Pancreatic Cancer Risk Stratification across Diabetes Stages: Development and Internal Validation of a Machine Learning Model.

BACKGROUND

Pancreatic cancer is diagnosed at advanced stages in patients with diabetes. Existing prediction models require complete historical data and focus on new-onset diabetes, limiting applicability. We developed a machine learning model to handle missing data and perform across the diabetes spectrum.

METHODS

This is a retrospective cohort study using TriNetX electronic health records. Patients with hemoglobin A1c ≥6.5% were included. Sixty-five clinical variables were extracted at 90-day intervals. An eXtreme Gradient Boosting (XGBoost) model was developed using patient-level 1:1 case-control sampling and compared with the Enriching New‑Onset Diabetes For Pancreatic Cancer (ENDPAC), Boursi, and Cheung models using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and lead time.

RESULTS

Among 3,213,551 patients (mean age 56.7 years; 46.8% female), 2,655 (0.08%) developed pancreatic cancer. XGBoost achieved an AUROC of 0.78 (95% confidence interval, 0.77-0.79). At 90% sensitivity, the specificity was 50% with a median 9-month lead time. The model scored 100% of the patients versus <10% for existing models. On matched patient subsets with complete data, XGBoost significantly outperformed ENDPAC (AUROC 0.79 vs. 0.63; P < 0.001) and Cheung (AUROC 0.84 vs. 0.75; P = 0.045) models. The Boursi model could not be reliably evaluated due to insufficient scorable patients.

CONCLUSIONS

This XGBoost model predicts pancreatic cancer among patients with both new-onset and prevalent diabetes in the setting of limited clinical information, achieving 100% patient coverage versus <10% for existing models. External validation is needed before clinical implementation.

IMPACT: Existing models focus exclusively on new-onset diabetes and require complete historical data, scoring fewer than 10% of patients. This model risk-stratifies both new-onset and prevalent diabetes with limited clinical information, achieving 100% patient coverage pending external validation.