the training set, fivefold cross-validation was used for hyperparameter optimization. The model with the highest area under the receiver operating characteristic curve (AUROC) was selected for each algorithm. After determining the base models, the hyperparameters of the RF meta-classifier were further optimized using the grid search method to identify the best configuration. Validation-set performance was assessed using the receiver operating characteristic (ROC) curve, accuracy, balanced accuracy, sensitivity, specificity, and negative predictive value (NPV). Bootstrap resampling was used to estimate 95% confidence intervals were estimated for all performance metrics in both the training and validation sets. Calibration curves were used to evaluate the agreement between predicted and observed risks, and decision curve analysis (DCA) was used to quantify the net clinical benefit across probability thresholds. A 14-variable CatBoost model demonstrated the best performance and was selected as the final classifier. Feature importance in the final model was interpreted using Shapley additive explanations (SHAP) to enhance its interpretability.

Proteomics data analysis

Protein biomarker data were obtained from independent sample collections of approximately 55,000 participants in the UKB. The samples were tested using the Olink Explore platform, which quantifies up to 2,923 proteins across four panels: cardiometabolism, inflammation, neurology, and oncology. Quality control procedures included outlier detection and data normalization. Protein abundance was expressed as log2-scaled normalized protein expression values. Proteins with more than 20% missing data were excluded, and remaining missing values were imputed using multivariate imputation by chained equations (MICE). The final analytical dataset included 2,912 proteins. For case-control analyses, 1,459 proteins were available in the NODM-PC group and 2,912 in the NODM-nonPC group, with 1,458 proteins shared across both groups. Differential expression analysis was performed using Student’s t-test, and proteins with a p-value < 0.05 and |log₂ fold change (FC)| > 0 were identified as significantly differentially expressed.

Metabolomics data analysis

Metabolite biomarker data were derived from plasma samples of approximately 118,000 participants in the UKB using a high-throughput nuclear magnetic resonance (NMR) platform (Nightingale). Plasma samples were collected in EDTA tubes at baseline and during follow-up and analyzed in two phases between 2019 and 2020. The Nightingale platform quantified 251 circulating metabolite biomarkers (including glucose, lactic acid, and spectrometer-corrected alanine), comprising 170 metabolites expressed in absolute concentrations (mmol/L) and 81 ratio indicators. These biomarkers traverse multiple pathways, including 14 lipoprotein subclasses, fatty acids, amino acids, ketone bodies, and glycolysis-related metabolites. Rigorous quality control was implemented during sample processing and analysis, and only samples and metabolite data meeting certified quality standards were retained. The proportion of missing values for the 251 metabolites was ≤ 20%, and the missing data were imputed using MICE. Differential analysis was performed using the Wilcoxon rank-sum test, and metabolites with a p-value < 0.05 and |log₂FC| > 0 were identified as significantly differential metabolites.

Clinical validation

A total of 62 PC tissue samples and 12 blood samples were obtained from patients with PC at Guangdong Provincial People’s Hospital (Guangzhou, Guangdong Province, China). The detailed clinical characteristics of these patients are presented in Table S6. Tissue samples were obtained after tumor resection and confirmed by two pathologists. Plasma samples were collected, processed, and stored at − 80 °C.

A multiplex immunofluorescence kit from Panovue (Panovue, Beijing, China; Cat#10002100050) was used for staining. The staining indicators included anti-PLTP antibody (Proteintech, Wuhan, China; Cat#84592-1-RR), anti-CRTAC1 antibody (Proteintech, Wuhan, China; Cat#13001-1-AP), and anti-ITGAV antibody (Proteintech, Wuhan, China; Cat#84883-5-RR), and nuclear staining used DAPI. Paraffin-embedded PC tissue sections were obtained, deparaffinized, hydrated, subjected to antigen retrieval, and treated to block endogenous peroxidase activity. The primary antibody was incubated at 37 °C for 1 h and the secondary antibody at 37 °C for 30 min. Fluorescence signal amplification was achieved using tyramide signal amplification. Finally, slides were re-stained using DAPI, washed, mounted using an anti-quenching medium, coverslipped, and sealed. The LSM900 confocal microscope system (Carl Zeiss AG) was then used for imaging.

Quantitative detection of PLTP, CRTAC1, and ITGAV in plasma was performed using human enzyme-linked immunosorbent assay (ELISA) kits (Abbexa, Cambridge, United Kingdom) following the manufacturer’s instructions (PLTP, Cat#abx152788; CRTAC1, Cat#abx520680; ITGAV, Cat#abx253772). Final concentrations were determined using a standard curve.

Web design of the online prediction model with Streamlit

To facilitate the clinical application of this model, an interactive web application was built using the Streamlit (v1.48.0) package based on Python 3.10 + for the online deployment of the final prediction model. The application enables real-time risk prediction by allowing users to input the required variables and obtain corresponding risk probability along with SHAP-based, case-level feature attributions. This deployment supports point-of-care accessibility, transparent interpretation, and straightforward integration into routine workflows.

Statistical analysis

Baseline characteristics were summarized as mean ± SD for continuous variables and percentages for categorical variables. Continuous variables were compared using t-tests, Mann–Whitney U tests or Wilcoxon rank-sum tests as appropriate, and categorical variables were compared using chi-square tests. The predictive performance of single and stacked ML models was evaluated using ROC curves. Model construction and validation were performed using Python, while other analyses were conducted using R software (version 4.3.1). All statistical tests were two-tailed, and a p-value < 0.05 was considered statistically significant.

Results

Study population and baseline characteristics

We conducted a descriptive analysis of the demographic and clinical characteristics of 50,862 individuals from the UKB. The cohort included individuals with NODM-PC (n = 190), LTDM-PC (n = 253), nonDM-PC (n = 1,749), NODM-nonPC (n = 7,202), and T2DM (n = 41,468 after excluding NODM-nonPC) (Fig. 2). Detailed comparisons of demographic, laboratory, and anthropometric parameters are presented in Table S7.

Fig. 2.

Study flow diagram. Flowcharts of patient inclusion and exclusion criteria for the UK Biobank study cohort

Feature selection and data selection

In this study, 519 baseline candidate features were compiled. After addressing missing data and assessing their availability, 229 variables were retained for differential comparisons (Table S8). To identify the early predictive features for PC associated with NODM, we compared NODM-PC with LTDM-PC, nonDM-PC, and T2DM groups, and selected intersecting variables. Combined with the important features of conventional modeling, 22 candidate predictors were obtained for modeling. The NODM-PC and NODM-nonPC populations were used for model development. Samples were stratified and randomly divided into training (NODM-PC, n = 133; NODM-nonPC, n = 5,041) and validation sets (NODM-PC, n = 57; NODM-nonPC, n = 2,161) (Fig. 3A-E, Table S9).

Fig. 3.

Feature selection and visualization. (A) Venn diagram showing differential features across three contrasts: NODM-PC vs. LTDM-PC, NODM-PC vs. NonDM-PC, and NODM-PC vs. T2DM. (B–C) Visualization of categorical (B) and continuous (C) features that differed consistently across all three comparisons. (D–E) Distributions of retained categorical (D) and continuous (E) features used for model construction, displayed for the NODM-PC and NODM-nonPC cohorts

ML model development and performance evaluation

We evaluated 22 candidate predictors across seven classifiers—LR, linear SVM, GaussianNB, RF, XGBoost, LightGBM, and CatBoost—using stepwise feature selection (forward and backward) with fivefold cross-validation for hyperparameter tuning and performance estimation. The optimal parameters and performance metrics for each iterative model were generated. Following model training, a stacked ensemble algorithm was constructed. The relationship between feature count and model performance, as well as the evaluation metrics for each algorithm, are presented in Table S10 and Figure S1–S7, including ROC curves, precision-recall curves, accuracy, balanced accuracy, sensitivity, specificity, and NPV. Among the single-algorithm models, CatBoost achieved the highest discrimination (AUROC = 0.844) on the independent validation set, outperforming the other candidate models and demonstrating robust efficacy in distinguishing NODM-PC from NODM populations (Fig. 4A–D). However, the stacked ensemble model, despite integrating five algorithms, did not achieve performance improvement. Furthermore, to assess robustness, we performed a sensitivity analysis excluding individuals diagnosed with T2DM and PC on the same day. The CatBoost model retained strong performance in the validation set (AUROC = 0.801; sensitivity = 86.1%), indicating stable discrimination after reducing potential temporal-order bias (Table S11 and Figure S8).

Fig. 4.

Performance of ML models for predicting NODM-PC. ROC curves for eight models in the training (A) and validation (B) sets; precision-recall curves for the same models in the training (C) and validation (D) sets

Interpretability analysis and development of a clinical web tool

To optimize early PC prediction in the NODM population, we compared alternative feature sets and identified a 14-variable CatBoost model that included glycated hemoglobin, total cholesterol, age, usual walking pace, and trunk fat mass as key predictors. This model demonstrated reliable and consistent performance for NODM-PC prediction and was suitable for early screening and risk assessment (Fig. 5A–D). At the prespecified cutoff of 0.257, predicted probabilities distinctly separated NODM-PC from NODM cases in both the training and validation sets (Fig. 5E–G). The model showed high sensitivity (0.877), supporting effective early detection of NODM-PC, and a specificity of 0.695, enabling reasonable exclusion to identify high-risk individuals. Furthermore, balanced accuracy was 0.786, indicating consistent overall discrimination. These results show that the CatBoost model prioritizes sensitivity while maintaining a high NPV, effectively ruling out low-risk NODM-nonPC cases and capturing most NODM-PC cases (Fig. 5H). DCA further demonstrated that across threshold probabilities of 2–55%, the CatBoost model yielded greater net benefit than treating all or none, supporting its utility in guiding targeted screening decisions in high-risk populations (Fig. 5I). SHAP analyses at both global and individual levels revealed that body composition indices (such as weight and waist circumference), functional phenotypes (such as walking speed), and several metabolic and lifestyle-related variables were the main risk drivers (Fig. 5J–L). To facilitate clinical use, we developed an interactive Streamlit-based web interface that accepts individual patient inputs, generates calibrated risk probabilities, and provides case-level feature attributions (waterfall and force plots), facilitating clinician-patient communication and decision-making (Fig. 5M). By inputting the actual values of the 14 required features, clinicians can automatically obtain an individualized PC risk prediction for patients with NODM. The online application can be accessed through the following link: https://ml-prediction-website-l7ccfnp2yhzfnkxsjqfugx.streamlit.app/.

Fig. 5.

Performance of the CatBoost model. (A) Relationship between the number of included features and AUROC across 144 CatBoost configurations. (B) AUROC of the top-performing CatBoost model at each feature-count level. (C) Relationship between the number of included features and accuracy across 144 CatBoost configurations. (D) Accuracy of the top-performing CatBoost model at each feature-count level. (E) Confusion matrices of the final CatBoost model in the training and validation sets. (F–G) Predicted probabilities for NODM-nonPC vs. NODM-PC in the training (F) and validation (G) sets. (H) Overall performance of the final CatBoost model (summary metrics). (I) DCA for the best CatBoost model. (J) Feature importance for the final CatBoost model. (K–L) SHAP summary plots for the validation set: bar plot (K) and beeswarm plot (L). (M) The web-based interface for predicting NODM-PC using the CatBoost model

Heatmap analysis of CatBoost model variables

Using the CatBoost model’s prespecified threshold, we stratified the NODM cohort into high- (n = 2,360) and low-risk (n = 5,032) groups. To visualize model performance within the NODM population, we generated heatmaps displaying the observed values of key predictors (such as age, HbA1c, and total body fat mass) across individuals. The heatmaps (Fig. 6A) revealed a clear separation between risk strata and concordant patterns across both the training and independent validation sets, supporting strong predictive accuracy and generalizability. These results highlight the model’s potential for early identification of individuals with NODM who are at elevated risk for PC in clinical settings.

Fig. 6.

Integrative multi-omics analysis. (A) Heatmaps of clinical features in the training and validation cohorts. (B) Volcano plots of differentially expressed proteins in the clinical and risk-stratified groups. (C) Venn diagrams showing overlaps of differential signals in proteomics under clinical groups vs. model-defined risk groups. (D–E) Gene Ontology (GO) enrichment results for the clinical (D) and risk-stratified group (E) analyses. (F–G) KEGG pathway enrichment results for the clinical (F) and risk-stratified group (G) analyses. (H) Venn diagrams showing overlaps of differential signals in metabolomics under clinical vs. risk-stratified groups. (I) Volcano plots of differentially expressed metabolites in the clinical and risk-stratified groups. (J) Metabolic pathways enrichment results for the clinical and risk-stratified group analyses

Multi-omics integrative analysis

To prioritize robust biomarkers and minimize noise, we employed a dual-validation strategy by intersecting differentially expressed molecules from the clinical grouping (representing the phenotypic ground truth) and the model-based risk grouping (representing the algorithm’s predictive logic). This convergent strategy isolates biologically stable targets that are consistent across both dimensions, ensuring biological robustness rather than mere model dependency. To further investigate the mechanisms underlying NODM-PC, we analyzed Olink proteomic data from 16 NODM-PC and 752 NODM-nonPC participants, focusing on 1,458 intersecting proteins. Based on the clinical grouping (NODM-PC vs. NODM-nonPC), 93 differentially expressed proteins were identified (including 37 upregulated and 56 downregulated proteins) (Fig. 6B-C and Table S12). Gene Ontology (GO) enrichment analysis (Fig. 6D) showed significant upregulation of pathways associated with leukocyte migration, collagen-containing extracellular matrix, and glycosaminoglycan binding, while pathways related to positive regulation of MAPK cascade, external side of the plasma membrane, and signaling receptor activator activity were significantly downregulated. The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway (Fig. 6F) analysis revealed significant upregulation of the PI3K-Akt signaling pathway, complement and coagulation cascades, and ECM-receptor interaction, along with downregulation of cytokine-cytokine receptor interaction and viral protein interaction with cytokines and their receptors. Based on the model-based risk stratification, the high-risk group included 230 samples with proteomic data (16 from the NODM-PC group and 214 from the NODM group), while the low-risk group comprised 538 samples with proteomic data (all from the NODM group) (Figure S9A-B). A total of 343 differentially expressed proteins were identified (117 upregulated and 226 downregulated proteins) (Fig. 6B-C and Table S13). GO enrichment analysis (Fig. 6E) indicated significant upregulation of pathways associated with viral entry into host cells, protein-lipid complexes, and virus receptor activity. In contrast, pathways associated with nervous system development, external side of the plasma membrane, and signaling receptor activator activity were significantly downregulated. KEGG pathway analysis (Fig. 6G) revealed upregulation of the PI3K-Akt signaling pathway, ether lipid metabolism, and cholesterol metabolism and downregulation of cytokine-cytokine receptor interaction and TNF signaling pathways. Intersecting the upregulated and downregulated results from both groups yielded 39 key proteins with consistent expression trends (13 upregulated and 26 downregulated proteins) (Fig. 6C). The expression trends of these proteins were similar in the heatmaps of the clinical and model-based risk groups (Figure S9E-F). Notably, proteins such as PLTP, CRTAC1, and ITGAV were highly expressed in NODM-PC. Functional enrichment analysis showed strong consistency between both groupings, particularly in pathways involving PI3K-Akt signaling and cytokine-cytokine receptor interaction. These findings further prove that the CatBoost model distinguishes PC in NODM populations with high stability at the molecular level.

Furthermore, 251 metabolites were detected using NMR metabolomics, including 103 NODM-PC cases and 3,921 NODM-nonPC cases. Based on the clinical grouping (NODM-PC vs. NODM-nonPC), 147 differential metabolites (119 upregulated and 28 downregulated) were identified (Fig. 6H-I and Table S14), followed by enrichment and pathway topological analysis. After model-based risk stratification, the high-risk group included 1,332 samples (99 NODM-PC and 1,233 NODM-nonPC), while the low-risk group comprised 2,692 cases (4 NODM-PC and 2,688 NODM-nonPC) (Figure S9C-D). The differential analysis of the two groups identified 210 metabolites (170 upregulated and 40 downregulated) (Fig. 6H-I and Table S15). The intersection of both groupings yielded 145 consistent metabolites (118 upregulated and 27 downregulated) (Figure S9G-H). In the clinical grouping, six KEGG pathways were significantly upregulated and six were significantly downregulated. Overall, 11 KEGG pathways (Fig. 6J) highly overlapped between the clinical and risk-based analyses: 6 upregulated and 5 downregulated, suggesting the stability of metabolic pathway perturbations across groupings. To further evaluate the relevance of these pathways in the metabolic network, we performed enrichment and topological analyses on the above 12 consistently significant KEGG pathways. Five significantly enriched pathways (p < 0.05) were identified, of which valine, leucine, and isoleucine biosynthesis; linoleic acid metabolism; and neomycin, kanamycin, and gentamicin biosynthesis simultaneously showed higher impact values (> 0.2), indicating that they occupy key topological positions within the metabolic network and have significant impacts on overall metabolic regulation.

Key molecules and clinical validation

To improve the clinical applicability of the early diagnostic model, we further investigated early biomarkers of NODM-PC in clinical samples. Based on the proteomic data from the Olink platform by Athanasiou et al. [24], 525 proteins were upregulated in NODM-PC compared with those in nonDM-PC. Additionally, analysis of the Clinical Proteomic Tumor Analysis Consortium (CPTAC; https://gdc.cancer.gov/about-gdc/contributed-genomic-data-cancer-research/clinical-proteomic-tumor-analysis-consortium-cptac) dataset revealed 2407 proteins upregulated in PC relative to the healthy population. Integrating these findings with our aforementioned clinical and risk-based groups, compared with NODM-nonPC, 13 upregulated proteins in NODM-PC were cross-analyzed with the above upregulated proteins, revealing 3 NODM-PC-related proteins: PLTP, CRTAC1, and ITGAV (Fig. 7A-B, Figure S10A-B). Across three independent analyses, all biomarkers showed positive associations (log(OR) > 0); CRTAC1 and ITGAV exhibited consistent and statistically significant effects across cohorts, whereas PLTP demonstrated a modest-to-moderate effect overall (Fig. 7C). We explored the correlation of PLTP, CRTAC1, and ITGAV in PC stages using CPTAC data. The results showed no significant differences in the expression of these proteins between early- and late-stage tumors compared with that in the healthy population, suggesting their potential stability as early-stage biomarkers for NODM-PC (Fig. 7D). Subsequently, we conducted a validation in our clinical cohort. ELISA of patient plasma revealed elevated PLTP, CRTAC1, and ITGAV levels in NODM-PC compared with those in nonDM-PC (Fig. 7E). Additionally, multiplex immunohistochemistry of patient tissues revealed similar overexpression patterns in NODM-PC compared with those in nonDM-PC (Fig. 7F). These findings indicate that PLTP, CRTAC1, and ITGAV are closely associated with NODM-PC and are promising biomarkers for its early diagnosis.

Fig. 7.

Biomarker prioritization and clinical validation. (A–B) Schematic illustration for screening. (A) and Venn diagram (B) of upregulated proteins in the NODM-PC population of this study, the PC population in the CPTAC database, and the Olink-based NODM-PC proteomic studies. (C) Forest plots for the three candidate biomarkers (PLTP, CRTAC1, ITGAV) summarizing effect sizes across four datasets: NODM-PC cohort of the clinical groups, NODM-PC cohort of the risk-stratified groups, CPTAC, and the Olink studies. (D) CPTAC expression of PLTP, CRTAC1, and ITGAV across healthy, Stage I–II, and Stage III–IV groups. (E) Plasma ELISA quantification of PLTP, CRTAC1, and ITGAV in nonDM-PC vs. NODM-PC (n = 6). (F) Representative HE and mIHC images with quantitative analyses of PLTP, CRTAC1, and ITGAV expression in nonDM-PC and NODM-PC. Scale bars, 50 μm. The statistical difference was assessed by one-way ANOVA followed by Dunnett’s tests in (D); and the two-tailed Student’s t test in (E). Data are indicated as mean ± SD from three independent experiments. (ns: not significant; *: p < 0.05; **: p < 0.01; ***: p < 0.001)

Discussion

This study used NODM as a high-yield entry point for population enrichment to develop and validate an interpretable ML framework based on population-level data, integrated with multi-omics data to clarify the biology of NODM-PC. The model demonstrated consistent performance across training and validation cohorts, enabling risk stratification within the NODM population. Combined with independent omics signals, it identified biologically plausible pathways and biomarkers. Thus, we propose a two-step early-screening strategy: “population first, then individual.” First, a clinical ML model stratifies risk at the population level. Second, high-risk individuals undergo confirmatory testing with serum biomarkers to facilitate the early diagnosis of NODM-PC. This strategy positions NODM as the “first screen”, providing a practical entry point to achieve a stage shift in NODM-PC detection.

Given the low incidence of PC in the general population, population-wide screening is inefficient and resource-intensive [25, 26]. Increasing diagnostic yield requires focusing on high-risk subgroups [27]. For instance, Boursi et al. [28]. proposed a model using routine clinical variables in which approximately 6.19% of individuals with NODM were selected for second-line screening at a 1% 3-year risk threshold, achieving an area under curve (AUC) of 0.82 and demonstrating a resource-feasible approach. Similarly, the Mayo Clinic’s END-PAC score, based on age, weight change, and glycemic change, achieved an AUC of 0.87 in validation and enriched 3-year incidence to approximately 3.6% in the high-risk stratum, underscoring the value of longitudinal metabolic trajectories [29]. Our work emphasizes the necessity of “further enrichment” in the NODM population. Compared to existing tools, our model optimizes variable complexity by relying only on single-timepoint admission data. It demonstrates robust discrimination with superior sensitivity (87.7%), minimizing missed diagnoses, and provides unique mechanistic insight through multi-omics validation, supporting its clinical utility. To mitigate overfitting and spurious findings, we applied conservative feature selection, five-fold cross-validation, and examined the biological concordance between risk stratification and clinical groupings. This approach establishes a closed-loop strategy that integrates population-level risk stratification with individual multi-omics validation for robust clinical model development. We further enrich within NODM using an interpretable ML model to stratify risk. SHAP attributions highlight directional contributions from weight change, glycemic indices, lipid metabolism, and anthropometric indicators, which can significantly increase the positive predictive value while reducing the cost and burden associated with broad screening.

Identifying high-performance biomarkers is another strategy for improving screening efficiency [30–32]. To effectively identify individuals at high risk of NODM-PC, our model establishes a bridge between prediction and biological mechanism through its interpretability and integration with multi-omics data. Previous studies have identified the tumor-secreted diabetogenic factors such as galectin-3 and S100A9, which distinguish PC-related diabetes (PCDM) from T2DM, achieving AUCs of approximately 0.83 and 0.77, respectively, in the validation set. These findings suggest that approximately 1% of NODM cases are actually PCDM, and approximately 50% of PC cases develop the PCDM phenotype within 24 months before diagnosis [33]. Therefore, we integrated proteomic and metabolomic data and observed coordinated perturbations along the inflammation-lipid metabolism-insulin signaling axis, consistent with the proposed mechanisms of PCDM. Additionally, clinical validation confirmed PLTP, CRTAC1, and ITGAV as early, actionable indicators of NODM-PC, thereby providing biological support for model interpretability and identifying practical biomarkers for triage and confirmation. Mechanistically, these biomarkers bridge systemic dysmetabolism with tumor aggressiveness. ITGAV drives PC metastasis and fibrosis-associated insulin resistance [34, 35], whereas PLTP facilitates lipid remodeling critical for tumor adaptation [36, 37]. Furthermore, CRTAC1, robustly linked to BMI, suggests a stromally mediated metabolic support system [38, 39]. Collectively, these markers reflect the inflammation–lipid–insulin axis underlying NODM-PC pathogenesis.

The prevention and treatment of PC are associated with extremely poor outcomes, primarily owing to difficulties in early detection, the lack of effective radical treatment methods, and limited improvement in prognosis [40–42]. Given the persistently low 5-year survival rate, meaningful progress will require shifting the diagnosis to earlier stages so that more than 10–15% of patients become eligible for curative resection. Our research provides a feasible framework toward this goal. Because population-wide screening is not feasible, efforts should focus on high-risk groups, using NODM as the first screening tool. To translate model performance into an actionable screening strategy, we identified an optimal risk threshold of 0.257 based on the Youden index and DCA curve. At this threshold, the model stratifies patients into a high-risk group warranting secondary screening strategies (e.g., imaging examinations and biomarker rechecks). Significantly, this targeted strategy effectively minimizes unnecessary imaging burden by excluding approximately 70% of low-risk individuals (specificity: 69.5%). Crucially, the model maintains a high sensitivity of 87.7%, and only NODM patients with a predicted risk > 25.7% are referred for follow-up imaging (e.g., CT/MRI). Clinicians may adjust this referral threshold based on local healthcare resources and risk tolerance, thereby improving the likelihood of earlier diagnosis and enhancing PC survival.

This study has several limitations. First, the number of NODM-PC cases, particularly within the multi-omics sub-cohorts, was relatively limited. Clinically, early-stage NODM-PC is difficult to ascertain because PC often presents insidiously, and approximately 80% of patients are diagnosed at intermediate/advanced stages [43]. Consequently, assembling real-world data for early-stage NODM-PC is inherently challenging, underscoring the need for NODM-based early-detection strategies. Given these constraints, we proactively implemented rigorous measures to mitigate overfitting, including stratified 5-fold cross-validation, conservative feature selection, and class-weighting. Furthermore, we quantified model uncertainty by reporting bootstrap-derived confidence intervals on an independent validation set comprising 30% of the cohort. Second, while our study lacked an external cohort with harmonized multi-omics data and the relatively small sample size for clinical validation, the practical feasibility of our model’s routine variables and the preliminary yet promising cross-platform consistency of PLTP, CRTAC1, and ITGAV support their potential for broader application, warranting further verification in larger, multicenter prospective studies. Third, we were unable to directly benchmark our model against established scores such as END-PAC or Boursi within this cohort. These legacy models require longitudinal data to calculate temporal trajectories (e.g., weight change over 12 months), whereas our model was designed to utilize cross-sectional baseline features for immediate risk stratification. Fourth, although pancreatogenic diabetes (type 3c DM) is widely regarded as an important early signal for PC screening, its biological basis is not fully understood. In addition, despite extensive covariate adjustment, residual confounding and reverse causation remain possible—particularly if occult tumors influence glycemic control and inflammatory markers before clinical diagnosis. Fifth, the UK Biobank is a volunteer cohort enriched for middle-aged adults of predominantly European ancestry, which may limit generalizability to younger individuals, non-European populations, and higher-risk clinical settings. Finally, studies should include multicenter prospective validation, hierarchical integration with polygenic risk scores, and the development of functional and biofluid-based assays for candidate biomarkers.

Conclusion

In conclusion, we used NODM as a high-yield entry point and applied an interpretable ML model to identify individuals at elevated risk for early PC. We combined this framework with integrated proteomic and metabolomic analyses to provide a biological rationale and identify actionable biomarkers and therapeutic targets. This closed-loop framework—progressing from population-level enrichment to model construction, mechanistic interrogation, and strategy formulation—provides a realistic pathway toward meaningful clinical translation and improved survival outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

PC

Pancreatic cancer

NODM

New-onset diabetes mellitus

ML

Machine learning

NODM-PC

Pancreatic cancer with new-onset diabetes mellitus

UKB

UK Biobank

T2DM

Type 2 diabetes mellitus

LTDM-PC

Pancreatic cancer with long-term diabetes mellitus

nonDM-PC

Pancreatic cancer without diabetes mellitus

NODM-nonPC

New-onset diabetes mellitus without pancreatic cancer

CatBoost

Categorical Boosting

GaussianNB

Gaussian Naïve Bayes

LightGBM

Light Gradient-Boosting Machine

Linear SVM

Linear Support Vector Machine

LR

Logistic Regression

RF

Random Forest

XGBoost

Extreme Gradient Boosting

AUROC

Area Under the Receiver Operating Characteristic Curve

ROC

Receiver operating characteristic curve

NPV

Negative predictive value

DCA

Decision curve analysis

SHAP

Shapley additive explanations

MICE

Multivariate imputation by chained equations

CPTAC

Clinical Proteomic Tumor Analysis Consortium

ELISA

Enzyme-linked immunosorbent assay

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

AUC

Area under curve

PCDM

PC-related diabetes mellitus

mIHC

Multiplex immunohistochemistry

Author contributions

J.Y. contributed to methodology, formal analysis, and to writing the manuscript; B.C., A.Y. and Y.Y.G. contributed to methodology, formal analysis and visualization; Y.S.G., H.Z., K.Z. and S.W. contributed to investigation and data collection; Q.H., J.Q.L., T.Y. and J.W.L. contributed to data curation and validation; Y.Z. and R.C. contributed to the conception, supervision and design of the study. All authors contributed substantially to the finalization of the manuscript and approved the submitted version, and Y.Z. and R.C., as guarantors of this work, had full access to all data and take responsibility for the integrity of the data and the accuracy of the analyses.

Funding

This work was funded by the National Key Research and Development Program of China (2023YFC2413401); the National Natural Science Foundation of China (82372858, 82072639); the Science and Technology Program of Guangzhou, China (2024B01J1379); the Precision Diagnosis and Treatment Center for Liver, Gallbladder, and Pancreas Diseases Project of Guangdong Provincial People’s Hospital, China (KY012023453); .

Data availability

These data were derived from the following resources available in the public domain: UKBiobank, https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access. The interactive Streamlit-based web tool is accessible via the following link: https://ml-prediction-website-l7ccfnp2yhzfnkxsjqfugx.streamlit.app/. Correspondingly, the complete source code for this tool is openly available on GitHub (https://github.com/AY0101AY/ML-Prediction-Website).

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki. Ethical approval was obtained from the UK NHS Research Ethics Committee (21/NW/0157) for the UK Biobank cohort and from the Ethics Committee of Guangdong Provincial People’s Hospital (approval number: KY2025-682-01) for the clinical validation cohort. Written informed consent was obtained from all participants involved in this study.

Research registration unique identifying number (UIN)

This research has been conducted using the UK Biobank Resource under Application Number-85244.

Conflicts of interest

The authors declare no conflict of interest.